artificial-intelligence-basics-a-non-technical-introduction - Inteligência Artificial (2025)

Prévia do material em texto

<p>A Non-Technical Introduction</p><p>―</p><p>Tom Taulli</p><p>Artificial</p><p>Intelligence</p><p>Basics</p><p>ARTIFICIAL INTELLIGENCE</p><p>BASICS</p><p>A NON-TECHNICAL INTRODUCTION</p><p>TomTaulli</p><p>Artificial Intelligence Basics: A Non-Technical Introduction</p><p>ISBN-13 (pbk): 978-1-4842-5027-3 ISBN-13 (electronic): 978-1-4842-5028-0</p><p>https://doi.org/10.1007/978-1-4842-5028-0</p><p>Copyright © 2019 by Tom Taulli</p><p>This work is subject to copyright. All rights are reserved by the Publisher, whether the whole</p><p>or part of the material is concerned, specifically the rights of translation, reprinting, reuse of</p><p>illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical</p><p>way, and transmission or information storage and retrieval, electronic adaptation, computer</p><p>software, or by similar or dissimilar methodology now known or hereafter developed.</p><p>Trademarked names, logos, and images may appear in this book. Rather than use a</p><p>trademark symbol with every occurrence of a trademarked name, logo, or image we use</p><p>the names, logos, and images only in an editorial fashion and to the benefit of the trademark</p><p>owner, with no intention of infringement of the trademark.</p><p>The use in this publication of trade names, trademarks, service marks, and similar terms,</p><p>even if they are not identified as such, is not to be taken as an expression of opinion as to</p><p>whether or not they are subject to proprietary rights.</p><p>While the advice and information in this book are believed to be true and accurate at the</p><p>date of publication, neither the authors nor the editors nor the publisher can accept any</p><p>legal responsibility for any errors or omissions that may be made. The publisher makes no</p><p>warranty, express or implied, with respect to the material contained herein.</p><p>Managing Director, Apress Media LLC: Welmoed Spahr</p><p>Acquisitions Editor: Shiva Ramachandran</p><p>Development Editor: Rita Fernando</p><p>Coordinating Editor: Rita Fernando</p><p>Cover designed by eStudioCalamar</p><p>Distributed to the book trade worldwide by Springer Science+Business Media NewYork,</p><p>233 Spring Street, 6th Floor, New York, NY 10013. Phone 1-800-SPRINGER, fax (201)</p><p>348-4505, e-mail orders-ny@springer-sbm.com, or visit www.springeronline.com. Apress</p><p>Media, LLC is a California LLC and the sole member (owner) is Springer Science + Business</p><p>Media Finance Inc (SSBM Finance Inc). SSBM Finance Inc is a Delaware corporation.</p><p>For information on translations, please e-mail rights@apress.com, or visit http://www.</p><p>apress.com/rights-permissions.</p><p>Apress titles may be purchased in bulk for academic, corporate, or promotional use. eBook</p><p>versions and licenses are also available for most titles. For more information, reference our</p><p>Print and eBook Bulk Sales web page at http://www.apress.com/bulk-sales.</p><p>Any source code or other supplementary material referenced by the author in this book is</p><p>available to readers on GitHub via the book's product page, located at www.apress.</p><p>com/978-1-4842-5027-3. For more detailed information, please visit http://www.apress.</p><p>com/source-code.</p><p>Printed on acid-free paper</p><p>TomTaulli</p><p>Monrovia, CA, USA</p><p>https://doi.org/10.1007/978-1-4842-5028-0</p><p>Chapter 1: AI Foundations � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � 1</p><p>Chapter 2: Data � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � �19</p><p>Chapter 3: Machine Learning � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � �39</p><p>Chapter 4: Deep Learning � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � �69</p><p>Chapter 5: Robotic Process Automation (RPA) � � � � � � � � � � � � � � � � � � � �91</p><p>Chapter 6: Natural Language Processing (NLP) � � � � � � � � � � � � � � � � � �103</p><p>Chapter 7: Physical Robots � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � �125</p><p>Chapter 8: Implementation of AI � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � �143</p><p>Chapter 9: The Future of AI � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � �161</p><p>Appendix: AI Resources � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � �177</p><p>Glossary � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � �179</p><p>Index � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � �185</p><p>Contents</p><p>About the Author � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � v</p><p>Foreword � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � vii</p><p>Introduction � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � �ix</p><p>About the Author</p><p>TomTaulli has been developing software since</p><p>the 1980s. In college, he started his first com-</p><p>pany, which focused on the development of</p><p>e-learning systems. He created other companies</p><p>as well, including Hypermart.net that was sold to</p><p>InfoSpace in 1996. Along the way, Tom has writ-</p><p>ten columns for online publications such as busi-</p><p>nessweek.com, techweb.com, and Bloomberg.</p><p>com. He also writes posts on artificial intelli-</p><p>gence for Forbes.com and is the advisor to vari-</p><p>ous companies in the AI space. You can reach</p><p>Tom on Twitter (@ttaulli) or through his web</p><p>site (www.taulli.com).</p><p>http://www.taulli.com/</p><p>Foreword</p><p>As this book demonstrates, the adoption of artificial intelligence (AI) will be a</p><p>major inflection point in human history. Like other similarly groundbreaking</p><p>technologies, how it's administered and who has access to it will shape society</p><p>for generations to come. However, AI stands out from the other transforma-</p><p>tive technologies of the nineteenth and twentieth centuries—think the steam</p><p>engine, the electrical grid, genomics, computers, and the Internet—because it</p><p>doesn't depend exclusively on critically expensive physical infrastructure to</p><p>enable adoption; after all, many of its benefits can be delivered through existing</p><p>hardware we all carry around in our pockets. Instead, the fundamental limiting</p><p>factor when it comes to the mass adoption of AI technology is our shared intel-</p><p>lectual infrastructure: education, understanding, and vision.</p><p>This is a crucial difference because, if handled correctly, AI can act as a sweep-</p><p>ing democratizing force. It has and will eliminate from our lives the drudgery</p><p>of the past and free up a tremendous amount of human energy and capital.</p><p>But that "if" is far from certain. AI executed irresponsibly has the power to</p><p>destabilize large parts of the world economy by causing, as many people fear,</p><p>a shrinking workforce, reduced purchasing power for the middle class, and an</p><p>economy without a wide and stable base fueled by an endless debt spiral.</p><p>However, before we succumb to pessimism on AI, we should take a look back.</p><p>Historic though AI's transformative capacity may be—and it is historic—these</p><p>same issues are and have been at play in the economic landscape for decades,</p><p>even centuries. AI is, after all, an extension of a trend toward automation that</p><p>has been at play since Henry Ford. In fact, Zoho itself was born from the ten-</p><p>sion between automation and egalitarian economic principles. Back in the</p><p>early 2000s, we came to a realization that has shaped our approach to tech-</p><p>nology: regular people—small business owners, here and abroad—should</p><p>have access to the same advanced business automations that the Fortune 500</p><p>companies have; otherwise, a huge swath of the population will be locked out</p><p>of the economy.</p><p>At the time, powerful digital software was almost unanimously gated behind</p><p>rigid contracts, exorbitant fee structures, and complicated on-premise imple-</p><p>mentations. Big companies could shoulder the burden of such systems, while</p><p>smaller operators were locked out, putting them at a tremendous disadvan-</p><p>tage. We sought to disrupt that by opening up the promise of technology to</p><p>wider and wider audiences. Over the last two decades, we've endeavored to</p><p>viii</p><p>increase</p><p>efforts.</p><p>It was not until 1993 that IBM came out with its own relational database,</p><p>DB2. But it was too late. By this time, Oracle was the leader in the database</p><p>market.</p><p>Through the 1980s and 1990s, the relational database was the standard for</p><p>mainframe and client-server systems. But when Big Data became a factor, the</p><p>technology had serious flaws like the following:</p><p>• Data Sprawl: Over time, different databases would spread</p><p>across an organization. The result was that it got tougher</p><p>to centralize the data.</p><p>• New Environments: Relational database technology was</p><p>not built for cloud computing, high-velocity data, or</p><p>unstructured data.</p><p>• High Costs: Relational databases can be expensive. This</p><p>means that it can be prohibitive to use the technology for</p><p>AI projects.</p><p>• Development Challenges: Modern software development</p><p>relies heavily on iterating. But relational databases have</p><p>proven challenging for this process.</p><p>In the late 1990s, there were open source projects developed to help create next-</p><p>generation database systems. Perhaps the most critical one came from Doug</p><p>Cutting who developed Lucene, which was for text searching. The technology</p><p>was based on a sophisticated index system that allowed for low-latency</p><p>performance. Lucene was an instant hit, and it started to evolve, such as with</p><p>Apache Nutch that efficiently crawled the Web and stored the data in an index.</p><p>But there was a big problem: To crawl the Web, there needed to be an</p><p>infrastructure that could hyperscale. So in late 2003, Cutting began</p><p>development on a new kind of infrastructure platform that could solve the</p><p>problem. He got the idea from a paper published from Google, which</p><p>described its massive file system. A year later, Cutting had built his new</p><p>platform, which allowed for sophisticated storage without the complexity. At</p><p>the core of this was MapReduce that allowed processing across multiple</p><p>servers. The results would then be merged, allowing for meaningful reports.</p><p>Chapter 2 | Data</p><p>27</p><p>Eventually, Cutting’s system morphed into a platform called Hadoop—and it</p><p>would be essential for managing Big Data, such as making it possible to create</p><p>sophisticated data warehouses. Initially, Yahoo! used it, and then it quickly</p><p>spread, as companies like Facebook and Twitter adopted the technology.</p><p>These companies were now able to get a 360 view of their data, not just</p><p>subsets. This meant there could be more effective data experiments.</p><p>But as an open source project, Hadoop still lacked the sophisticated systems</p><p>for enterprise customers. To deal with this, a startup called Hortonworks</p><p>built new technologies like YARN on top of the Hadoop platform. It had</p><p>features like in-memory analytic processing, online data processing, and</p><p>interactive SQL processing. Such capabilities supported adoption of Hadoop</p><p>across many corporations.</p><p>But of course, there emerged other open source data warehouse projects.</p><p>The well-known ones, like Storm and Spark, focused on streaming data.</p><p>Hadoop, on the other hand, was optimized for batch processing.</p><p>Besides data warehouses, there was also innovation of the traditional database</p><p>business. Often these were known as NoSQL systems. Take MongoDB. It</p><p>started as an open source project and has turned into a highly successful</p><p>company, which went public in October 2017. The MongoDB database, which</p><p>has over 40 million downloads, is built to handle cloud, on-premise, and hybrid</p><p>environments.9 There is also much flexibility structuring the data, which is</p><p>based on a document model. MongoDB can even manage structured and</p><p>unstructured data at high petabyte scale.</p><p>Even though startups have been a source of innovation in database systems</p><p>and storage, it’s important to note that the mega tech operators have also</p><p>been critical. Then again, companies like Amazon.com and Google have had</p><p>to find ways to deal with the huge scale of data because of the need for</p><p>managing their massive platforms.</p><p>One of the innovations has been the data lake, which allows for seamless</p><p>storage of structured and unstructured data. Note that there is no need to</p><p>reformat the data. A data lake will handle this and allow you to quickly perform</p><p>AI functions. According to a study from Aberdeen, companies who use this</p><p>technology have an average of 9% organic growth compared to those who do</p><p>not.10</p><p>Now this does not mean you have to get rid of your data warehouses. Rather,</p><p>both serve particular functions and use cases. A data warehouse is generally</p><p>good for structured data, whereas a data lake is better for diverse environments.</p><p>What’s more, it’s likely that a large portion of the data will never be used.</p><p>9 www.mongodb.com/what-is-mongodb</p><p>10 https://aws.amazon.com/big-data/datalakes-and-analytics/</p><p>what-is-a-data-lake/</p><p>Artificial Intelligence Basics</p><p>http://www.mongodb.com/what-is-mongodb</p><p>https://aws.amazon.com/big-data/datalakes-and-analytics/what-is-a-data-lake/</p><p>https://aws.amazon.com/big-data/datalakes-and-analytics/what-is-a-data-lake/</p><p>28</p><p>For the most part, there are a myriad of tools. And expect more to be</p><p>developed as data environments get more complex.</p><p>But this does not mean you should chose the latest technology. Again, even</p><p>older relational databases can be quite effective with AI projects. The key is</p><p>understanding the pros/cons of each and then putting together a clear strategy.</p><p>Data Process</p><p>The amount of money shelled out on data is enormous. According to IDC,</p><p>the spending on Big Data and analytics solutions is forecasted to go from $166</p><p>billion in 2018 to $260 billion by 2022.11 This represents an 11.9% compound</p><p>annual growth rate. The biggest spenders include banks, discrete manufacturers,</p><p>process manufacturers, professional service firms, and the federal government.</p><p>They account for close to half the overall amount.</p><p>Here’s what IDC’s Jessica Goepfert—the program vice president (VP) of</p><p>Customer Insights and Analysis—said:</p><p>At a high level, organizations are turning to Big Data and analytics</p><p>solutions to navigate the convergence of their physical and digital worlds.</p><p>This transformation takes a different shape depending on the industry.</p><p>For instance, within banking and retail—two of the fastest growth areas</p><p>for Big Data and analytics—investments are all about managing and</p><p>reinvigorating the customer experience. Whereas in manufacturing, firms</p><p>are reinventing themselves to essentially be high tech companies, using</p><p>their products as a platform to enable and deliver digital services.12</p><p>But a high level of spending does not necessarily translate into good results.</p><p>A Gartner study estimates that roughly 85% of Big Data projects are</p><p>abandoned before they get to the pilot stage.13 Some of the reasons include</p><p>the following:</p><p>• Lack of a clear focus</p><p>• Dirty data</p><p>• Investment in the wrong IT tools</p><p>• Problems with data collection</p><p>• Lack of buy-in from key stakeholders and champions in</p><p>the organization</p><p>11 www.idc.com/getdoc.jsp?containerId=prUS44215218</p><p>12 www.idc.com/getdoc.jsp?containerId=prUS44215218</p><p>13 www.techrepublic.com/article/85-of-big-data-projects-fail-</p><p>but-your-developers-can-help-yours-succeed/</p><p>Chapter 2 | Data</p><p>http://www.idc.com/getdoc.jsp?containerId=prUS44215218</p><p>http://www.idc.com/getdoc.jsp?containerId=prUS44215218</p><p>http://www.techrepublic.com/article/85-of-big-data-projects-fail-but-your-developers-can-help-yours-succeed/</p><p>http://www.techrepublic.com/article/85-of-big-data-projects-fail-but-your-developers-can-help-yours-succeed/</p><p>29</p><p>In light of this, it is critical to have a data process. Notwithstanding there are</p><p>many approaches—often extoled by software vendors—there is one that has</p><p>widespread acceptance. A group of experts, software developers, consultants,</p><p>and academics created the CRISP-DM Process in the late 1990s. Take a look</p><p>at Figure2-1 for a visual.</p><p>In this chapter, we’ll take a look at steps #1 through #3. Then in the rest of</p><p>the book, we’ll cover the remaining ones (that is, we will look at Modelling</p><p>and Evaluation in Chapter 3 and Deployment</p><p>in Chapter 8).</p><p>Note that steps #1–#3 can account for 80% of the time of the data</p><p>process, which is based on the experience of Atif Kureishy, who is the</p><p>global VP of Emerging Practices at Teradata.14 This is due to factors like:</p><p>14 This is from the author’s interview with Atif Kureishy in February 2019.</p><p>Figure 2-1. The CRISP-DM Process</p><p>Artificial Intelligence Basics</p><p>30</p><p>The data is not well organized and comes from different sources (whether</p><p>from different vendors or silos in the organization), there is not enough</p><p>focus on automation tools, and the initial planning was insufficient for the</p><p>scope of the project.</p><p>It’s also worth keeping in mind that the CRISP-DM Process is not a strict</p><p>linear process. When dealing with data, there can be much iteration. For</p><p>example, there may be multiple attempts at coming up with the right data and</p><p>testing it.</p><p>Step #1—Business Understanding</p><p>You should come up with a clear view of the business problem to be solved.</p><p>Some examples:</p><p>• How might a price adjustment impact your sales?</p><p>• Will a change in copy lead to improved conversion of</p><p>digital ads?</p><p>• Does a fall in engagement mean there will be an increase</p><p>in churn?</p><p>Then, you must establish how you will measure success. Might it be that sales</p><p>should increase by at least 1% or that conversions should rise by 5%?</p><p>Here’s a case from Prasad Vuyyuru, who is a partner at the Enterprise Insights</p><p>Practice of Infosys Consulting:</p><p>Identifying which business problem to solve using AI and assessing what</p><p>value will be created are critical for the success of all AI projects. Without</p><p>such diligent focus on business value, AI projects risk not getting adopted</p><p>in the organization. AB Inbev’s experience in using AI to identify packag-</p><p>ing line motors that are likely to fail is a great example of how AI is creat-</p><p>ing practical value. ABInbev installed 20 wireless sensors to measure</p><p>vibrations at packaging lines motors. They compared sounds with nor-</p><p>mally functioning motors to identify anomalies which predicted eventual</p><p>failure of the motors.15</p><p>Regardless of the goal, it’s essential that the process be free of any prejudgments</p><p>or bias. The focus is to find the best results. No doubt, in some cases, there</p><p>will not be a satisfactory result.</p><p>Or, in other situations, there may be big surprises. A famous example of this</p><p>comes from the book Moneyball by Michael Lewis, which was also made into</p><p>a movie in 2011 that starred Brad Pitt. It’s a true story of how the Oakland</p><p>A’s used data science techniques to recruit players. The tradition in baseball</p><p>15 This is from the author’s interview of Prasad Vuyyuru in February 2019.</p><p>Chapter 2 | Data</p><p>31</p><p>was to rely on metrics like batting averages. But when using sophisticated data</p><p>analytics techniques, there were some startling results. The Oakland A’s</p><p>realized that the focus should be on slugging and on-base percentages. With</p><p>this information, the team was able to recruit high-performing players at</p><p>lower compensation levels.</p><p>The upshot is that you need to be open minded and willing to experiment.</p><p>In step #1, you should also assemble the right team for the project. Now</p><p>unless you work at a company like Facebook or Google, you will not have the</p><p>luxury of selecting a group of PhDs in machine learning and data science. Such</p><p>talent is quite rare—and expensive.</p><p>But you also do not need an army of top-notch engineers for an AI project</p><p>either. It is actually getting easier to apply machine learning and deep learning</p><p>models, because of open source systems like TensorFlow and cloud-based</p><p>platforms from Google, Amazon.com, and Microsoft. In other words, you may</p><p>only need a couple people with a background in data science.</p><p>Next, you should find people—likely from your organization—who have the</p><p>right domain expertise for the AI project. They will need to think through the</p><p>workflows, models, and the training data—with a particular understanding of</p><p>the industry and customer requirements.</p><p>Finally, you will need to evaluate the technical needs. What infrastructure and</p><p>software tools will be used? Will there be a need to increase capacity or</p><p>purchase new solutions?</p><p>Step #2—Data Understanding</p><p>In this step, you will look at the data sources for the project. Consider that</p><p>there are three main ones, which include the following:</p><p>• In-House Data: This data may come from a web site,</p><p>beacons in a store location, IoT sensors, mobile apps, and</p><p>so on. A major advantage of this data is that it is free and</p><p>customized to your business. But then again, there are</p><p>some risks. There can be problems if there has not been</p><p>enough attention on the data formatting or what data</p><p>should be selected.</p><p>• Open Source Data: This is usually freely available, which is</p><p>certainly a nice benefit. Some examples of open source</p><p>data include government and scientific information. The</p><p>data is often accessed through an API, which makes the</p><p>process fairly straightforward. Open source data is also</p><p>usually well formatted. However, some of the variables</p><p>may not be clear, and there could be bias, such as being</p><p>skewed to a certain demographic.</p><p>Artificial Intelligence Basics</p><p>32</p><p>• Third-Party Data: This is data from a commercial vendor.</p><p>But the fees can be high. In fact, the data quality, in some</p><p>cases, may be lacking.</p><p>According to Teradata—based on the firm’s own AI engagements—about</p><p>70% of data sources are in-house, 20% from open source, and the rest from</p><p>commercial vendors.16 But despite the source, all data must be trusted. If not,</p><p>there will likely be the problem of “garbage in, garbage out.”</p><p>To evaluate the data, you need to answer questions like the following:</p><p>• Is the data complete? What might be missing?</p><p>• Where did the data come from?</p><p>• What were the collection points?</p><p>• Who touched the data and processed it?</p><p>• What have been the changes in the data?</p><p>• What are the quality issues?</p><p>If you are working with structured data, then this stage should be easier.</p><p>However, when it comes to unstructured and semi-structured data, you will</p><p>need to label the data—which can be a protracted process. But there are</p><p>some tools emerging in the market that can help automate this process.</p><p>Step #3—Data Preparation</p><p>The first step in the data preparation process is to decide what datasets</p><p>to use.</p><p>Let’s take a look at a scenario: Suppose you work for a publishing company</p><p>and want to put together a strategy to improve customer retention. Some</p><p>of the data that should help would include demographic information on</p><p>the customer base like age, sex, income, and education. To provide more</p><p>color, you can also look at browser information. What type of content</p><p>interests customers? What’s the frequency and duration? Any other</p><p>interesting patterns—say accessing information during weekends? By</p><p>combining the sources of information, you can put together a powerful</p><p>model. For example, if there is a drop-off in activity in certain areas, it</p><p>could pose a risk of cancellation. This would alert sales people to reach</p><p>out to the customers.</p><p>16 This is from the author’s interview with Atif Kureishy in February 2019.</p><p>Chapter 2 | Data</p><p>33</p><p>While this is a smart process, there are still landmines. Including or excluding</p><p>even one variable can have a significant negative impact on an AI model. To</p><p>see why, look back at the financial crisis. The models for underwriting</p><p>mortgages were sophisticated and based on huge amounts of data. During</p><p>normal economic times, they worked quite well as major financial institutions</p><p>like Goldman Sachs, JP Morgan, and AIG relied on them heavily.</p><p>But there was a problem: The models did not account for falling housing</p><p>prices! The main reason was that—for decades—there had never been a</p><p>national drop. The assumption was that housing was mostly a local</p><p>phenomenon.</p><p>Of course, housing prices more than just fell—they plunged. The models then</p><p>proved to be far off the mark, and billions of dollars in losses nearly took</p><p>down the US financial system. The federal</p><p>government had little choice but to</p><p>lend $700 billion for a bailout of Wall Street.</p><p>Granted, this is an extreme case. But it does highlight the importance of data</p><p>selection. This is where having a solid team of domain experts and data</p><p>scientists can be essential.</p><p>Next, when in the data preparation stage, there will need to be data cleansing.</p><p>The fact is that all data has issues. Even companies like Facebook have gaps,</p><p>ambiguities, and outliers in their datasets. It’s inevitable.</p><p>So here are some actions you can take to cleanse the data:</p><p>• De-duplication: Set tests to identify any duplications and</p><p>delete the extraneous data.</p><p>• Outliers: This is data that is well beyond the range of most</p><p>of the rest of the data. This may indicate that the</p><p>information is not helpful. But of course, there are</p><p>situations where the reverse is true. This would be for</p><p>fraud deduction.</p><p>• Consistency: Make sure you have clear definitions for the</p><p>variables. Even terms like “revenue” or “customer” can</p><p>have multiple meanings.</p><p>• Validation Rules: As you look at the data, try to find the</p><p>inherent limitations. For example, you can have a flag for</p><p>the age column. If it is over 120in many cases, then the</p><p>data has some serious issues.</p><p>• Binning: Certain data may not need to be specific. Does it</p><p>really matter if someone is 35 or 37? Probably not. But</p><p>comparing those from 30–40 to 41–50 probably would.</p><p>Artificial Intelligence Basics</p><p>34</p><p>• Staleness: Is the data timely and relevant?</p><p>• Merging: In some cases, the columns of data may have</p><p>very similar information. Perhaps one has height in inches</p><p>and another in feet. If your model does not require a</p><p>more detailed number, you can just use the one for feet.</p><p>• One-Hot Encoding: This is a way to replace categorical</p><p>data as numbers. Example: Let’s say we have a database</p><p>with a column that has three possible values: Apple,</p><p>Pineapple, and Orange. You could represent Apple as 1,</p><p>Pineapple as 2, and Orange as 3. Sounds reasonable,</p><p>right? Perhaps not. The problem is that an AI algorithm</p><p>may think that Orange is greater than Apple. But with</p><p>one-hot encoding, you can avoid this problem. You will</p><p>create three new columns: is_Apple, is_Pineapple, and</p><p>is_Orange. For each row in the data, you’ll put 1 for</p><p>where the fruit exists and 0 for the rest.</p><p>• Conversion Tables: You can use this when translating data</p><p>from one standard to another. This would be the case if</p><p>you have data in the decimal system and want to move</p><p>over to the metric system.</p><p>These steps will go a long way in improving the quality of the data. There are</p><p>also automation tools that can help out, such as from companies like SAS,</p><p>Oracle, IBM, Lavastorm Analytics, and Talend. Then there are open source</p><p>projects, such as OpenRefine, plyr, and reshape2.</p><p>Regardless, the data will not be perfect. No data source is. There will likely</p><p>still be gaps and inaccuracies.</p><p>This is why you need to be creative. Look at what Eyal Lifshitz did, who is the</p><p>CEO of BlueVine. His company leverages AI to provide financing to small</p><p>businesses. “One of our data sources is credit information of our customers,”</p><p>he said. “But we’ve found that small business owners incorrectly identify their</p><p>type of business. This could mean bad results for our underwriting. To deal</p><p>with this, we scrape data from the customer website with AI algorithms,</p><p>which helps identify the industry.”17</p><p>Data cleansing approaches will also depend on the use cases for the AI project.</p><p>For example, if you are building a system for predictive maintenance in</p><p>manufacturing, the challenge will be to handle the wide variations from</p><p>different sensors. The result is that a large amount of the data may have little</p><p>value and be mostly noise.</p><p>17 This is from the author’s interview with Eyal Lifshitz in February 2019.</p><p>Chapter 2 | Data</p><p>35</p><p>Ethics andGovernance</p><p>You need to be mindful of any restrictions on the data. Might the vendor</p><p>prohibit you from using the information for certain purposes? Perhaps your</p><p>company will be on the hook if something goes wrong? To deal with these</p><p>issues, it is advisable to have the legal department brought in.</p><p>For the most part,data must be treated with care. After all, there are many</p><p>high-profile cases where companies have violated privacy. A prominent</p><p>example of this is Facebook. One of the company’s partners, Cambridge</p><p>Analytica, accessed millions of data points from profiles without the permission</p><p>of users. When a whistleblower uncovered this, Facebook stock plunged—</p><p>losing more than $100 billion in value. The company also came under pressure</p><p>from the US and European governments.18</p><p>Something else to be wary of is scraping data from public sources. True, this</p><p>is often an efficient way to create large datasets. There are also many tools</p><p>that can automate the process. But scraping could expose your company to</p><p>legal liability as the data may be subject to copyrights or privacy laws.</p><p>There are also some precautions that may ironically have inherent flaws. For</p><p>example, a recent study from MIT shows that anonymized data may not be</p><p>very anonymized. The researchers found that it was actually quite easy to</p><p>reconstruct this type of data and identify the individuals—such as by merging</p><p>two datasets. This was done by using data in Singapore from a mobile network</p><p>(GPS tracking) and a local transportation system. After about 11 weeks of</p><p>analysis, the researchers were able to identify 95% of the individuals.19</p><p>Finally, make sure you take steps to secure the data. The instances of</p><p>cyberattacks and threats continue to increase at an alarming rate. In 2018,</p><p>there were 53,000+ incidents and about 2,200 breaches, according to</p><p>Verizon.20 The report also noted the following:</p><p>• 76% of the breaches were financially motivated.</p><p>• 73% were from those outside the company.</p><p>• About half came from organized criminal groups and 12%</p><p>from nation-state or state-affiliated actors.</p><p>The increasing use of cloud and on-premise data can subject a company to</p><p>gaps in security as well. Then there is the mobile workforce, which can mean</p><p>access to data that could expose it to breaches.</p><p>18 h t t p s : / / v e n t u r e b e a t . c o m / 2 0 1 8 / 0 7 / 0 2 / u - s - a g e n c i e s - w i d e n -</p><p>investigation-into-what-facebook-knew-about-cambridge-analytica/</p><p>19 http://news.mit.edu/2018/privacy-risks-mobility-data-1207</p><p>20 https://enterprise.verizon.com/resources/reports/dbir/</p><p>Artificial Intelligence Basics</p><p>https://venturebeat.com/2018/07/02/u-s-agencies-widen-investigation-into-what-facebook-knew-about-cambridge-analytica/</p><p>https://venturebeat.com/2018/07/02/u-s-agencies-widen-investigation-into-what-facebook-knew-about-cambridge-analytica/</p><p>http://news.mit.edu/2018/privacy-risks-mobility-data-1207</p><p>https://enterprise.verizon.com/resources/reports/dbir/</p><p>36</p><p>The attacks are also getting much more damaging. The result is that a company</p><p>can easily suffer penalties, lawsuits, and reputational damage.</p><p>Basically, when putting together an AI project, make sure there is a security</p><p>plan and that it is followed.</p><p>How Much Data Do YouNeed forAI?</p><p>The more data, the better, right? This is usually the case. Look at something</p><p>called Hughes Phenomenon. This posits that as you add features to a model,</p><p>the performance generally increases.</p><p>But quantity is not the end-all, be-all. There may come a point where the data</p><p>starts to degrade. Keep in mind that you may run into something called the</p><p>curse of dimensionality. According to Charles Isbell, who is the professor and</p><p>senior associate dean of the School of Interactive Computing at Georgia Tech,</p><p>“As the number of features or dimensions grows, the amount of data we need</p><p>to generalize accurately grows exponentially.”21</p><p>What is the practical impact? It could make it impossible to have a good</p><p>model since there may not be enough data. This is why that when it comes to</p><p>applications like vision recognition, the curse of dimensionality can be quite</p><p>problematic. Even when analyzing</p><p>RGB images, the number of dimensions is</p><p>roughly 7,500. Just imagine how intensive the process would be using real-</p><p>time, high-definition video.</p><p>More Data Terms andConcepts</p><p>When engaging in data analysis, you should know the basic terms. Here are</p><p>some that you’ll often hear:</p><p>Categorical Data: This is data that does not have a numerical meaning. Rather,</p><p>it has a textual meaning like a description of a group (race and gender).</p><p>Although, you can assign numbers to each of the elements.</p><p>Data Type: This is the kind of information a variable represents, such as a</p><p>Boolean, integer, string, or floating point number.</p><p>Descriptive Analytics: This is analyzing data to get a better understanding of the</p><p>current status of a business. Some examples of this include measuring what</p><p>products are selling better or determining risks in customer support. There</p><p>are many traditional software tools for descriptive analytics, such as BI</p><p>applications.</p><p>21 www.kdnuggets.com/2017/04/must-know-curse-dimensionality.html</p><p>Chapter 2 | Data</p><p>http://www.kdnuggets.com/2017/04/must-know-curse-dimensionality.html</p><p>37</p><p>Diagnostic Analytics: This is querying data to see why something has happened.</p><p>This type of analytics uses techniques like data mining, decision trees, and</p><p>correlations.</p><p>ETL (Extraction, Transformation, and Load): This is a form of data integration</p><p>and is typically used in a data warehouse.</p><p>Feature: This is a column of data.</p><p>Instance: This is a row of data.</p><p>Metadata: This is data about data—that is, descriptions. For example, a music file</p><p>can have metadata like the size, length, date of upload, comments, genre, artist,</p><p>and so on. This type of data can wind up being quite useful for an AI project.</p><p>Numerical Data: This is any data that can be represented by a number. But</p><p>numerical data can have two forms. There is discrete data, which is an</p><p>integer—that is, a number without a decimal point. Then there is continuous</p><p>data that has a flow, say temperature or time.</p><p>OLAP (Online Analytical Processing): This is technology that allows you to analyze</p><p>information from various databases.</p><p>Ordinal Data: This is a mix of numerical and categorical data. A common</p><p>example of this is the five-star rating on Amazon.com. It has both a star and</p><p>a number associated with it.</p><p>Predictive Analytics: This involves using data to make forecasts. The models for</p><p>this are usually sophisticated and rely on AI approaches like machine learning.</p><p>To be effective, it is important to update the underlying model with new data.</p><p>Some of the tools for predictive analytics include machine learning approaches</p><p>like regressions.</p><p>Prescriptive Analytics: This is about leveraging Big Data to make better decisions.</p><p>This is not only focued on predicting outcomes—but understanding the</p><p>rationales. And this is where AI plays a big part.</p><p>Scalar Variables: These are variables that hold single values like name or credit</p><p>card number.</p><p>Transactional Data: This is data that is recorded on financial, business, and</p><p>logistical actions. Examples include payments, invoices, and insurance claims.</p><p>Conclusion</p><p>Being successful with AI means having a data-driven culture. This is what has</p><p>been critical for companies like Amazon.com, Google, and Facebook. When</p><p>making decisions, they look to the data first. There should also be wide</p><p>availability of data across the organization.</p><p>Artificial Intelligence Basics</p><p>38</p><p>Without this approach, success with AI will be fleeting, regardless of your</p><p>planning. Perhaps this helps explain that—according to a study from</p><p>NewVantage Partners—about 77% of respondents say that “business</p><p>adoption” of Big Data and AI remain challenges.22</p><p>Key Takeaways</p><p>• Structured data is labeled and formatted—and is often</p><p>stored in a relational database or spreadsheet.</p><p>• Unstructured data is information that has no predefined</p><p>formatting.</p><p>• Semi-structured data has some internal tags that help</p><p>with categorization.</p><p>• Big Data describes a way to handle huge amounts of</p><p>volumes of information.</p><p>• A relational database is based on relationships of data.</p><p>But this structure can prove difficult for modern-day</p><p>applications, such as AI.</p><p>• A NoSQL database is more free-form, being based on a</p><p>document model. This has made it better able to deal</p><p>with unstructured and semi-structured data.</p><p>• The CRISP-DM Process provides a way to manage data</p><p>for a project, with steps that include business</p><p>understanding, data understanding, data preparation,</p><p>modelling, evaluation, and deployment.</p><p>• Quantity of data is certainly important, but there also</p><p>needs to be much work on the quality. Even small errors</p><p>can have a huge impact on the results of an AI model.</p><p>22 http://newvantage.com/wp-content/uploads/2018/12/Big-Data-Executive-</p><p>Survey-2019-Findings-Updated-010219-1.pdf</p><p>Chapter 2 | Data</p><p>http://newvantage.com/wp-content/uploads/2018/12/Big-Data-Executive-Survey-2019-Findings-Updated-010219-1.pdf</p><p>http://newvantage.com/wp-content/uploads/2018/12/Big-Data-Executive-Survey-2019-Findings-Updated-010219-1.pdf</p><p>© Tom Taulli 2019</p><p>T. Taulli, Artif icial Intelligence Basics,</p><p>https://doi.org/10.1007/978-1-4842-5028-0_3</p><p>C H A P T E R</p><p>3</p><p>Machine</p><p>Learning</p><p>Mining Insights from Data</p><p>A breakthrough in machine learning would be worth ten Microsofts.</p><p>—Bill Gates1</p><p>While Katrina Lake liked to shop online, she knew the experience could be</p><p>much better. The main problem: It was tough to find fashions that were</p><p>personalized.</p><p>So began the inspiration for Stitch Fix, which Katrina launched in her</p><p>Cambridge apartment while attending Harvard Business School in 2011 (by</p><p>the way, the original name for the company was the less catchy “Rack Habit”).</p><p>The site had a Q&A for its users—asking about size and fashion styles, just to</p><p>name a few factors—and expert stylists would then put together curated</p><p>boxes of clothing and accessories that were sent out monthly.</p><p>The concept caught on quickly, and the growth was robust. But it was tough to</p><p>raise capital as many venture capitalists did not see the potential in the business.</p><p>Yet Katrina persisted and was able to create a profitable operation—fairly</p><p>quickly.</p><p>1 Steve Lohr, “Microsoft, Amid Dwindling Interest, Talks Up Computing as a Career:</p><p>Enrollment in Computing Is Dwindling,” New York Times, March 1, 2004, start page C1,</p><p>quote page C2, column 6.</p><p>40</p><p>Along the way, Stitch Fix was collecting enormous amounts of valuable data,</p><p>such as on body sizes and style preferences. Katrina realized that this would</p><p>be ideal for machine learning. To leverage on this, she hired Eric Colson, who</p><p>was the vice president of Data Science and Engineering at Netflix, his new</p><p>title being chief algorithms officer.</p><p>This change in strategy was pivotal. The machine learning models got better</p><p>and better with their predictions, as Stitch Fix collected more data—not only</p><p>from the initial surveys but also from ongoing feedback. The data was also</p><p>encoded in the SKUs.</p><p>The result: Stitch Fix saw ongoing improvement in customer loyalty and</p><p>conversion rates. There were also improvements in inventory turnover, which</p><p>helped to reduce costs.</p><p>But the new strategy did not mean firing the stylists. Rather, the machine</p><p>learning greatly augmented their productivity and effectiveness.</p><p>The data also provided insights on what types of clothing to create. This led</p><p>to the launch of Hybrid Designs in 2017, which is Stitch Fix’s private-label</p><p>brand. This proved effective in dealing with the gaps in inventory.</p><p>By November 2017, Katrina took Stitch Fix public, raising $120 million. The</p><p>valuation of the company was a cool $1.63 billion—making her one of the</p><p>richest women in the United States.2 Oh, and at the time, she had a 14-month-</p><p>old son!</p><p>Fast forward to today, Stitch Fix has 2.7 million customers in the United</p><p>States and generates over $1.2 billion in revenues. There are also more than</p><p>100 data scientists on staff and a majority of them have PhDs in areas like</p><p>neuroscience, mathematics, statistics, and AI.3</p><p>According to</p><p>the company’s 10-K filing:</p><p>Our data science capabilities fuel our business. These capabilities consist</p><p>of our rich and growing set of detailed client and merchandise data and</p><p>our proprietary algorithms. We use data science throughout our</p><p>business, including to style our clients, predict purchase behavior,</p><p>forecast demand, optimize inventory and design new apparel.4</p><p>2 www.cnbc.com/2017/11/16/stitch-fix-ipo-sees-orders-coming-in-under-</p><p>range.html</p><p>3 https://investors.stitchfix.com/static-files/2b398694-f553-4586-</p><p>b763-e942617e4dbf</p><p>4 www.sec.gov/Archives/edgar/data/1576942/000157694218000003/</p><p>stitchfix201810k.htm</p><p>Chapter 3 | Machine Learning</p><p>http://www.cnbc.com/2017/11/16/stitch-fix-ipo-sees-orders-coming-in-under-range.html</p><p>http://www.cnbc.com/2017/11/16/stitch-fix-ipo-sees-orders-coming-in-under-range.html</p><p>https://investors.stitchfix.com/static-files/2b398694-f553-4586-b763-e942617e4dbf</p><p>https://investors.stitchfix.com/static-files/2b398694-f553-4586-b763-e942617e4dbf</p><p>http://www.sec.gov/Archives/edgar/data/1576942/000157694218000003/stitchfix201810k.htm</p><p>http://www.sec.gov/Archives/edgar/data/1576942/000157694218000003/stitchfix201810k.htm</p><p>41</p><p>No doubt, the story of Stitch Fix clearly shows the incredible power of</p><p>machine learning and how it can disrupt an industry. In an interview with</p><p>digiday.com, Lake noted:</p><p>Historically, there’s been a gap between what you give to companies and</p><p>how much the experience is improved. Big data is tracking you all over</p><p>the web, and the most benefit you get from that right now is: If you</p><p>clicked on a pair of shoes, you’ll see that pair of shoes again a week from</p><p>now. We’ll see that gap begin to close. Expectations are very different</p><p>around personalization, but importantly, an authentic version of it. Not,</p><p>‘You abandoned your cart and we’re recognizing that.’ It will be genuinely</p><p>recognizing who you are as a unique human. The only way to do this</p><p>scalably is through embracing data science and what you can do through</p><p>innovation.5</p><p>OK then, what is machine learning really about? Why can it be so impactful?</p><p>And what are some of the risks to consider?</p><p>In this chapter, we’ll answer these questions—and more.</p><p>What Is Machine Learning?</p><p>After stints at MIT and Bell Telephone Laboratories, Arthur L.Samuel joined</p><p>IBM in 1949 at the Poughkeepsie Laboratory. His efforts helped boost the</p><p>computing power of the company’s machines, such as with the development</p><p>of the 701 (this was IBM’s first commercialized computer system).</p><p>But he also programmed applications. And there was one that would make</p><p>history—that is, his computer checkers game. It was the first example of a</p><p>machine learning system (Samuel published an influential paper on this in</p><p>19596). IBM CEO Thomas J.Watson, Sr., said that the innovation would add</p><p>15 points to the stock price!7</p><p>Then why was Samuel’s paper so consequential? By looking at checkers, he</p><p>showed how machine learning works—in other words, a computer could</p><p>learn and improve by processing data without having to be explicitly</p><p>programmed. This was possible by leveraging advanced concepts of statistics,</p><p>especially with probability analysis. Thus, a computer could be trained to</p><p>make accurate predictions.</p><p>5 https://digiday.com/marketing/stitch-fix-ceo-katrina-lake-</p><p>predicts-ais-impact-fashion/</p><p>6 Arthur L.Samuel, “Some Studies in Machine Learning Using the Game of Checkers,” in</p><p>Edward A. Feigenbaum and Julian Feldman, eds., Computers and Thought (New York:</p><p>McGraw-Hill, 1983), pp.71–105.</p><p>7 https://history.computer.org/pioneers/samuel.html</p><p>Artificial Intelligence Basics</p><p>https://digiday.com/marketing/stitch-fix-ceo-katrina-lake-predicts-ais-impact-fashion/</p><p>https://digiday.com/marketing/stitch-fix-ceo-katrina-lake-predicts-ais-impact-fashion/</p><p>https://history.computer.org/pioneers/samuel.html</p><p>42</p><p>This was revolutionary as software development, at this time, was mostly</p><p>about a list of commands that followed a workflow of logic.</p><p>To get a sense of how machine learning works, let’s use an example from the</p><p>HBO TV comedy show Silicon Valley. Engineer Jian-Yang was supposed to</p><p>create a Shazam for food. To train the app, he had to provide a massive dataset</p><p>of food pictures. Unfortunately, because of time constraints, the app only</p><p>learned how to identify…hot dogs. In other words, if you used the app, it</p><p>would only respondwith “hot dog” and “not hot dog.”</p><p>While humorous, the episode did a pretty good job of demonstrating machine</p><p>learning. In essence, it is a process of taking in labeled data and finding</p><p>relationships. If you train the system with hot dogs—such as thousands of</p><p>images—it will get better and better at recognizing them.</p><p>Yes, even TV shows can teach valuable lessons about AI!</p><p>But of course, you still need much more. In the next section of the chapter,</p><p>we’ll take a deeper look at the core statistics you need to know about machine</p><p>learning. This includes the standard deviation, the normal distribution, Bayes’</p><p>theorem, correlation, and feature extraction.</p><p>Then we’ll cover topics like the use cases for machine learning, the general</p><p>process, and the common algorithms.</p><p>Standard Deviation</p><p>The standard deviation measures the average distance from the mean. In fact,</p><p>there is no need to learn how to calculate this (the process involves multiple</p><p>steps) since Excel or other software can do this for you easily.</p><p>To understand the standard deviation, let’s take an example of the home</p><p>values in your neighborhood. Suppose that the average is $145,000 and the</p><p>standard deviation is $24,000. This means that one standard deviation below</p><p>the average would be $133,000 ($145,000 – $12,000) and one standard</p><p>deviation above the mean would come to $157,000 ($145,000 + $12,000).</p><p>This gives us a way to quantify the variation in the data. That is, there is a</p><p>spread of $24,000 from the average.</p><p>Next, let’s take a look at the data if, well, Mark Zuckerberg moves into your</p><p>neighborhood and, as a result, the average jumps to $850,000 and the standard</p><p>deviation is $175,000. But do these statistical metrics reflect the valuations?</p><p>Not really. Zuckerberg’s purchase is an outlier. In this situation, the best</p><p>approach may be instead to exclude his home.</p><p>Chapter 3 | Machine Learning</p><p>43</p><p>The Normal Distribution</p><p>When plotted on a graph, the normal distribution looks like a bell (this is why</p><p>another name for it is the “bell curve”). It represents the sum of probabilities</p><p>for a variable. Interestingly enough, the normal curve is common in the natural</p><p>world, as it reflects distributions of such things like height and weight.</p><p>A general approach when interpreting a normal distribution is to use the</p><p>68-95-99.7 rule. This estimates that 68% of the data items will fall within one</p><p>standard deviation, 95% within two standard deviations, and 99.7% within</p><p>three standard deviations.</p><p>A way to understand this is to use IQ scores. Suppose the mean score is 100</p><p>and the standard deviation is 15. We’d have this for the three standard</p><p>deviations, as shown in Figure3-1.</p><p>Note that the peak in this graph is the average. So, if a person has an IQ of</p><p>145, then only 0.15% will have a higher score.</p><p>Now the curve may have different shapes, depending on the variation in the</p><p>data. For example, if our IQ data has a large number of geniuses, then the</p><p>distribution will skew to the right.</p><p>Bayes’ Theorem</p><p>As the name implies, descriptive statistics provides information about your</p><p>data. We’ve already seen this with such things as averages and standard</p><p>deviations.</p><p>But of course, you can go well beyond this—basically, by using the Bayes’</p><p>theorem. This approach is common in analyzing medical diseases, in which</p><p>cause and effect are key—say for FDA (Federal Drug Administration) trials.</p><p>Figure 3-1. Normal distribution of IQ scores</p><p>Artificial Intelligence Basics</p><p>44</p><p>To understand how Bayes’ theorem works, let’s take an example. A researcher</p><p>comes up with a test for a certain type of cancer, and it has proven to be</p><p>accurate 80% of the</p><p>time. This is known as a true positive.</p><p>But 9.6% of the time, the test will identify the person as having the cancer</p><p>even though he or she does not have it, which is known as a false positive.</p><p>Keep in mind that—in some drug tests—this percentage may be higher than</p><p>the accuracy rate!</p><p>And finally, 1% of the population has the cancer.</p><p>In light of all this, if a doctor uses the test on you and it shows that you have</p><p>the cancer, what is the probability that you really have the cancer? Well, Bayes’</p><p>theorem will show the way. This calculation uses factors like accuracy rates,</p><p>false positives, and the population rate to come up with a probability:</p><p>• Step #1: 80% accuracy rate × the chance of having the</p><p>cancer (1%) = 0.008.</p><p>• Step #2: The chance of not having the cancer (99%) × the</p><p>9.6% false positive = 0.09504.</p><p>• Step #3: Then plug the above numbers into the following</p><p>equation: 0.008 / (0.008 + 0.09504) = 7.8%.</p><p>Sounds kind of out of whack, right? Definitely. After all, how is it that a test,</p><p>which is 90% accurate, has only a 7.8% probability of being right? But remember</p><p>the accuracy rate is based on the measure of those who have the flu. And this</p><p>is a small number since only 1% of the population has the flu. What’s more,</p><p>the test is still giving off false positives. So Bayes’ theorem is a way to provide</p><p>a better understanding of results—which is critical for systems like AI.</p><p>Correlation</p><p>A machine learning algorithm often involves some type of correlation among</p><p>the data. A quantitative way to describe this is to use the Pearson correlation,</p><p>which shows the strength of the relationship between two variables that</p><p>range from 1 to –1 (this is the coefficient).</p><p>Here’s how it works:</p><p>• Greater than 0: This is where an increase in one variable</p><p>leads to the increase in another. For example: Suppose</p><p>that there is a 0.9 correlation between income and</p><p>spending. If income increases by $1,000, then spending</p><p>will be up by $900 ($1,000 X 0.9).</p><p>• 0: There is no correlation between the two variables.</p><p>Chapter 3 | Machine Learning</p><p>45</p><p>• Less than 0: Any increase in the variable means a decrease</p><p>in another and vice versa. This describes an inverse</p><p>relationship.</p><p>Then what is a strong correlation? As a general rule of thumb, it’s if the</p><p>coefficient is +0.7 or so. And if it is under 0.3, then the correlation is tenuous.</p><p>All this harkens the old saying of “Correlation is not necessarily causation.”</p><p>Yet when it comes to machine learning, this concept can easily be ignored and</p><p>lead to misleading results.</p><p>For example, there are many correlations that are just random. In fact, some</p><p>can be downright comical. Check out the following from Tylervigen.com:8</p><p>• The divorce rate in Maine has a 99.26% correlation with</p><p>per capita consumption of margarine.</p><p>• The age of Miss America has an 87.01% correlation with</p><p>the murders by steam, hot vapors, and hot tropics.</p><p>• The US crude oil imports from Norway have a 95.4%</p><p>correlation with drivers killed in collision with a railway</p><p>train.</p><p>There is a name for this: patternicity. This is the tendency to find patterns in</p><p>meaningless noise.</p><p>Feature Extraction</p><p>In Chapter 2, we looked at selecting the variables for a model. The process is</p><p>often called feature extraction or feature engineering.</p><p>An example of this would be a computer model that identifies a male or</p><p>female from a photo. For humans, this is fairly easy and quick. It’s something</p><p>that is intuitive. But if someone asked you to describe the differences, would</p><p>you be able to? For most people, it would be a difficult task. However, if we</p><p>want to build an effective machine learning model, we need to get feature</p><p>extraction right—and this can be subjective.</p><p>8 www.tylervigen.com/spurious-correlations</p><p>Artificial Intelligence Basics</p><p>http://www.tylervigen.com/spurious-correlations</p><p>46</p><p>Table 3-1 shows some ideas about how a man’s face may differ from a woman’s.</p><p>This just scratches the surface as I’m sure you have your own ideas or</p><p>approaches. And this is normal. But this is also why such things as facial</p><p>recognition are highly complex and subject to error.</p><p>Feature extraction also has some nuanced issues. One is the potential for</p><p>bias. For example, do you have preconceptions of what a man or woman</p><p>looks like? If so, this can result in models that give wrong results.</p><p>Because of all this, it’s a good idea to have a group of experts who can</p><p>determine the right features. And if the feature engineering proves too</p><p>complex, then machine learning is probably not a good option.</p><p>But there is another approach to consider: deep learning. This involves</p><p>sophisticated models that find features in a data. Actually, this is one of the</p><p>reasons that deep learning has been a major breakthrough in AI.We’ll learn</p><p>more about this in the next chapter.</p><p>What Can YouDo withMachine Learning?</p><p>As machine learning has been around for decades, there have been many uses</p><p>for this powerful technology. It also helps that there are clear benefits, in</p><p>terms of cost savings, revenue opportunities, and risk monitoring.</p><p>To give a sense of the myriad applications, here’s a look at some examples:</p><p>• Predictive Maintenance: This monitors sensors to forecast</p><p>when equipment may fail. This not only helps to reduce</p><p>costs but also lessens downtime and boosts safety. In</p><p>fact, companies like PrecisionHawk are actually using</p><p>drones to collect data, which is much more efficient. The</p><p>technology has proven quite effective for industries like</p><p>energy, agriculture, and construction. Here’s what</p><p>PrecisionHawk notes about its own drone-based</p><p>predictive maintenance system: “One client tested the</p><p>use of visual line of sight (VLOS) drones to inspect a</p><p>Table 3-1. Facial features</p><p>Features Male</p><p>Eyebrows Thicker and straighter</p><p>Face shape Longer and larger, with more of a square shape</p><p>Jawbone Square, wider, and sharper</p><p>Neck Adam’s apple</p><p>Chapter 3 | Machine Learning</p><p>47</p><p>cluster of 10 well pads in a three-mile radius. Our client</p><p>determined that the use of drones reduced inspection</p><p>costs by approximately 66%, from $80–$90 per well pad</p><p>from traditional inspection methodology to $45–$60 per</p><p>well pad using VLOS drone missions.”9</p><p>• Recruiting Employees: This can be a tedious process since</p><p>many resumes are often varied. This means it is easy to</p><p>pass over great candidates. But machine learning can help</p><p>in a big way. Take a look at CareerBuilder, which has</p><p>collected and analyzed more than 2.3 million jobs, 680</p><p>million unique profiles, 310 million unique resumes, 10</p><p>million job titles, 1.3 billion skills, and 2.5 million</p><p>background checks to build Hello to Hire. It’s a platform</p><p>that has leveraged machine learning to reduce the number</p><p>of job applications—for a successful hire—to an average</p><p>of 75. The industry average, on the other hand, is about</p><p>150.10 The system also automates the creation of job</p><p>descriptions, which even takes into account nuances</p><p>based on the industry and location!</p><p>• Customer Experience: Nowadays, customers want a</p><p>personalized experience. They have become accustomed</p><p>to this by using services like Amazon.com and Uber. With</p><p>machine learning, a company can leverage its data to gain</p><p>insight—learning about what really works. This is so</p><p>important that it led Kroger to buy a company in the</p><p>space, called 84.51°. It is definitely key that it has data on</p><p>more than 60 million US households. Here’s a quick case</p><p>study: For most of its stores, Kroger had bulk avocados,</p><p>and only a few carried 4-packs. The conventional wisdom</p><p>was that 4-packs had to be discounted because of the size</p><p>disparity with the bulk items. But when applying machine</p><p>learning analysis, this proved to be incorrect, as the 4-packs</p><p>attracted new and different households like Millennials</p><p>and ClickList shoppers. By expanding 4-packs across the</p><p>chain, there was an overall increase in avocado sales.11</p><p>• Finance: Machine learning can detect discrepancies, say</p><p>with billing. But there is a new category of technology,</p><p>called RPA (Robotic</p><p>Process Automation), that can help</p><p>9 www.precisionhawk.com/blog/in-oil-gas-the-economics-of-bvlos-drone-</p><p>operations</p><p>10 This information is from the author’s interview in February 2019 with Humair Ghauri,</p><p>who is the chief product officer at CareerBuilder.</p><p>11 www.8451.com/case-study/avocado</p><p>Artificial Intelligence Basics</p><p>http://www.precisionhawk.com/blog/in-oil-gas-the-economics-of-bvlos-drone-operations</p><p>http://www.precisionhawk.com/blog/in-oil-gas-the-economics-of-bvlos-drone-operations</p><p>http://www.8451.com/case-study/avocado</p><p>48</p><p>with this (we’ll cover this topic in Chapter5). It automates</p><p>routine processes in order to help reduce errors. RPA</p><p>also may use machine learning to detect abnormal or</p><p>suspicious transactions.</p><p>• Customer Service: The past few years has seen the growth</p><p>in chatbots, which use machine learning to automate</p><p>interactions with customers. We’ll cover this in</p><p>Chapter6.</p><p>• Dating: Machine learning could help find your soul mate!</p><p>Tinder, one of the largest dating apps, is using the</p><p>technology to help improve the matches. For instance,</p><p>it has a system that automatically labels more than 10</p><p>billion photos that are uploaded on a daily basis.</p><p>Figure 3-2 shows some of the applications for machine learning.</p><p>The Machine Learning Process</p><p>To be successful with applying machine learning to a problem, it’s important</p><p>to take a systematic approach. If not, the results could be way off base.</p><p>First of all, you need to go through a data process, which we covered in the</p><p>prior chapter. When this is finished, it’s a good idea to do a visualization of the</p><p>data. Is it mostly scattered? Or are there some patterns? If the answer is yes,</p><p>then the data could be a good candidate for machine learning.</p><p>Figure 3-2. Applications for machine learning</p><p>Chapter 3 | Machine Learning</p><p>49</p><p>The goal of the machine learning process is to create a model, which is based</p><p>on one or more algorithms. We develop this by training it. The goal is that the</p><p>model should provide a high-degree of predictability.</p><p>Now let’s take a closer look at this (by the way, this will also be applicable for</p><p>deep learning, which we’ll cover in the next chapter):</p><p>Step #1—Data Order</p><p>If your data is sorted, then this could skew the results. That is, the machine</p><p>learning algorithm may detect this as a pattern! This is why it’s a good idea to</p><p>randomize the order of the data.</p><p>Step #2—Choose aModel</p><p>You will need to select an algorithm. This will be an educated guess, which will</p><p>involve a process of trial and error. In this chapter, we’ll look at the various</p><p>algorithms available.</p><p>Step #3—Train theModel</p><p>The training data, which will be about 70% of the complete dataset, will be</p><p>used to create the relationships in the algorithm. For example, suppose you</p><p>are building a machine learning system to find the value of a used car. Some of</p><p>the features will include the year manufactured, make, model, mileage, and</p><p>condition. By processing this training data, the algorithm will calculate the</p><p>weights for each of these factors.</p><p>Example: Suppose we are using a linear regression algorithm, which has the</p><p>following format:</p><p>y = m * x + b</p><p>In the training phase, the system will come up with the values for m (which is</p><p>the slope on a graph) and b (which is the y-intercept).</p><p>Step #4—Evaluate theModel</p><p>You will put together test data, which is the remaining 30% of the dataset. It</p><p>should be representative of the ranges and type of information in the training</p><p>data.</p><p>With the test data, you can see if the algorithm is accurate. In our used car</p><p>example, are the market values consistent with what’s happening in the real</p><p>world?</p><p>Artificial Intelligence Basics</p><p>50</p><p>■ Note With the training and test data, there must not be any intermingling. This can easily lead</p><p>to distorted results. Interestingly enough, this is a common mistake.</p><p>Now accuracy is one measure of the success of the algorithm. But this can, in</p><p>some cases, be misleading. Consider the situation with fraud deduction. There</p><p>are usually a small number of features when compared to a dataset. But missing</p><p>one could be devastating, costing a company millions of dollars in losses.</p><p>This is why you might want to use other approaches like Bayes’ theorem.</p><p>Step #5—Fine-Tune theModel</p><p>In this step, we can adjust the values of the parameters in the algorithm. This</p><p>is to see if we can get better results.</p><p>When fine-tuning the model, there may also be hyperparameters. These are</p><p>parameters that cannot be learned directly from the training process.</p><p>Applying Algorithms</p><p>Some algorithms are quite easy to calculate, while others require complex</p><p>steps and mathematics. The good news is that you usually do not have to</p><p>compute an algorithm because there are a variety of languages like Python and</p><p>R that make the process straightforward.</p><p>As for machine learning, an algorithm is typically different from a traditional</p><p>one. The reason is that the first step is to process data—and then, the</p><p>computer will start to learn.</p><p>Even though there are hundreds of machine learning algorithms available, they</p><p>can actually be divided into four major categories: supervised learning,</p><p>unsupervised learning, reinforcement learning, and semi-supervised learning.</p><p>We’ll take a look at each.</p><p>Supervised Learning</p><p>Supervised learning uses labeled data. For example, suppose we have a set of</p><p>photos of thousands of dogs. The data is considered to be labeled if each</p><p>photo identifies each for the breed. For the most part, this makes it easier to</p><p>analyze since we can compare our results with the correct answer.</p><p>One of the keys with supervised learning is that there should be large amounts</p><p>of data. This helps to refine the model and produce more accurate results.</p><p>Chapter 3 | Machine Learning</p><p>51</p><p>But there is a big issue: The reality is that much of the data available is not</p><p>labeled. In addition, it could be time consuming to provide labels if there is a</p><p>massive dataset.</p><p>Yet there are creative ways to deal with this, such as with crowdfunding. This</p><p>is how the ImageNet system was built, which was a breakthrough in AI</p><p>innovation. But it still took several years to create it.</p><p>Or, in some cases, there can be automated approaches to label data. Take the</p><p>example of Facebook. In 2018, the company announced—at its F8 developers</p><p>conference—it leveraged its enormous database of photos from Instagram,</p><p>which were labeled with hashtags.12</p><p>Granted, this approach had its flaws. A hashtag may give a nonvisual description</p><p>of the photo—say #tbt (which is “throwback Thursday)—or could be too vague,</p><p>like #party. This is why Facebook called its approach “weakly supervised data.”</p><p>But the talented engineers at the company found some ways to improve the</p><p>quality, such as by building a sophisticated hashtag prediction model.</p><p>All in all, things worked out quite well. Facebook’s machine learning model,</p><p>which included 3.5 billion photos, had an accuracy rate of 85.4%, which was</p><p>based on the ImageNet recognition benchmark. It was actually the highest</p><p>recorded in history, by 2%.</p><p>This AI project also required innovative approaches for building the</p><p>infrastructure. According to the Facebook blog post:</p><p>Since a single machine would have taken more than a year to complete</p><p>the model training, we created a way to distribute the task across up to</p><p>336 GPUs, shortening the total training time to just a few weeks. With</p><p>ever-larger model sizes—the biggest in this research is a ResNeXt 101-</p><p>32x48d with over 861 million parameters—such distributed training is</p><p>increasingly essential. In addition, we designed a method for removing</p><p>duplicates to ensure we don’t accidentally train our models on images</p><p>that we want to evaluate them on, a problem that plagues similar</p><p>research in this area.13</p><p>Going forward, Facebook sees potential in using its approach to various areas,</p><p>including the following:</p><p>• Improved ranking in the newsfeed</p><p>• Better detection of objectionable content</p><p>• Auto generation of captions for the visually</p><p>impaired</p><p>12 www.engadget.com/2018/05/02/facebook-trained-image-recognition-ai-</p><p>instagram-pics/</p><p>13 https://code.fb.com/ml-applications/advancing-state-of-the-art-image-</p><p>recognition-with-deep-learning-on-hashtags/</p><p>Artificial Intelligence Basics</p><p>http://www.engadget.com/2018/05/02/facebook-trained-image-recognition-ai-instagram-pics/</p><p>http://www.engadget.com/2018/05/02/facebook-trained-image-recognition-ai-instagram-pics/</p><p>https://code.fb.com/ml-applications/advancing-state-of-the-art-image-recognition-with-deep-learning-on-hashtags/</p><p>https://code.fb.com/ml-applications/advancing-state-of-the-art-image-recognition-with-deep-learning-on-hashtags/</p><p>52</p><p>Unsupervised Learning</p><p>Unsupervised learning is when you are working with unlabeled data. This</p><p>means that you will use deep learning algorithms to detect patterns.</p><p>By far, the most common approach for unsupervised learning is clustering,</p><p>which takes unlabeled data and uses algorithms to put similar items into</p><p>groups. The process usually starts with guesses, and then there are iterations</p><p>of the calculations to get better results. At the heart of this is finding data</p><p>items that are close together, which can be accomplished with a variety of</p><p>quantitative methods:</p><p>• Euclidean Metric: This is a straight line between two data</p><p>points. The Euclidean metric is quite common with</p><p>machine learning.</p><p>• Cosine Similarity Metric: As the name implies, you will use</p><p>a cosine to measure the angle. The idea is to find</p><p>similarities between two data points in terms of the</p><p>orientation.</p><p>• Manhattan Metric: This involves taking the sum of the</p><p>absolute distances of two points on the coordinates of</p><p>a graph. It’s called the “Manhattan” because it references</p><p>the city’s street layout, which allows for shorter distances</p><p>for travel.</p><p>In terms of use cases for clustering, one of the most common is customer</p><p>segmentation, which is to help better target marketing messages. For the</p><p>most part, a group that has similar characteristics is likely to share interests</p><p>and preferences.</p><p>Another application is sentiment analysis, which is where you mine social</p><p>media data and find the trends. For a fashion company, this can be crucial in</p><p>understanding how to adapt the styles of the upcoming line of clothes.</p><p>Now there are other approaches than just clustering. Here’s a look at three</p><p>more:</p><p>• Association: The basic concept is that if X happens, then Y</p><p>is likely to happen. Thus, if you buy my book on AI, you</p><p>will probably want to buy other titles in the genre. With</p><p>association, a deep learning algorithm can decipher these</p><p>kinds of relationships. This can result in powerful</p><p>recommendation engines.</p><p>Chapter 3 | Machine Learning</p><p>53</p><p>• Anomaly Detection: This identifies outliers or anomalous</p><p>patterns in the dataset, which can be helpful with</p><p>cybersecurity applications. According to Asaf Cidon, who</p><p>is the VP of Email Security at Barracuda Networks:</p><p>“We’ve found that by combining many different signals—</p><p>such as the email body, header, the social graph of</p><p>communications, IP logins, inbox forwarding rules, etc.—</p><p>we’re able to achieve an extremely high precision in</p><p>detecting social engineering attacks, even though the</p><p>attacks are highly personalized and crafted to target a</p><p>particular person within a particular organization. Machine</p><p>learning enables us to detect attacks that originate from</p><p>within the organization, whose source is a legitimate</p><p>mailbox of an employee, which would be impossible to do</p><p>with a static one-size-fits-all rule engine.”14</p><p>• Autoencoders: With this, the data will be put into a</p><p>compressed form, and then it will be reconstructed.</p><p>From this, new patterns may emerge. However, the use of</p><p>autoencoders is rare. But it could be shown to be useful</p><p>in helping with applications like reducing noise in data.</p><p>Consider that many AI researchers believe that unsupervised learning will</p><p>likely be critical for the next level of achievements. According to a paper in</p><p>Nature by Yann LeCun, Geoffrey Hinton, and Yoshua Bengio, “We expect</p><p>unsupervised learning to become far more important in the longer term.</p><p>Human and animal learning is largely unsupervised: we discover the structure</p><p>of the world by observing it, not by being told the name of every object.”15</p><p>Reinforcement Learning</p><p>When you were a kid and wanted to play a new sport, chances were you did</p><p>not read a manual. Instead, you observed what other people were doing and</p><p>tried to figure things out. In some situations, you made mistakes and lost the</p><p>ball as your teammates would show their displeasure. But in other cases, you</p><p>made the right moves and scored. Through this trial-and-error process, your</p><p>learning was improved based on positive and negative reinforcement.</p><p>14 This is from the author’s interview in February 2019 with Asaf Cidon, who is the VP of</p><p>Email Security at Barracuda Networks.</p><p>15 https://towardsdatascience.com/simple-explanation-of-semi-supervised-</p><p>learning-and-pseudo-labeling-c2218e8c769b</p><p>Artificial Intelligence Basics</p><p>https://towardsdatascience.com/simple-explanation-of-semi-supervised-learning-and-pseudo-labeling-c2218e8c769b</p><p>https://towardsdatascience.com/simple-explanation-of-semi-supervised-learning-and-pseudo-labeling-c2218e8c769b</p><p>54</p><p>At a high level, this is analogous to reinforcement learning. It has been key for</p><p>some of the most notable achievements in AI, such as the following:</p><p>• Games: They are ideal for reinforcement learning since</p><p>there are clear-cut rules, scores, and various constraints</p><p>(like a game board). When building a model, you can test</p><p>it with millions of simulations, which means that the</p><p>system will quickly get smarter and smarter. This is how</p><p>a program can learn to beat the world champion of Go</p><p>or chess.</p><p>• Robotics: A key is being able to navigate within a space—</p><p>and this requires evaluating the environment at many</p><p>different points. If the robot wants to move to, say, the</p><p>kitchen, it will need to navigate around furniture and</p><p>other obstacles. If it runs into things, there will be a</p><p>negative reinforcement action.</p><p>Semi-supervised Learning</p><p>This is a mix of supervised and unsupervised learning. This arises when you</p><p>have a small amount of unlabeled data. But you can use deep learning systems</p><p>to translate the unsupervised data to supervised data—a process that is called</p><p>pseudo-labeling. After this, you can then apply the algorithms.</p><p>An interesting use case of semi-supervised learning is the interpretation of</p><p>MRIs. A radiologist can first label the scans, and after this, a deep learning</p><p>system can find the rest of the patterns.</p><p>Common Types ofMachine Learning</p><p>Algorithms</p><p>There is simply not enough room in this book to cover all the machine learning</p><p>algorithms! Instead, it’s better to focus on the most common ones.</p><p>In the remaining part of this chapter, we’ll take a look at those for the following:</p><p>• Supervised Learning: You can boil down the algorithms to</p><p>two variations. One is classification, which divides the</p><p>dataset into common labels. Examples of the algorithms</p><p>include Naive Bayes Classifier and k-Nearest Neighbor</p><p>(neural networks will be covered in Chapter 4). Next,</p><p>there is regression, which finds continuous patterns in</p><p>the data. For this, we’ll take a look at linear regression,</p><p>ensemble modelling, and decision trees.</p><p>Chapter 3 | Machine Learning</p><p>55</p><p>• Unsupervised Learning: In this category, we’ll look at</p><p>clustering. For this, we’ll cover k-Means clustering.</p><p>Figure 3-3 shows a general framework for machine learning algorithms.</p><p>Naïve Bayes Classifier (Supervised</p><p>Learning/Classification)</p><p>Earlier in this chapter, we looked at Bayes’ theorem. As for machine learning,</p><p>this has been modified into something called the Naïve Bayes Classifier. It is</p><p>“naïve” because the assumption is that the variables are independent from</p><p>each other—that is, the occurrence of one variable has nothing to do with</p><p>the others. True, this may seem like a drawback. But the fact is that the Naïve</p><p>Bayes Classifier has proven to be quite</p><p>effective and fast to develop.</p><p>There is another assumption to note as well: the a priori assumption. This</p><p>says that the predictions will be wrong if the data has changed.</p><p>There are three variations on the Naïve Bayes Classifier:</p><p>• Bernoulli: This is if you have binary data (true/false, yes/no).</p><p>• Multinomial: This is if the data is discrete, such as the</p><p>number of pages of a book.</p><p>• Gaussian: This is if you are working with data that</p><p>conforms to a normal distribution.</p><p>Figure 3-3. General framework for machine learning algorithms</p><p>Artificial Intelligence Basics</p><p>56</p><p>A common use case for Naïve Bayes Classifiers is text analysis. Examples</p><p>include email spam detection, customer segmentation, sentiment analysis,</p><p>medical diagnosis, and weather predictions. The reason is that this approach</p><p>is useful in classifying data based on key features and patterns.</p><p>To see how this is done, let’s take an example: Suppose you run an e-commerce</p><p>site and have a large database of customer transactions. You want to see how</p><p>variables like product review ratings, discounts, and time of year impact sales.</p><p>Table 3-2 shows a look at the dataset.</p><p>You will then organize this data into frequency tables, as shown in Tables 3-3</p><p>and 3-4.</p><p>Table 3-2. Customer transactions dataset</p><p>Discount Product Review Purchase</p><p>Yes High Yes</p><p>Yes Low Yes</p><p>No Low No</p><p>No Low No</p><p>No Low No</p><p>No High Yes</p><p>Yes High No</p><p>Yes Low Yes</p><p>No High Yes</p><p>Yes High Yes</p><p>No High No</p><p>No Low Yes</p><p>Yes High Yes</p><p>Yes Low No</p><p>Table 3-3. Discount frequency table</p><p>Purchase</p><p>Yes No</p><p>Discount Yes 19 1</p><p>Yes 5 5</p><p>Chapter 3 | Machine Learning</p><p>57</p><p>When looking at this, we call the purchase an event and the discount and</p><p>product reviews as independent variables. Then we can make a probability</p><p>table for one of the independent variables, say the product reviews. See</p><p>Table3-5.</p><p>Using this chart, we can see that the probability of a purchase when there is</p><p>a low product review is 7/24 or 29%. In other words, the Naïve Bayes Classifier</p><p>allows more granular predictions within a dataset. It is also relatively easy to</p><p>train and can work well with small datasets.</p><p>K-Nearest Neighbor (Supervised</p><p>Learning/Classification)</p><p>The k-Nearest Neighbor (k-NN) is a method for classifying a dataset (k</p><p>represents the number of neighbors). The theory is that those values that are</p><p>close together are likely to be good predictors for a model. Think of it as</p><p>“Birds of a feather flock together.”</p><p>A use case for k-NN is the credit score, which is based on a variety of factors</p><p>like income, payment histories, location, home ownership, and so on. The</p><p>algorithm will divide the dataset into different segments of customers. Then,</p><p>when there is a new customer added to the base, you will see what cluster he</p><p>or she falls into—and this will be the credit score.</p><p>K-NN is actually simple to calculate. In fact, it is called lazy learning because</p><p>there is no training process with the data.</p><p>Table 3-4. Product review frequency table</p><p>Purchase</p><p>Yes No Total</p><p>Product Review High 21 2 11</p><p>Low 3 4 8</p><p>Total 24 6 19</p><p>Table 3-5. Product review probability table</p><p>Purchase</p><p>Yes No</p><p>Product Reviews High 9/24 2/6 11/30</p><p>Low 7/24 1/6 8/30</p><p>24/30 6/30</p><p>Artificial Intelligence Basics</p><p>58</p><p>To use k-NN, you need to come up with the distance between the nearest</p><p>values. If the values are numerical, it could be based on a Euclidian distance,</p><p>which involves complicated math. Or, if there is categorical data, then you can</p><p>use an overlap metric (this is where the data is the same or very similar).</p><p>Next, you’ll need to identify the number of neighbors. While having more will</p><p>smooth the model, it can also mean a need for huge amount of computational</p><p>resources. To manage this, you can assign higher weights to data that are</p><p>closer to their neighbors.</p><p>Linear Regression (Supervised</p><p>Learning/Regression)</p><p>Linear regression shows the relationship between certain variables. The</p><p>equation—assuming there is enough quality data—can help predict outcomes</p><p>based on inputs.</p><p>Example: Suppose we have data on the number of hours spent studying for an</p><p>exam and the grade. See Table3-6.</p><p>As you can see, the general relationship is positive (this describes the tendency</p><p>where a higher grade is correlated with more hours of study). With the</p><p>regression algorithm, we can plot a line that has the best fit (this is done by</p><p>using a calculation called “least squares,” which minimizes the errors). See</p><p>Figure3-4.</p><p>Table 3-6. Chart for hours of study and grades</p><p>Hours of Study Grade Percentage</p><p>1 0.75</p><p>1 0.69</p><p>1 0.71</p><p>3 0.82</p><p>3 0.83</p><p>4 0.86</p><p>5 0.85</p><p>5 0.89</p><p>5 0.84</p><p>6 0.91</p><p>6 0.92</p><p>7 0.95</p><p>Chapter 3 | Machine Learning</p><p>59</p><p>From this, we get the following equation:</p><p>Grade = Number of hours of study × 0.03731 + 0.6889</p><p>Then, let’s suppose you study 4 hours for the exam. What will be your</p><p>estimated grade? The equation tells us how:</p><p>0.838 = 4 × 0.03731 + 0.6889</p><p>How accurate is this? To help answer this question, we can use a calculation</p><p>called R-squared. In our case, it is 0.9180 (this ranges from 0 to 1). The closer</p><p>the value is to 1, the better the fit. So 0.9180 is quite high. It means that the</p><p>hours of study explains 91.8% of the grade on the exam.</p><p>Now it’s true that this model is simplistic. To better reflect reality, you can</p><p>add more variables to explain the grade on the exam—say the student’s</p><p>attendance. When doing this, you will use something called multivariate</p><p>regression.</p><p>■ Note If the coefficient for a variable is quite small, then it might be a good idea to not include</p><p>it in the model.</p><p>Sometimes data may not be in a straight line either, in which case the regression</p><p>algorithm will not work. But you can use a more complex version, called</p><p>polynomial regression.</p><p>Figure 3-4. This is a plot of a linear regression model that is based on hours of study</p><p>Artificial Intelligence Basics</p><p>60</p><p>Decision Tree (Supervised</p><p>Learning/Regression)</p><p>No doubt, clustering may not work on some datasets. But the good news is</p><p>that there are alternatives, such as a decision tree. This approach generally</p><p>works better with nonnumerical data.</p><p>The start of a decision tree is the root node, which is at the top of the flow</p><p>chart. From this point, there will be a tree of decision paths, which are</p><p>called splits. At these points, you will use an algorithm to make a decision,</p><p>and there will be a probability computed. At the end of the tree will be the</p><p>leaf (or the outcome).</p><p>A famous example—in machine learning circles—is to use a decision tree for</p><p>the tragic sinking of the Titanic. The model predicts the survival of a passenger</p><p>based on three features: sex, age, and the number of spouses or children</p><p>along (sibsp). Here’s how it looks, in Figure3-5.</p><p>There are clear advantages for decision trees. They are easy to understand,</p><p>work well with large datasets, and provide transparency with the model.</p><p>However, decision trees also have drawbacks. One is error propagation. If one</p><p>of the splits turns out to be wrong, then this error can cascade throughout</p><p>the rest of the model!</p><p>Figure 3-5. This is a basic decision tree algorithm for predicting the survival of the Titanic</p><p>Chapter 3 | Machine Learning</p><p>61</p><p>Next, as the decision trees grow, there will be more complexity as there will</p><p>be a large number of algorithms. This could ultimately result in lower</p><p>performance for the model.</p><p>Ensemble Modelling (Supervised</p><p>Learning/Regression)</p><p>Ensemble modelling means using more than one model for your predictions.</p><p>Even though this increases the complexity, this approach has been shown to</p><p>generate strong results.</p><p>To see this in action, take a look at the “Netflix Prize,” which began in 2006.</p><p>The company announced it would pay $1 million to anyone or any team that</p><p>could improve the accuracy of its movie recommendation system by 10% or</p><p>more. Netflix also provided a dataset of over 100 million ratings of 17,770</p><p>movies from 480,189 users.16 There would ultimately be more than 30,000</p><p>downloads.</p><p>Why did Netflix do all this? A big</p><p>reason is that the company’s own engineers</p><p>were having trouble making progress. Then why not give it to the crowd to</p><p>figure out? It turned out to be quite ingenious—and the $1 million payout was</p><p>really modest compared to the potential benefits.</p><p>The contest certainly stirred up a lot of activity from coders and data</p><p>scientists, ranging from students to employees at companies like AT&T.</p><p>Netflix also made the contest simple. The main requirement was that the</p><p>teams had to disclose their methods, which helped boost the results (there</p><p>was even a dashboard with rankings of the teams).</p><p>But it was not until 2009 that a team—BellKor’s Pragmatic Chaos—won the</p><p>prize. Then again, there were considerable challenges.</p><p>So how did the winning team pull it off? The first step was to create a baseline</p><p>model that smoothed out the tricky issues with the data. For example, some</p><p>movies only had a handful of ratings, whereas others had thousands. Then</p><p>there was the thorny problem where there were users who would always rate</p><p>a movie with one star. To deal with these matters, BellKor used machine</p><p>learning to predict ratings in order to fill the gaps.</p><p>16 www.thrillist.com/entertainment/nation/the-netflix-prize</p><p>Artificial Intelligence Basics</p><p>http://www.thrillist.com/entertainment/nation/the-netflix-prize</p><p>62</p><p>Once the baseline was finished, there were more tough challenges to tackle</p><p>like the following:</p><p>• A system may wind up recommending the same films to</p><p>many users.</p><p>• Some movies may not fit well within genres. For example,</p><p>Alien is really a cross of science fiction and horror.</p><p>• There were movies, like Napoleon Dynamite, that proved</p><p>extremely difficult for algorithms to understand.</p><p>• Ratings of a movie would often change over time.</p><p>The winning team used ensemble modelling, which involved hundreds of</p><p>algorithms. They also used something called boosting, which is where you</p><p>build consecutive models. With this, the weights in the algorithms are adjusted</p><p>based on the results of the previous model, which help the predictions get</p><p>better over time (another approach, called bagging, is when you build different</p><p>models in parallel and then select the best one).</p><p>But in the end, BellKor found the solutions. However, despite this, Netflix did</p><p>not use the model! Now it’s not clear why this was the case. Perhaps it was</p><p>that Netflix was moving away from five-star ratings anyway and was more</p><p>focused on streaming. The contest also had blowback from people who</p><p>thought there may have been privacy violations.</p><p>Regardless, the contest did highlight the power of machine learning—and the</p><p>importance of collaboration.</p><p>K-Means Clustering (Unsupervised/Clustering)</p><p>The k-Means clustering algorithm, which is effective for large datasets, puts</p><p>similar, unlabeled data into different groups. The first step is to select k, which</p><p>is the number of clusters. To help with this, you can perform visualizations of</p><p>that data to see if there are noticeable grouping areas.</p><p>Here’s a look at sample data, in Figure3-6:</p><p>Figure 3-6. The initial plot for a dataset</p><p>Chapter 3 | Machine Learning</p><p>63</p><p>For this example, we assume there will be two clusters, and this means there</p><p>will also be two centroids. A centroid is the midpoint of a cluster. We will</p><p>assign each randomly, which you can see in Figure3-7.</p><p>As you can see, the centroid at the top left looks way off, but the one on the</p><p>right side is better. The k-Means algorithm will then calculate the average</p><p>distances of the centroids and then change their locations. This will be iterated</p><p>until the errors are fairly minimal—a point that is called convergence, which</p><p>you can see with Figure3-8.</p><p>Granted, this is a simple illustration. But of course, with a complex dataset, it</p><p>will be difficult to come up with the number of initial clusters. In this situation,</p><p>you can experiment with different k values and then measure the average</p><p>distances. By doing this multiple times, there should be more accuracy.</p><p>Then why not just have a high number for k? You can certainly do this. But</p><p>when you compute the average, you’ll notice that there will be only incremental</p><p>improvements. So one method is to stop at the point where this starts to</p><p>occur. This is seen in Figure3-9.</p><p>Figure 3-7. This chart shows two centroids—represented by circles—that are randomly placed</p><p>Figure 3-8. Through iterations, the k-Means algorithm gets better at grouping the data</p><p>Artificial Intelligence Basics</p><p>64</p><p>However, k-Means has its drawbacks. For instance, it does not work well with</p><p>nonspherical data, which is the case with Figure3-10.</p><p>Figure 3-9. This shows the optimal point of the k value in the k-Means algorithm</p><p>Figure 3-10. Here’s a demonstration where k-Means does not work with nonspherical data</p><p>With this, the k-Means algorithm would likely not pick up on the surrounding</p><p>data, even though it has a pattern. But there are some algorithms that can</p><p>help, such as DBScan (density-based spatial clustering of applications with</p><p>Chapter 3 | Machine Learning</p><p>65</p><p>noise), which is meant to handle a mix of widely varying sizes of datasets.</p><p>Although, DBScan can require lots of computational power.</p><p>Next, there is the situation where there are some clusters with lots of data</p><p>and others with little. What might happen? There is a chance that the k-Means</p><p>algorithm will not pick up on the light one. This is the case with Figure3-11.</p><p>Conclusion</p><p>These algorithms can get complicated and do require strong technical skills.</p><p>But it is important to not get too bogged down in the technology. After all,</p><p>the focus is to find ways to use machine learning to accomplish clear objectives.</p><p>Again, Stich Fix is a good place to get guidance on this. In the November issue</p><p>of the Harvard Business Review, the company’s chief algorithms officer, Eric</p><p>Colson, published an article, “Curiosity-Driven Data Science.”17 In it, he</p><p>provided his experiences in creating a data-driven organization.</p><p>At the heart of this is allowing data scientists to explore new ideas, concepts,</p><p>and approaches. This has resulted in AI being implemented across core</p><p>functions of the business like inventory management, relationship management,</p><p>logistics, and merchandise buying. It has been transformative, making the</p><p>17 https://hbr.org/2018/11/curiosity-driven-data-science</p><p>Figure 3-11. If there are areas of thin data, the k-Means algorithm may not pick them up</p><p>Artificial Intelligence Basics</p><p>https://hbr.org/2018/11/curiosity-driven-data-science</p><p>66</p><p>organization more agile and streamlined. Colson also believes it has provided</p><p>“a protective barrier against competition.”</p><p>His article also provides other helpful advice for data analysis:</p><p>• Data Scientists: They should not be part of another</p><p>department. Rather, they should have their own, which</p><p>reports directly to the CEO.This helps with focusing on</p><p>key priorities as well as having a holistic view of the needs</p><p>of the organization.</p><p>• Experiments: When a data scientist has a new idea, it</p><p>should be tested on a small sample of customers. If</p><p>there is traction, then it can be rolled out to the rest of</p><p>the base.</p><p>• Resources: Data scientists need full access to data and</p><p>tools. There should also be ongoing training.</p><p>• Generalists: Hire data scientists who span different</p><p>domains like modelling, machine learning, and analytics</p><p>(Colson refers to these people as “full-stack data</p><p>scientists”). This leads to small teams—which are often</p><p>more efficient and productive.</p><p>• Culture: Colson looks for values like “learning by doing,</p><p>being comfortable with ambiguity, balancing long-and</p><p>short-term returns.”</p><p>Key Takeaways</p><p>• Machine learning, whose roots go back to the 1950s, is</p><p>where a computer can learn without being explicitly</p><p>programmed. Rather, it will ingest and process data by</p><p>using sophisticated statistical techniques.</p><p>• An outlier is data that is far outside the rest of the</p><p>numbers in the dataset.</p><p>• The standard deviation measures the average distance</p><p>from the mean.</p><p>• The</p><p>the value of our products without increasing the price by tapping</p><p>into the scalability of cloud technology. Our goal is to empower people at all</p><p>levels of society by pushing down the price of business software while expand-</p><p>ing the power of the tools. Access to capital shouldn't limit success; busi-</p><p>nesses should rise or fall based on the strength of their vision for the future.</p><p>Viewed this way, AI is the fulfillment of the promise of technology. It frees</p><p>people from the constraints of time by enabling them to offload tedious or</p><p>unpleasant rote labor. It helps them identify patterns at microscopic and mac-</p><p>roscopic scales, which humans are not naturally well suited to perceive. It can</p><p>forecast problems, and it can correct errors. It can save money, time, and even</p><p>lives.</p><p>Seeking to democratize these benefits just as we did for general business</p><p>software, Zoho has threaded AI throughout our suite of apps. We spent the</p><p>last six years quietly developing our own internal AI technology, built on the</p><p>bedrock of our own principles. The result is Zia, an AI assistant who is smart,</p><p>but not clever. This is a crucial distinction. A smart system has the informa-</p><p>tion and functionality to empower the unique vision and intuition of an active</p><p>operator. A clever system obfuscates the internal workings of the process,</p><p>reducing the human to a passive user who simply consumes the insights pro-</p><p>vided by the machine. AI should be a tool to be wielded, not a lens through</p><p>which we view the world. To steer such a powerful tool, we must be equipped</p><p>with the knowledge to understand and operate it without eroding the human</p><p>quality of our human systems.</p><p>The need to stay current on this technology is exactly why a book like Artif icial</p><p>Intelligence Basics is so important in today's world. It is the intellectual infra-</p><p>structure that will enable people—regular people—to tap into the power of</p><p>AI.Without these kinds of initiatives, AI will tip the balance of power in favor</p><p>of big companies with big budgets. It's crucial that the general population</p><p>equip themselves with the skills to understand AI systems, because these</p><p>systems will increasingly define how we interact with and navigate through the</p><p>world. Soon, the information contained in this book won't be merely a topic</p><p>of interest; it will be a prerequisite for participation in the modern economy.</p><p>This is how the average person can enjoy the fruits of the AI revolution. In the</p><p>years to come, how we define work and which activities carry economic value</p><p>will change. We have to embrace the fact that the future of work may be as</p><p>foreign to us as a desk job would be to our distant ancestors. But we have</p><p>to—and should—have faith in the human capacity to innovate new forms of</p><p>work, even if that work doesn't look like the work we're familiar with. But the</p><p>first step, before everything else, is to learn more about this new, exciting, and</p><p>fundamentally democratizing technology.</p><p>—Sridhar Vembu, co-founder and CEO of Zoho</p><p>Foreword</p><p>Introduction</p><p>On the face of it, the Uber app is simple. With just a couple clicks, you can</p><p>hail a driver within a few minutes.</p><p>But behind the scenes, there is an advanced technology platform, which relies</p><p>heavily on artificial intelligence (AI). Here are just some of the capabilities:</p><p>• A Natural Language Processing (NLP) system that can</p><p>understand conversations, allowing for a streamlined</p><p>experience</p><p>• Computer vision software that verifies millions of images</p><p>and documents like drivers’ licenses and restaurant</p><p>menus</p><p>• Sensor processing algorithms that help improve the accu-</p><p>racy in dense urban areas, including automatic crash</p><p>detection by sensing unexpected movement from the</p><p>phone of a driver or passenger</p><p>• Sophisticated machine learning algorithms that predict</p><p>driver supply, rider demand, and ETAs</p><p>Such technologies are definitely amazing, but they are also required. There is</p><p>no way that Uber could have scaled its growth—which has involved handling</p><p>over 10 billion trips—without AI.In light of this, it should be no surprise that</p><p>the company spends hundreds of millions on the technology and has a large</p><p>group of AI experts on staff.1</p><p>But AI is not just for fast-charging startups. The technology is also proving a</p><p>critical priority for traditional companies. Just look at McDonald’s. In 2019,</p><p>the company shelled out $300 million to acquire a tech startup, Dynamic</p><p>Yield. It was the company’s largest deal since it purchased Boston Market in</p><p>1999.2</p><p>1www.sec.gov/Archives/edgar/data/1543151/000119312519120759/d647752ds1a.</p><p>htm#toc647752_11</p><p>2https://news.mcdonalds.com/news-releases/news-release-details/dynamic-yield-</p><p>acquisition-release</p><p>http://www.sec.gov/Archives/edgar/data/1543151/000119312519120759/d647752ds1a.htm#toc647752_11</p><p>http://www.sec.gov/Archives/edgar/data/1543151/000119312519120759/d647752ds1a.htm#toc647752_11</p><p>https://news.mcdonalds.com/news-releases/news-release-details/dynamic-yield-acquisition-release</p><p>https://news.mcdonalds.com/news-releases/news-release-details/dynamic-yield-acquisition-release</p><p>x</p><p>Dynamic Yield, which was founded in 2011, is a pioneer in leveraging AI for</p><p>creating personalized customer interactions across the Web, apps, and email.</p><p>Some of its customers include the Hallmark Channel, IKEA, and Sephora.</p><p>As for McDonald’s, it has been undergoing a digital transformation—and AI is</p><p>a key part of the strategy. With Dynamic Yield, the company plans to use the</p><p>technology to reimagine its Drive Thru, which accounts for a majority of its</p><p>revenues. By analyzing data, such as the weather, traffic, and time of day, the</p><p>digital menus will be dynamically changed to enhance the revenue opportuni-</p><p>ties. It also looks like McDonald’s will use geofencing and even image recogni-</p><p>tion of license plates to enhance the targeting.</p><p>But this will just be the start. McDonald’s expects to use AI for in-store</p><p>kiosks and signage as well as the supply chain.</p><p>The company realizes that the future is both promising and dangerous. If com-</p><p>panies are not proactive with new technologies, they may ultimately fail. Just</p><p>look at how Kodak was slow to adapt to digital cameras. Or consider how the</p><p>taxi industry did not change when faced with the onslaught of Uber and Lyft.</p><p>On the other hand, new technologies can be almost an elixir for a company.</p><p>But there needs to be a solid strategy, a good understanding of what’s possi-</p><p>ble, and a willingness to take risks. So in this book, I’ll provide tools to help</p><p>with all this.</p><p>OK then, how big will AI get? According to a study from PWC, it will add a</p><p>staggering $15.7 trillion to the global GDP by 2030, which is more than the</p><p>combined output of China and India. The authors of the report note: “AI</p><p>touches almost every aspect of our lives. And it’s only just getting started.”3</p><p>True, when it comes to predicting trends, there can be a good deal of hype.</p><p>However, AI may be different because it has the potential for turning into</p><p>a general-purpose technology. A parallel to this is what happened in the</p><p>nineteenth century with the emergence of electricity, which had a transfor-</p><p>mative impact across the world.</p><p>As a sign of the strategic importance of AI, tech companies like Google,</p><p>Microsoft, Amazon.com, Apple, and Facebook have made substantial invest-</p><p>ments in this industry. For example, Google calls itself an “AI-first” company</p><p>and has spent billions buying companies in the space as well as hiring thou-</p><p>sands of data scientists.</p><p>In other words, more and more jobs will require knowledge of AI.Granted,</p><p>this does not mean you’ll need to learn programming languages or understand</p><p>advanced statistics. But it will be critical to have a solid foundation of the</p><p>fundamentals.</p><p>3www.pwc.com/gx/en/issues/data-and-analytics/publications/artificial-</p><p>intelligence-study.html</p><p>Introduction</p><p>http://www.pwc.com/gx/en/issues/data-and-analytics/publications/artificial-intelligence-study.html</p><p>http://www.pwc.com/gx/en/issues/data-and-analytics/publications/artificial-intelligence-study.html</p><p>normal distribution—which has a shape like a bell—</p><p>represents the sum of probabilities for a variable.</p><p>• The Bayes’ theorem is a sophisticated statistical technique</p><p>that provides a deeper look at probabilities.</p><p>Chapter 3 | Machine Learning</p><p>67</p><p>• A true positive is when a model makes a correct</p><p>prediction. A false positive, on the other hand, is when a</p><p>model prediction shows that the result is true even</p><p>though it is not.</p><p>• The Pearson correlation shows the strength of the</p><p>relationship between two variables that range from</p><p>1 to -1.</p><p>• Feature extraction or feature engineering describes the</p><p>process of selecting variables for a model. This is critical</p><p>since even one wrong variable can have amajor impact</p><p>on the results.</p><p>• Training data is what is used to create the relationships in</p><p>an algorithm. The test data, on the other hand, is used to</p><p>evaluate the model.</p><p>• Supervised learning uses labeled data to create a model,</p><p>whereas unsupervised learning does not. There is also</p><p>semi-supervised learning, which uses a mix of both</p><p>approaches.</p><p>• Reinforcement learning is a way to train a model by</p><p>rewarding accurate predictions and punishing those that</p><p>are not.</p><p>• The k-Nearest Neighbor (k-NN) is an algorithm based</p><p>on the notion that values that are close together are</p><p>good predictors for a model.</p><p>• Linear regression estimates the relationship between</p><p>certain variables. The R-squared will indicate the strength</p><p>of the relationship.</p><p>• A decision tree is a model that is based on a workflow of</p><p>yes/no decisions.</p><p>• An ensemble model uses more than one model for the</p><p>predictions.</p><p>• The k-Means clustering algorithm puts similar unlabeled</p><p>data into different groups.</p><p>Artificial Intelligence Basics</p><p>© Tom Taulli 2019</p><p>T. Taulli, Artif icial Intelligence Basics,</p><p>https://doi.org/10.1007/978-1-4842-5028-0_4</p><p>C H A P T E R</p><p>4</p><p>Deep Learning</p><p>The Revolution in AI</p><p>Take any old classif ication problem where you have a lot of data, and it’s</p><p>going to be solved by deep learning. There’s going to be thousands of</p><p>applications of deep learning.</p><p>—Geoffrey Hinton,</p><p>English Canadian cognitive psychologist and computer scientist1</p><p>Fei-Fei Li, who got a BA degree in physics from Princeton in 1999 with high</p><p>honors and a PhD in electrical engineering from Caltech in 2005, focused her</p><p>brilliance on developing AI models. But she had a major challenge: finding</p><p>quality datasets. At first, she looked at creating them by hand, such as with</p><p>graduate students who downloaded images from the Internet. But the process</p><p>was too slow and tedious.</p><p>One day a student mentioned to Li that Amazon.com’s Mechanical Turk, an</p><p>online service that uses crowdsourcing to solve problems, could be a good</p><p>way to scale the process. It would allow for fast and accurate labeling of the</p><p>data.</p><p>Li gave it a try, and it worked out quite well. By 2010, she had created</p><p>ImageNet, which had 3.2 million images across over 5,200 categories.</p><p>1 Siddhartha Mukherjee, “The Algorithm Will See You Now,” The NewYorker, April 3, 2017,</p><p>https://www.newyorker.com/magazine/2017/04/03/ai-versus-md.</p><p>https://www.newyorker.com/magazine/2017/04/03/ai-versus-md</p><p>70</p><p>Yet it got a tepid response from the academic community. But this did not</p><p>deter Li. She continued to work tirelessly to evangelize the dataset. In 2012,</p><p>she put together a contest as a way to encourage researchers to create more</p><p>effective models and push the boundaries of innovation. It would turn out to</p><p>be a game changer, and the contest would become an annual event.</p><p>In the first contest, professors from the University of Toronto—Geoffrey</p><p>Hinton, Ilya Sutskever, and Alex Krizhevsky—used sophisticated deep learning</p><p>algorithms. And the results were standout. The system they built, which was</p><p>called AlexNet, beat all the other contestants by a margin of 10.8%.2</p><p>This was no fluke. In the years after this, deep learning continued to show</p><p>accelerated progress with ImageNet. As of now, the error rate for deep</p><p>learning is a mere 2% or so—which is better than humans.</p><p>By the way, Li has since gone on to become a professor at Stanford and</p><p>co-director of the school’s AI lab. She is also Google’s chief scientist of AI and</p><p>Machine Learning. Needless to say, whenever she has new ideas now, people listen!</p><p>In this chapter, we’ll take a look at deep learning, which is clearly the hottest</p><p>area of AI.It has led to major advances in areas like self-driving cars and virtual</p><p>assistants like Siri.</p><p>Yes, deep learning can be a complicated subject, and the field is constantly</p><p>changing. But we’ll take a look at the main concepts and trends—without</p><p>getting into the technical details.</p><p>Difference Between Deep Learning</p><p>andMachine Learning</p><p>There is often confusion between deep learning and machine learning. And</p><p>this is reasonable. Both topics are quite complex, and they do share many</p><p>similarities.</p><p>So to understand the differences, let’s first take a look at two high-level</p><p>aspects of machine learning and how they relate to deep learning. First of all,</p><p>while both usually require large amounts of data, the types are generally</p><p>different.</p><p>Take the following example: Suppose we have photos of thousands of animals</p><p>and want to create an algorithm to find the horses. Well, machine learning</p><p>cannot analyze the photos themselves; instead, the data must be labeled. The</p><p>machine learning algorithm will then be trained to recognize horses, through</p><p>a process known as supervised learning (covered in Chapter 3).</p><p>2 https://qz.com/1034972/the-data-that-changed-the-direction-of-ai-research-and-</p><p>possibly-the-world/</p><p>Chapter 4 | Deep Learning</p><p>https://qz.com/1034972/the-data-that-changed-the-direction-of-ai-research-and-possibly-the-world/</p><p>https://qz.com/1034972/the-data-that-changed-the-direction-of-ai-research-and-possibly-the-world/</p><p>71</p><p>Even though machine learning will likely come up with good results, they will</p><p>still have limitations. Wouldn’t it be better to look at the pixels of the images</p><p>themselves—and find the patterns? Definitely.</p><p>But to do this with machine learning, you need to use a process called feature</p><p>extraction. This means you must come up with the kinds of characteristics of</p><p>a horse—such as the shape, the hooves, color, and height—which the algorithms</p><p>will then try to identify.</p><p>Again, this is a good approach—but it is far from perfect. What if your</p><p>features are off the mark or do not account for outliers or exceptions? In</p><p>such cases, the accuracy of the model will likely suffer. After all, there are</p><p>many variations to a horse. Feature extraction also has the drawback of</p><p>ignoring a large amount of the data. This can be exceedingly complicated—if</p><p>not impossible—for certain use cases. Look at computer viruses. Their</p><p>structures and patterns, which are known as signatures, are constantly</p><p>changing so as to infiltrate systems. But with feature extraction, a person</p><p>would somehow have to anticipate this, which is not practical. This is why</p><p>cybersecurity software is often about collecting signatures after a virus has</p><p>exacted damage.</p><p>But with deep learning, we can solve these problems. This approach analyzes</p><p>all the data—pixel by pixel—and then finds the relationships by using a neural</p><p>network, which mimics the human brain.</p><p>Let’s take a look.</p><p>So What Is Deep Learning Then?</p><p>Deep learning is a subfield of machine learning. This type of system allows for</p><p>processing huge amounts of data to find relationships and patterns that</p><p>humans are often unable to detect. The word “deep” refers to the number of</p><p>hidden layers in the neural network, which provide much of the power to</p><p>learn.</p><p>When it comes to the topic of AI, deep learning is at the cutting-edge and</p><p>often generates most of the buzz in mainstream media. “[Deep learning] AI is</p><p>the new electricity,” extolled Andrew Yan-Tak Ng, who is the former chief</p><p>scientist at Baidu and co-founder of Google Brain.3</p><p>But it is also important to remember that deep learning is still in the early</p><p>stages of development and commercialization.</p><p>For example, it was not until</p><p>about 2015 that Google started using this technology for its search engine.</p><p>3 https://medium.com/@GabriellaLeone/the-best-explanation-machine-learning-vs-</p><p>deep-learning-d5c123405b11</p><p>Artificial Intelligence Basics</p><p>https://medium.com/@GabriellaLeone/the-best-explanation-machine-learning-vs-deep-learning-d5c123405b11</p><p>https://medium.com/@GabriellaLeone/the-best-explanation-machine-learning-vs-deep-learning-d5c123405b11</p><p>72</p><p>As we saw in Chapter 1, the history of neural networks was full of ebbs and</p><p>flows. It was Frank Rosenblatt who created the perceptron, which was a fairly</p><p>basic system. But real academic progress with neural networks did not occur</p><p>until the 1980s, such as with the breakthroughs with backpropagation,</p><p>convolutional neural networks, and recurrent neural networks. But for deep</p><p>learning to have an impact on the real world, it would take the staggering</p><p>growth in data, such as from the Internet, and the surge in computing power.</p><p>The Brain andDeep Learning</p><p>Weighing only about 3.3 pounds, the human brain is an amazing feat of</p><p>evolution. There are about 86 billion neurons—often called gray matter—</p><p>that are connected with trillions of synapses. Think of neurons as CPUs</p><p>(Central Processing Units) that take in data. The learning occurs with the</p><p>strengthening or weakening of the synapses.</p><p>The brain is made up of three regions: the forebrain, the midbrain, and the</p><p>hindbrain. Among these, there are a variety of areas that perform different</p><p>functions. Some of the main ones include the following:</p><p>• Hippocampus: This is where your brain stores memories.</p><p>In fact, this is the part that fails when a person has</p><p>Alzheimer’s disease, in which a person loses the ability to</p><p>form short-term memories.</p><p>• Frontal Lobe: Here the brain focuses on emotions, speech,</p><p>creativity, judgment, planning, and reasoning.</p><p>• Cerebral Cortex: This is perhaps the most important when</p><p>it comes to AI.The cerebral cortex helps with thinking</p><p>and other cognitive activities. According to research</p><p>from Suzana Herculano-Houzel, the level of intelligence</p><p>is related to the number of neurons in this area of the</p><p>brain.</p><p>Then how does deep learning compare to the human brain? There are some</p><p>tenuous similarities. At least in areas like the retina, there is a process of</p><p>ingesting data and processing them through a complex network, which is</p><p>based on assigning weights. But of course, this is only a minute part of the</p><p>learning process. Besides, there are still many mysteries about the human</p><p>brain, and of course, it is not based on things like digital computing (instead,</p><p>it appears that it is more of an analogue system). However, as the research</p><p>continues to advance, the discoveries in neuroscience could help build new</p><p>models for AI.</p><p>Chapter 4 | Deep Learning</p><p>73</p><p>Artificial Neural Networks (ANNs)</p><p>At the most basic level, an artificial neural network (ANN) is a function that</p><p>includes units (which may also be called neurons, perceptrons, or nodes).</p><p>Each unit will have a value and a weight, which indicates the relative importance,</p><p>and will go into the hidden layer. The hidden layer uses a function, with the</p><p>result becoming the output. There is also another value, called bias, which is</p><p>a constant and is used in the calculation of the function.</p><p>This type of training of a model is called a feed-forward neural network. In</p><p>other words, it only goes from input to the hidden layer to the output. It does</p><p>not cycle back. But it could go to a new neural network, with the output</p><p>becoming the input.</p><p>Figure 4-1 shows a chart of a feed-forward neural network.</p><p>Let’s go deeper on this by taking an example. Suppose you are creating a</p><p>model to predict whether a company’s stock will increase. The following are</p><p>what the variables represent as well as the values and weights assigned:</p><p>• X1: Revenues are growing at a minimum of 20% a year.</p><p>The value is 2.</p><p>• X2: The profit margin is at least 20%. The value is 4.</p><p>• W1: 1.9.</p><p>• W2: 9.6.</p><p>• b: This is the bias (the value is 1), which helps smooth out</p><p>the calculations.</p><p>You’ll then sum the weights, and then the function will process the information.</p><p>This will often involve an activation function, which is non-linear. This is more</p><p>reflective of the real world since data is usually not in a straight line.</p><p>Figure 4-1. A basic feed-forward neural network</p><p>Artificial Intelligence Basics</p><p>74</p><p>Now there are a variety of activation functions to choose from. One of the</p><p>most common is the sigmoid. This compresses the input value into a range of</p><p>0–1. The closer it is to 1, the more accurate the model.</p><p>When you graph this function, it will look like an S shape. See Figure4-2.</p><p>As you can see, the system is relatively simplistic and will not be helpful in</p><p>high-end AI models. To add much more power, there usually needs to be</p><p>multiple hidden layers. This results in a multilayered perceptron (MLP). It also</p><p>helps to use something called backpropagation, which allows for the output</p><p>to be circled back into the neural network.</p><p>Backpropagation</p><p>One of the major drawbacks with artificial neural networks is the process of</p><p>making adjustments to the weights in the model. Traditional approaches, like</p><p>the use of the mutation algorithm, used random values that proved to be time</p><p>consuming.</p><p>Given this, researchers looked for alternatives, such as backpropagation. This</p><p>technique had been around since the 1970s but got little interest as the</p><p>performance was lacking. But David Rumelhart, Geoffrey Hinton, and Ronald</p><p>Williams realized that backpropagation still had potential, so long as it was</p><p>refined. In 1986, they wrote a paper entitled “Learning Representations by</p><p>Back-propagating Errors,” and it was a bombshell in the AI community.4 It</p><p>clearly showed that backpropagation could be much faster but also allow for</p><p>more powerful artificial neural networks.</p><p>As should be no surprise, there is a lot of math involved in backpropagation.</p><p>But when you boil things down, it’s about adjusting the neural network when</p><p>errors are found and then iterating the new values through the neural network</p><p>again. Essentially, the process involves slight changes that continue to optimize</p><p>the model.</p><p>4 David E.Rumelhart, Geoffrey E.Hinton, and Ronald J.Williams, “Learning Representations</p><p>by Back-propagating Errors,” Nature 323 (1986): 533–536.</p><p>Figure 4-2. A typical sigmoid activation function</p><p>Chapter 4 | Deep Learning</p><p>75</p><p>For example, let’s say one of the inputs has an output of 0.6. This means that</p><p>the error is 0.4 (1.0 minus 0.6), which is subpar. But we can then backpropogate</p><p>the output, and perhaps the new output may get to 0.65. This training will go</p><p>on until the value is much closer to 1.</p><p>Figure 4-3 illustrates this process. At first, there is a high level of errors</p><p>because the weights are too large. But by making iterations, the errors will</p><p>gradually fall. However, doing too much of this could mean an increase in</p><p>errors. In other words, the goal of backpropagation is to find the midpoint.</p><p>As a gauge of the success of backpropagation, there were a myriad of</p><p>commercial applications that sprung up. One was called NETtalk, which was</p><p>developed by Terrence Sejnowski and Charles Rosenberg in the mid-1980s.</p><p>The machine was able to learn how to pronounce English text. NETtalk was</p><p>so interesting that it was even demoed on the Today show.</p><p>There were also a variety of startups that were created that leveraged</p><p>backpropagation, such as HNC Software. It built models that detected credit</p><p>card fraud. Up until this point—when HNC was founded in the late 1980s—</p><p>the process was done mostly by hand, which led to costly errors and low</p><p>volumes of issuances. But by using deep learning approaches, credit card</p><p>companies were able to save billions of dollars.</p><p>In 2002, HNC was acquired by Fair, Isaac and valued at $810 million.5</p><p>5 www.insurancejournal.com/news/national/2002/05/01/16857.htm</p><p>Figure 4-3. The optimal value for a backpropagation function is at</p><p>the bottom of the graph</p><p>Artificial Intelligence Basics</p><p>http://www.insurancejournal.com/news/national/2002/05/01/16857.htm</p><p>76</p><p>The Various Neural Networks</p><p>The most basic type of a neural network is a fully connected neural network.</p><p>As the name implies, it is where all the neurons have connections from layer</p><p>to layer. This network is actually quite popular since it means having to use</p><p>little judgment when creating the model.</p><p>Then what are some of the other neural networks? The common ones include</p><p>the recurrent neural network (RNN), the convolutional neural network</p><p>(CNN), and the generative adversarial network (GAN), which we’ll cover</p><p>next.</p><p>Recurrent Neural Network</p><p>With a recurrent neural network (RNN), the function not only processes the</p><p>input but also prior inputs across time. An example of this is what happens</p><p>when you enter characters in a messaging app. As you begin to type, the</p><p>system will predict the words. So if you tap out “He,” the computer will</p><p>suggest “He,” “Hello,” and “Here’s.” The RNN is essentially a string of neural</p><p>networks that feed on each other based on complex algorithms.</p><p>There are variations on the model. One is called LSTM, which stands for long</p><p>short-term memory. This came about from a paper written by professors</p><p>Sepp Hochreiter and Jürgen Schmidhuber in 1997.6 In it, they set forth a way</p><p>to effectively use inputs that are separated from each other for long time</p><p>periods, allowing the use of more datasets.</p><p>Of course, RNNs do have drawbacks. There is the vanishing gradient problem,</p><p>which means that the accuracy decays as the models get larger. The models</p><p>can also take longer to train.</p><p>To deal with this, Google developed a new model called the Transformer,</p><p>which is much more efficient since it processes the inputs in parallel. It also</p><p>results in more accurate results.</p><p>Google has gained much insight about RNNs through its Translate app, which</p><p>handles over 100 languages and processes over 100 billion words a day.7</p><p>Launched in 2006, it initially used machine learning systems. But in 2016,</p><p>Google switched to deep learning by creating Google Neural Machine</p><p>Translation.8 All in all, it has resulted in much higher accuracy rates.9</p><p>6 Sepp Hochreiter and Jürgen Schmidhuber, “Long Short-Term Memory,” Neural Computation</p><p>9, no. 8 (1997): 1735-80.</p><p>7 www.argotrans.com/blog/accurate-google-translate-2018/</p><p>8 www.techspot.com/news/75637-google-translate-not-monetized-despite-</p><p>converting-over-100.html</p><p>9 www.argotrans.com/blog/accurate-google-translate-2018/</p><p>Chapter 4 | Deep Learning</p><p>http://www.argotrans.com/blog/accurate-google-translate-2018/</p><p>http://www.techspot.com/news/75637-google-translate-not-monetized-despite-converting-over-100.html</p><p>http://www.techspot.com/news/75637-google-translate-not-monetized-despite-converting-over-100.html</p><p>http://www.argotrans.com/blog/accurate-google-translate-2018/</p><p>77</p><p>Consider how Google Translate has helped out doctors who work with</p><p>patients who speak other languages. According to a study from the University</p><p>of California, San Francisco (UCSF), that was published in JAMA Internal</p><p>Medicine, the app had a 92% accuracy rate with English-to-Spanish</p><p>translations. This was up from 60% over the past couple years.10</p><p>Convolutional Neural Network (CNN)</p><p>Intuitively, it makes sense to have all the units in a neural network to be</p><p>connected. This works well with many applications.</p><p>But there are scenarios where it is far from optimal, such as with image</p><p>recognition. Just imagine how complex a model would be where every pixel</p><p>is a unit! It could quickly become unmanageable. There would also be other</p><p>complications like overfitting. This is where the data is not reflective of</p><p>what is being tested or there is a focus on the wrong features.</p><p>To deal with all this, you can use a convolutional neural network (CNN).</p><p>The origins of this go back to professor Yann LeCun in 1998, when he</p><p>published a paper called “Gradient-Based Learning Applied to Document</p><p>Recognition.”11 Despite its strong insights and breakthroughs, it got little</p><p>traction. But as deep learning started to show significant progress in 2012,</p><p>researchers revisited the model.</p><p>LeCun got his inspiration for the CNN from Nobel Prize winners David</p><p>Hubel and Torsten Wiesel who studied neurons of the visual cortex. This</p><p>system takes an image from the retina and processes it in different stages—</p><p>from easy to more complex. Each of the stages is called a convolution. For</p><p>example, the first level would be to identify lines and angles; next, the visual</p><p>cortex will find the shapes; and then it will detect the objects.</p><p>This is analogous to how a computer-based CNN works. Let’s take an</p><p>example: Suppose you want to build a model that can identify a letter. The</p><p>CNN will have input in the form of an image that has 3,072 pixels. Each of</p><p>the pixels will have a value that is from 0 to 255, which indicates the overall</p><p>intensity. By using a CNN, the computer will go through multiple variations</p><p>to identify the features.</p><p>The first is the convolutional layer, which is a filter that scans the image. In</p><p>our example, this could be 5 × 5 pixels. The process will create a feature map,</p><p>which is a long array of numbers. Next, the model will apply more filters to</p><p>the image. By doing this, the CNN will identify the lines, edges and shapes—all</p><p>10 https://gizmodo.com/google-translate-can-help-doctors-bridge-the-language-</p><p>g-1832881294</p><p>11 Yann LeCun et al., “Gradient-Based Learning Applied to Document Recognition,”</p><p>Proceedings of the IEEE 86 no. 11 (1998): 2278-2324.</p><p>Artificial Intelligence Basics</p><p>https://gizmodo.com/google-translate-can-help-doctors-bridge-the-language-g-1832881294</p><p>https://gizmodo.com/google-translate-can-help-doctors-bridge-the-language-g-1832881294</p><p>78</p><p>expressed in numbers. With the various output layers, the model will use</p><p>pooling, which combines them to generate a single output, and then create a</p><p>fully connected neural network.</p><p>A CNN can definitely get complex. But it should be able to accurately identify</p><p>the numbers that are input into the system.</p><p>Generative Adversarial Networks (GANs)</p><p>Ian Goodfellow, who got his masters in computer science at Stanford and</p><p>his PhD in machine learning at the Université de Montréal, would go on to</p><p>work at Google. In his 20s, he co-authored one of the top books in AI,</p><p>called Deep Learning,12 and also made innovations with Google Maps.</p><p>But it was in 2014 that he had his most impactful breakthrough. It actually</p><p>happened in a pub in Montreal when he talked with some of his friends about</p><p>how deep learning could create photos.13 At the time, the approach was to</p><p>use generative models, but they were often blurry and nonsensical.</p><p>Goodfellow realized that there had to be a better why. So why not use game</p><p>theory? That is, have two models compete against each other in a tight</p><p>feedback loop. This could also be done with unlabeled data.</p><p>Here’s a basic workflow:</p><p>• Generator: This neural network creates a myriad of new</p><p>creations, such as photos or sentences.</p><p>• Discriminator: This neural network looks at the creations</p><p>to see which ones are real.</p><p>• Adjustments: With the two results, a new model would</p><p>change the creations to make them as realistic as possible.</p><p>Through many iterations, the discriminator will no longer</p><p>need to be used.</p><p>He was so excited about the idea that after he left the pub he started to code</p><p>his ideas. The result was a new deep learning model: the generative adversarial</p><p>network or GAN.And the results were standout. He would soon become an</p><p>AI rock star.</p><p>12 Ian Goodfellow, Yoshua Bengio, and Aaron Courville, Deep Learning (Cambridge, MA:</p><p>The MIT Press, 2016).</p><p>13 www.technologyreview.com/s/610253/the-ganfather-the-man-whos-given-</p><p>machines-the-gift-of-imagination/</p><p>Chapter 4 | Deep Learning</p><p>http://www.technologyreview.com/s/610253/the-ganfather-the-man-whos-given-machines-the-gift-of-imagination/</p><p>http://www.technologyreview.com/s/610253/the-ganfather-the-man-whos-given-machines-the-gift-of-imagination/</p><p>79</p><p>GAN research has already spurred over 500 academic papers.14 Companies</p><p>like Facebook have also used this technology, such as for its photo analysis and</p><p>processing. The company’s chief AI scientist, Yann LeCun, noted that GANs</p><p>are the “the coolest idea in deep learning in the last 20 years.”15</p><p>GANs have also been shown to help with sophisticated scientific research.</p><p>For example, they have helped improve the accuracy of detecting behavior of</p><p>subatomic particles in the Large Hadron Collider at CERN in Switzerland.16</p><p>While still in the early innings, this technology could lead to such things as a</p><p>computer that can develop new types of fashion items or maybe a new-fangled</p><p>wearable. Perhaps a GAN could even come up with a hit rap song.</p><p>And it could be sooner than you think. As a teenager, Robbie Barrat taught</p><p>himself how to use deep learning systems and built a model to rap in the style</p><p>of Kanye West.</p><p>But this was just the beginning of his AI wizardry. As a researcher at Stanford,</p><p>he developed his own GAN platform, which processed roughly 10,000 nude</p><p>portraits. The system then would create truly mesmerizing new works of art</p><p>(you can find them at his Twitter account at @DrBeef_).</p><p>Oh, and he also made his system open source at his GitHub account. This</p><p>caught the attention of a collective of French artists, called Obvious, that used</p><p>the technology to create portraits of an eighteenth-century fictional family. It</p><p>was based on processing 15,000 portraits from the fourteenth to the twentieth</p><p>centuries.</p><p>In 2018, Obvious put its artwork at a Christie’s auction, fetching a cool</p><p>$432,000. 17</p><p>But unfortunately, when it comes to GANs, there have been uses that have</p><p>been less than admirable. One example is to use them for deepfakes, which</p><p>involve leveraging neural networks to create images or videos that are</p><p>misleading. Some of this is just kind of playful. For example, one GAN makes</p><p>it possible to have Barack Obama say anything you tell him!</p><p>Yet there are lots of risks. Researchers at New York University and the</p><p>Michigan State University wrote a paper that focused on “DeepMasterPrints.”18</p><p>14 https://github.com/hindupuravinash/the-gan-zoo</p><p>15 https://trendsandevents4developers.wordpress.com/2017/04/24/</p><p>the-coolest-idea-in-deep-learning-in-20-years-and-more/</p><p>16 www.hpcwire.com/2018/08/14/cern-incorporates-ai-into-physics-</p><p>based-simulations/</p><p>17 www.washingtonpost.com/nation/2018/10/26/year-old-developed-code-ai-</p><p>portrait-that-sold-christies/?utm_term=.b2f366a4460e</p><p>18 www.cnbc.com/2018/12/28/research-claims-fake-fingerprints-could-hack-</p><p>a-third-of-smartphones.html</p><p>Artificial Intelligence Basics</p><p>https://github.com/hindupuravinash/the-gan-zoo</p><p>https://trendsandevents4developers.wordpress.com/2017/04/24/the-coolest-idea-in-deep-learning-in-20-years-and-more/</p><p>https://trendsandevents4developers.wordpress.com/2017/04/24/the-coolest-idea-in-deep-learning-in-20-years-and-more/</p><p>http://www.hpcwire.com/2018/08/14/cern-incorporates-ai-into-physics-based-simulations/</p><p>http://www.hpcwire.com/2018/08/14/cern-incorporates-ai-into-physics-based-simulations/</p><p>http://www.washingtonpost.com/nation/2018/10/26/year-old-developed-code-ai-portrait-that-sold-christies/?utm_term=.b2f366a4460e</p><p>http://www.washingtonpost.com/nation/2018/10/26/year-old-developed-code-ai-portrait-that-sold-christies/?utm_term=.b2f366a4460e</p><p>http://www.cnbc.com/2018/12/28/research-claims-fake-fingerprints-could-hack-a-third-of-smartphones.html</p><p>http://www.cnbc.com/2018/12/28/research-claims-fake-fingerprints-could-hack-a-third-of-smartphones.html</p><p>80</p><p>It showed how a GAN can develop fake fingerprints to unlock three types of</p><p>smartphones!</p><p>Then there was the incident of a so-called deepfake video of actress Jennifer</p><p>Lawrence at a Golden Globes press conference. Her face was merged with</p><p>Steve Buscemi’s.19</p><p>Deep Learning Applications</p><p>With so much money and resources being devoted to deep learning, there</p><p>has been a surge in innovations. It seems that every day there is something</p><p>amazing that is being announced.</p><p>Then what are some of the applications? Where has deep learning proven to</p><p>be a game changer? Let’s take a look at some that cover areas like healthcare,</p><p>energy, and even earthquakes</p><p>Use Case: Detecting Alzheimer’s Disease</p><p>Despite decades of research, a cure for Alzheimer’s disease remains elusive.</p><p>Although, scientists have developed drugs that have slowed down the</p><p>progression of the disease.</p><p>In light of this, early diagnosis is critical—and deep learning can potentially be</p><p>a big help. Researchers at the UCSF Department of Radiology and Biomedical</p><p>Imaging have used this technology to analyze brain screens—from the</p><p>Alzheimer’s Disease Neuroimaging Initiative public dataset—and to detect</p><p>changes in the levels of glucose.</p><p>The result: The model can diagnose Alzheimer’s disease up to six years before</p><p>a clinical diagnosis. One of the tests showed a 92% accuracy rate, and another</p><p>was 98%.</p><p>Now this is still in the beginning phases—and there will need to be more</p><p>datasets analyzed. But so far, the results are very encouraging.</p><p>According to Dr. Jae Ho Sohn, who authored the study:</p><p>This is an ideal application of deep learning because it is particularly</p><p>strong at finding very subtle but diffuse processes. Human radiologists</p><p>are really strong at identifying tiny focal finding like a brain tumor, but we</p><p>struggle at detecting more slow, global changes. Given the strength of</p><p>deep learning in this type of application, especially compared to humans,</p><p>it seemed like a natural application.20</p><p>19 http://fortune.com/2019/01/31/what-is-deep-fake-video/</p><p>20 www.ucsf.edu/news/2018/12/412946/artificial-intelligence-can-detect-</p><p>alzheimers-disease-brain-scans-six-years</p><p>Chapter 4 | Deep Learning</p><p>http://fortune.com/2019/01/31/what-is-deep-fake-video/</p><p>http://www.ucsf.edu/news/2018/12/412946/artificial-intelligence-can-detect-alzheimers-disease-brain-scans-six-years</p><p>http://www.ucsf.edu/news/2018/12/412946/artificial-intelligence-can-detect-alzheimers-disease-brain-scans-six-years</p><p>81</p><p>Use Case: Energy</p><p>Because of its massive data center infrastructure, Google is one of the largest</p><p>consumers of energy. Even a small improvement in efficiency can lead to a</p><p>sizeable impact on the bottom line. But there could also be the benefits of</p><p>less carbon emissions.</p><p>To help with these goals, Google’s DeepMind unit has been applying deep</p><p>learning, which has involved better management of wind power. Even</p><p>though this is a clean source of energy, it can be tough to use because of</p><p>the changes in weather.</p><p>But DeepMind’s deep learning algorithms have been critical. Applied to 700</p><p>megawatts of wind power in the United States, they were able to make</p><p>accurate forecasts for output with a lead time of 36 hours.</p><p>According to DeepMind’s blog:</p><p>This is important, because energy sources that can be scheduled (i.e.</p><p>can deliver a set amount of electricity at a set time) are often more</p><p>valuable to the grid…To date, machine learning has boosted the value of</p><p>our wind energy by roughly 20 percent, compared to the baseline</p><p>scenario of no time-based commitments to the grid.21</p><p>But of course, this deep learning system could be more than just about</p><p>Google—it could have a wide-ranging impact on energy use across the</p><p>world.</p><p>Use Case: Earthquakes</p><p>Earthquakes are extremely complicated to understand. They are also</p><p>exceedingly difficult to predict. You need to evaluate faults, rock formations</p><p>and deformations, electromagnetic activity, and changes in the groundwater.</p><p>Hey, there is even evidence that animals have the ability to sense an earthquake!</p><p>But over the decades, scientists have collected huge amounts of data on this</p><p>topic. In other words, this could be an application for deep learning, right?</p><p>Absolutely.</p><p>Seismologists at Caltech, which include Yisong Yue, Egill Hauksson, Zachary</p><p>Ross, and Men-Andrin Meier, have been doing considerable research on this,</p><p>using convolutional neural networks and recurrent neural networks. They are</p><p>trying to build</p><p>an effective early-warning system.</p><p>21 https://deepmind.com/blog/machine-learning-can-boost-value-</p><p>wind-energy/</p><p>Artificial Intelligence Basics</p><p>https://deepmind.com/blog/machine-learning-can-boost-value-wind-energy/</p><p>https://deepmind.com/blog/machine-learning-can-boost-value-wind-energy/</p><p>82</p><p>Here’s what Yue had to say:</p><p>AI can [analyze earthquakes] faster and more accurately than humans</p><p>can, and even find patterns that would otherwise escape the human eye.</p><p>Furthermore, the patterns we hope to extract are hard for rule-based</p><p>systems to adequately capture, and so the advanced pattern-matching</p><p>abilities of modern deep learning can offer superior performance than</p><p>existing automated earthquake monitoring algorithms.22</p><p>But the key is improving data collection. This means more analysis of small</p><p>earthquakes (in California, there is an average of 50 each day). The goal is to</p><p>create an earthquake catalog that can lead to the creation of a virtual</p><p>seismologist, who can make evaluations of an earthquake faster than a human.</p><p>This could allow for faster lead times when an earthquake strikes, which may</p><p>help to save lives and property.</p><p>Use Case: Radiology</p><p>PET scans and MRIs are amazing technology. But there are definitely downsides.</p><p>A patient needs to stay within a confining tube for 30 minutes to an hour. This</p><p>is uncomfortable and means being exposed to gadolinium, which has been</p><p>shown to have harmful side effects.</p><p>Greg Zaharchuk and Enhao Gong, who met at Stanford, thought there could</p><p>be a better way. Zaharchuk was an MD and PhD, with a specialization in</p><p>radiology. He was also the doctoral advisor of Gong, who was an electrical</p><p>engineering PhD in deep learning and medical image reconstruction.</p><p>In 2017, they co-founded Subtle Medical and hired some of the brightest</p><p>imaging scientists, radiologists, and AI experts. Together, they set themselves</p><p>to the challenge of improving PET scans and MRIs. Subtle Medical created a</p><p>system that not only reduces the time for an MRI and PET scans by up to ten</p><p>times, but the accuracy has been much higher. This was powered by high-end</p><p>NVIDIA GPUs.</p><p>Then in December 2018, the system received FDA (Federal Drug</p><p>Administration) 510(k) clearance and a CE mark approval for the European</p><p>market.23 It was the first ever AI-based nuclear medical device to achieve both</p><p>of these designations.</p><p>22 www.caltech.edu/about/news/qa-creating-virtual-seismologist-84789</p><p>23 https://subtlemedical.com/subtle-medical-receives-fda-510k-</p><p>clearance-and-ce-mark-approval-for-subtlepet/</p><p>Chapter 4 | Deep Learning</p><p>http://www.caltech.edu/about/news/qa-creating-virtual-seismologist-84789</p><p>https://subtlemedical.com/subtle-medical-receives-fda-510k-clearance-and-ce-mark-approval-for-subtlepet/</p><p>https://subtlemedical.com/subtle-medical-receives-fda-510k-clearance-and-ce-mark-approval-for-subtlepet/</p><p>83</p><p>Subtle Medical has more plans to revolutionize the radiology business. As of</p><p>2019, it is developing SubtleMRTM, which will be even more powerful than</p><p>the company’s current solution, and SubtleGADTM, which will reduce</p><p>gadolinium dosages.24</p><p>Deep Learning Hardware</p><p>Regarding chip systems for deep learning, GPUs have been the primary choice.</p><p>But as AI gets more sophisticated—such as with GANs—and the datasets</p><p>much larger, there is certainly more room for new approaches. Companies</p><p>also have custom needs, such as in terms of functions and data. After all, an</p><p>app for a consumer is usually quite different than one that is focused on the</p><p>enterprise.</p><p>As a result, some of the mega tech companies have been developing their own</p><p>chipsets:</p><p>• Google: In the summer of 2018, the company announced</p><p>its third version of its Tensor Processing Unit (TPU; the</p><p>first chip was developed in 2016).25 The chips are so</p><p>powerful—handling over 100 petaflops for training of</p><p>models—there needs to be liquid cooling in the data</p><p>centers. Google has also announced a version of its TPU</p><p>for devices. Essentially, it means that processing will have</p><p>less latency because there will be no need to access the</p><p>cloud.</p><p>• Amazon: In 2018, the company announced AWS</p><p>Inferentia.26 The technology, which has come out of the</p><p>acquisition of Annapurna in 2015, is focused on handling</p><p>complex inference operations. In other words, this is</p><p>what happens after a model has been trained.</p><p>• Facebook and Intel: These companies have joined forces to</p><p>create an AI chip.27 But the initiative is still in the initial</p><p>phases. Intel also has been getting traction with an AI</p><p>chip called the Nervana Neural Network Processor</p><p>(NNP).</p><p>24 www.streetinsider.com/Press+Releases/Subtle+Medical+Receives+FDA+510%2</p><p>8k%29+Clearance+and+CE+Mark+Approval+for+SubtlePET™/14892974.html</p><p>25 www.theregister.co.uk/2018/05/09/google_tpu_3/</p><p>26 https://aws.amazon.com/about-aws/whats-new/2018/11/announcing-</p><p>amazon-inferentia-machine-learning-inference-microchip/</p><p>27 www.analyticsindiamag.com/inference-chips-are-the-next-big-</p><p>battlefield-for-nvidia-and-intel/</p><p>Artificial Intelligence Basics</p><p>http://www.streetinsider.com/Press+Releases/Subtle+Medical+Receives+FDA+510%28k%29+Clearance+and+CE+Mark+Approval+for+SubtlePET%E2%84%A2/14892974.html</p><p>http://www.streetinsider.com/Press+Releases/Subtle+Medical+Receives+FDA+510%28k%29+Clearance+and+CE+Mark+Approval+for+SubtlePET%E2%84%A2/14892974.html</p><p>http://www.theregister.co.uk/2018/05/09/google_tpu_3/</p><p>https://aws.amazon.com/about-aws/whats-new/2018/11/announcing-amazon-inferentia-machine-learning-inference-microchip/</p><p>https://aws.amazon.com/about-aws/whats-new/2018/11/announcing-amazon-inferentia-machine-learning-inference-microchip/</p><p>http://www.analyticsindiamag.com/inference-chips-are-the-next-big-battlefield-for-nvidia-and-intel/</p><p>http://www.analyticsindiamag.com/inference-chips-are-the-next-big-battlefield-for-nvidia-and-intel/</p><p>84</p><p>• Alibaba: The company has created its own AI chip</p><p>company called Pingtouge.28 It also has plans to build a</p><p>quantum computer processor, which is based on qubits</p><p>(they represent subatomic particles like electrons and</p><p>photons).</p><p>• Tesla: Elon Musk has developed his own AI chip. It has 6</p><p>billion transistors and can process 36 trillion operations</p><p>per second.29</p><p>There are a variety of startups that are making a play for the AI chip market</p><p>as well. Among the leading companies is Untether AI, which is focused on</p><p>creating chips that boost the transfer speeds of data (this has been a particularly</p><p>difficult part of AI). In one of the company’s prototypes, this process was</p><p>more than 1,000 faster than a typical AI chip.30 Intel, along with other investors,</p><p>participated in a $13 million round of funding in 2019.</p><p>Now when it comes to AI chips, NVIDIA has the dominant market share. But</p><p>because of the importance of this technology, it seems inevitable that there</p><p>will be more and more offerings that will come to the market.</p><p>When toUse Deep Learning?</p><p>Because of the power of deep learning, there is the temptation to use this</p><p>technology first when creating an AI project. But this can be a big mistake.</p><p>Deep learning still has narrow use cases, such as for text, video, image, and</p><p>time-series datasets. There is also a need for large amounts of data and high-</p><p>powered computer systems.</p><p>Oh, and deep learning is better when outcomes can be quantified and verified.</p><p>To see why, let’s consider the following example. A team of researchers, led</p><p>by Thomas Hartung (a toxicologist at Johns Hopkins University), created a</p><p>dataset of about 10,000 chemicals that were based on 800,000 animal tests.</p><p>By using deep learning, the results showed that the model was more predictive</p><p>than many animal tests for toxicity.31 Remember that animal tests can not only</p><p>be costly and require safety measures but also have inconsistent results</p><p>because of repeated testing on the same chemical.</p><p>28 www.technologyreview.com/s/612190/why-alibaba-is-investing-in-ai-</p><p>chips-and-quantum-computing/</p><p>29 www.technologyreview.com/f/613403/tesla-says-its-new-self-driving-</p><p>chip-will-help-make-its-cars-autonomous/</p><p>30 www.technologyreview.com/f/613258/intel-buys-into-an-ai-chip-</p><p>that-can-transfer-data-1000-times-faster/</p><p>31 www.nature.com/articles/d41586-018-05664-2</p><p>Chapter 4 | Deep Learning</p><p>http://www.technologyreview.com/s/612190/why-alibaba-is-investing-in-ai-chips-and-quantum-computing/</p><p>http://www.technologyreview.com/s/612190/why-alibaba-is-investing-in-ai-chips-and-quantum-computing/</p><p>http://www.technologyreview.com/f/613403/tesla-says-its-new-self-driving-chip-will-help-make-its-cars-autonomous/</p><p>http://www.technologyreview.com/f/613403/tesla-says-its-new-self-driving-chip-will-help-make-its-cars-autonomous/</p><p>http://www.technologyreview.com/f/613258/intel-buys-into-an-ai-chip-that-can-transfer-data-1000-times-faster/</p><p>http://www.technologyreview.com/f/613258/intel-buys-into-an-ai-chip-that-can-transfer-data-1000-times-faster/</p><p>http://www.nature.com/articles/d41586-018-05664-2</p><p>85</p><p>“The first scenario illustrates the predictive power of deep learning, and its</p><p>ability to unearth correlations from large datasets that a human would never</p><p>find,” said Sheldon Fernandez, who is the CEO of DarwinAI.32</p><p>So where’s a scenario in which deep learning falls short? Actually, an illustration</p><p>of this is the 2018 FIFA World Cup in Russia, which France won. Many</p><p>researchers tried to predict the outcomes of all 64 matches, but the results</p><p>were far from accurate:33</p><p>• One group of researchers employed the bookmaker</p><p>consensus model that indicated that Brazil would win.</p><p>• Another group of researchers used algorithms such as</p><p>random forest and Poisson ranking to forecast that Spain</p><p>would prevail.</p><p>The problem here is that it is tough to find the right variables that have</p><p>predictive power. In fact, deep learning models are basically unable to handle</p><p>the complexity of features for certain events, especially those that have</p><p>elements of being chaotic.</p><p>However, even if you have the right amount of data and computing power, you still</p><p>need to hire people who have a background in deep learning, which is not easy.</p><p>Keep in mindthat it is a challenge to select the right model and fine-tune it.</p><p>How many hyperparameters should there be? What should be the number of</p><p>hidden layers? And how do you evaluate the model? All of these are highly</p><p>complex.</p><p>Even experts can get things wrong. Here’s the following from Sheldon:</p><p>One of our automotive clients encountered some bizarre behavior in</p><p>which a self-driving car would turn left with increasing regularity when</p><p>the sky was a certain shade of purple. After months of painful debugging,</p><p>they determined the training for certain turning scenarios had been</p><p>conducted in the Nevada desert when the sky was a particular hue.</p><p>Unbeknownst to its human designers, the neural network had established</p><p>a correlation between its turning behavior and the celestial tint.34</p><p>There are some tools that are helping with the deep learning process, such as</p><p>Amazon.com’s SageMaker, Google’s HyperTune, and SigOpt. But there is still</p><p>a long way to go.</p><p>If deep learning is not a fit, then you may want to consider machine learning,</p><p>which often requires relatively less data. Furthermore, the models tend to be</p><p>much simpler, but the results may still be more effective.</p><p>32 This is from the author’s interview with Sheldon Fernandez, the CEO of DarwinAI.</p><p>33 https://medium.com/futuristone/artificial-intelligence-failed-</p><p>in-world-cup-2018-6af10602206a</p><p>34 This is from the author’s interview with Sheldon Fernandez, the CEO of DarwinAI.</p><p>Artificial Intelligence Basics</p><p>https://medium.com/futuristone/artificial-intelligence-failed-in-world-cup-2018-6af10602206a</p><p>https://medium.com/futuristone/artificial-intelligence-failed-in-world-cup-2018-6af10602206a</p><p>86</p><p>Drawbacks withDeep Learning</p><p>Given all the innovations and breakthroughs, it’s reasonable that many people</p><p>consider deep learning to be a silver bullet. It will mean we no longer have to</p><p>drive a car. It may even mean that we’ll cure cancer.</p><p>How is it not possible to be excited and optimistic? This is natural and</p><p>reasonable. But it is important to note that deep learning is still in a nascent</p><p>stage and there are actually many nagging issues. It’s a good idea to temper</p><p>expectations.</p><p>In 2018, Gary Marcus wrote a paper entitled “Deep Learning: A Critical</p><p>Appraisal,” in which he clearly set forth the challenges.35 In his paper, he notes:</p><p>Against a background of considerable progress in areas such as speech</p><p>recognition, image recognition, and game playing, and considerable</p><p>enthusiasm in the popular press, I present ten concerns for deep learning,</p><p>and suggest that deep learning must be supplemented by other techniques</p><p>if we are to reach Artificial General Intelligence.36</p><p>Marcus definitely has the right pedigree to present his concerns, as he has</p><p>both an academic and business background in AI.Before becoming a professor</p><p>at the Department of Psychology at NewYork University, he sold his startup,</p><p>called Geometric Intelligence, to Uber. Marcus is also the author of several</p><p>bestselling books like The Haphazard Construction of the Human Mind.37</p><p>Here’s a look at some of his worries about deep learning:</p><p>• Black Box: A deep learning model could easily have</p><p>millions of parameters that involve many hidden layers.</p><p>Having a clear understanding of this is really beyond a</p><p>person’s capabilities. True, this may not necessarily be a</p><p>problem with recognizing cats in a dataset. But it could</p><p>definitely be an issue with models for medical diagnosis</p><p>or determining the safety of an oil rig. In these situations,</p><p>regulators will want to have a good understanding of the</p><p>transparency of the models. Because of this, researchers</p><p>are looking at creating systems to determine</p><p>“explainability,” which provides an understanding of deep</p><p>learning models.</p><p>35 Gary Marcus, “Deep Learning: A Critical Appraisal,” arXiv, 1801.00631v1 [cs.AI]:1–27,</p><p>2018.</p><p>36 https://arxiv.org/ftp/arxiv/papers/1801/1801.00631.pdf</p><p>37 Gary Marcus, Kluge: The Haphazard Construction of the Human Mind (Houghton Mifflin,</p><p>2008).</p><p>Chapter 4 | Deep Learning</p><p>https://arxiv.org/ftp/arxiv/papers/1801/1801.00631.pdf</p><p>87</p><p>• Data: The human brain has its flaws. But there are some</p><p>functions that it does extremely well like the ability to</p><p>learn by abstraction. For example, suppose Jan, who is</p><p>five years old, goes to a restaurant with her family. Her</p><p>mother points out an item on the plate and says it is a</p><p>“taco.” She does not have to explain it or provide any</p><p>information about it. Instead, Jan’s brain will instantly</p><p>process this information and understand the overall</p><p>pattern. In the future, when she sees another taco—even</p><p>if it has differences, such as with the dressing—she will</p><p>know what it is. For the most part, this is intuitive. But</p><p>unfortunately, when it comes to deep learning, there is</p><p>no taco learning by abstraction! The system has to</p><p>process enormous amounts of information to recognize</p><p>it. Of course, this is not a problem for companies like</p><p>Facebook, Google, or even Uber. But many companies</p><p>have much more limited datasets. The result is that deep</p><p>learning may not be a good option.</p><p>• Hierarchical Structure: This way of organizing does not</p><p>exist in deep learning. Because of this, language</p><p>understanding still has a long way to go (especially with</p><p>long discussions).</p><p>• Open-Ended Inference: Marcus notes that deep learning</p><p>cannot understand the nuances between “John promised</p><p>Mary to leave” and “John promised to leave Mary.” What’s</p><p>more, deep learning is far away from being able to, for</p><p>instance, read Jane Austen’s Pride and Prejudice and be</p><p>able to divine Elizabeth Bennet’s character motivations.</p><p>• Conceptual Thinking: Deep learning cannot have an</p><p>understanding of concepts like democracy, justice, or</p><p>happiness. It also does not have imagination, thinking of</p><p>new ideas or plans.</p><p>• Common Sense: This is something deep learning does not</p><p>do well. If anything, this means a model can be easily</p><p>confused. For example, let’s say you ask an AI system, “Is</p><p>it possible to make a computer with a sponge?” For the</p><p>most part, it will probably not</p><p>know that this is a</p><p>ridiculous question.</p><p>• Causation: Deep learning is unable to determine this. It’s</p><p>all about finding correlations.</p><p>Artificial Intelligence Basics</p><p>88</p><p>• Prior Knowledge: CNNs can help with some prior</p><p>information, but this is limited. Deep learning is still fairly</p><p>self-contained, as it only solves one problem at a time. It</p><p>cannot take in the data and create algorithms that span</p><p>various domains. In addition, a model does not adapt. If</p><p>there is change in the data, then a new model needs to</p><p>be trained and tested. And finally, deep learning does not</p><p>have prior understanding of what people know</p><p>instinctively—such as basic physics of the real world. This</p><p>is something that has to be explicitly programmed into an</p><p>AI system.</p><p>• Static: Deep learning works best in environments that</p><p>are fairly simple. This is why AI has been so effective</p><p>with board games, which have a clear set of rules and</p><p>boundaries. But the real world is chaotic and</p><p>unpredictable. This means that deep learning may fall</p><p>short with complex problems, even with self-driving</p><p>cars.</p><p>• Resources: A deep learning model often requires a</p><p>tremendous amount of CPU power, such as with GPUs.</p><p>This can get costly. Although, one option is to use a</p><p>third-party cloud service.</p><p>This is quite a lot? It’s true. But the paper still has left out some drawbacks.</p><p>Here are a couple other ones:</p><p>• Butterfly Effect: Because of the complexity of the data,</p><p>networks, and connections, a minute change can have a</p><p>major impact in the results of the deep learning model.</p><p>This could easily lead to conclusions that are wrong or</p><p>misleading.</p><p>• Overfitting: We explained this concept earlier in the</p><p>chapter.</p><p>As for Marcus, his biggest fear is that AI could “get trapped in a local minimum,</p><p>dwelling too heavily in the wrong part of intellectual space, focusing too much</p><p>on the detailed exploration of a particular class of accessible but limited</p><p>models that are geared around capturing low-hanging fruit—potentially</p><p>neglecting riskier excursions that might ultimately lead to a more robust</p><p>path.”</p><p>However, he is not a pessimist. He believes that researchers need to go</p><p>beyond deep learning and find new techniques that can solve tough problems.</p><p>Chapter 4 | Deep Learning</p><p>89</p><p>Conclusion</p><p>While Marcus has pointed out the flaws in deep learning, the fact is that this</p><p>AI approach is still extremely powerful. In less than a decade, it has</p><p>revolutionized the tech world—and is also significantly impacting areas like</p><p>finance, robotics, and healthcare.</p><p>With the surge in investments from large tech companies and VCs, there</p><p>will be further innovation with the models. This will also encourage</p><p>engineers to get postgraduate degrees, creating a virtuous cycle of</p><p>breakthroughs.</p><p>Key Takeaways</p><p>• Deep learning, which is a subfield of machine learning,</p><p>processes huge amounts of data to detect relationships</p><p>and patterns that humans are often unable to detect. The</p><p>word “deep” describes the number of hidden layers.</p><p>• An artificial neural network (ANN) is a function that</p><p>includes units that have weights and are used to predict</p><p>values in an AI model.</p><p>• A hidden layer is a part of a model that processes</p><p>incoming data.</p><p>• A feed-forward neural network has data that goes only</p><p>from input to the hidden layer to the output. The results</p><p>do not cycle back. Yet they can go into another neural</p><p>network.</p><p>• An activation function is non-linear. In other words, it</p><p>tends to do a better job of reflecting the real world.</p><p>• A sigmoid is an activation function that compresses the</p><p>input value into a range of 0–1, which makes it easier for</p><p>analysis.</p><p>• Backpropagation is a sophisticated technique to adjust</p><p>the weights in a neural network. This approach has been</p><p>critical for the growth in deep learning.</p><p>• A recurrent neural network (RNN) is a function that not</p><p>only processes the input but also prior inputs across</p><p>time.</p><p>Artificial Intelligence Basics</p><p>90</p><p>• A convolutional neural network (CNN) analyzes data</p><p>section by section (that is, by convolutions). This model</p><p>is geared for complex applications like image recognition.</p><p>• A generative adversarial network or GAN is where two</p><p>neural networks compete with each other in a tight</p><p>feedback loop. The result is often the creation of a new</p><p>object.</p><p>• Explainability describes techniques for transparency with</p><p>complex deep learning models.</p><p>Chapter 4 | Deep Learning</p><p>© Tom Taulli 2019</p><p>T. Taulli, Artif icial Intelligence Basics,</p><p>https://doi.org/10.1007/978-1-4842-5028-0_5</p><p>C H A P T E R</p><p>5</p><p>Robotic Process</p><p>Automation</p><p>(RPA)</p><p>An Easier Path to AI</p><p>By interacting with applications just as a human would, software robots</p><p>can open email attachments, complete e-forms, record and re-key data,</p><p>and perform other tasks that mimic human action.</p><p>—Kaushik Iyengar,</p><p>director of Digital Transformation and Optimization at AT&T1</p><p>Back in 2005, Daniel Dines and Marius Tirca founded UiPath, which was</p><p>located in Bucharest, Romania. The company focused mostly on providing</p><p>integration services for applications from Google, Microsoft, and IBM.But it</p><p>was a struggle as the company relied mostly on custom work for clients.</p><p>1 www2.deloitte.com/insights/us/en/focus/signals-for-strategists/cognitive-</p><p>enterprise-robotic-process-automation.html</p><p>http://www2.deloitte.com/insights/us/en/focus/signals-for-strategists/cognitive-enterprise-robotic-process-automation.html</p><p>http://www2.deloitte.com/insights/us/en/focus/signals-for-strategists/cognitive-enterprise-robotic-process-automation.html</p><p>92</p><p>By 2013, UiPath was close to being shut down. But the founders did not give</p><p>up as they saw this as an opportunity to rethink the business and find a new</p><p>opportunity.2 To this end, they started to build a platform for Robotic</p><p>Process Automation (RPA). The category, which had been around since</p><p>2000, was about automating routine and mundane tasks within a company.</p><p>Yet RPA was actually a backwater area in the tech world—as seen with the</p><p>slow growth rates. However, Dines and Tirca were convinced that they could</p><p>transform the industry. One of the key reasons: the rise of AI and the cloud.</p><p>The new strategy was spot-on, and growth took off. Dines and Tirca also</p><p>were aggressive with seeking funding, innovating its RPA platform, and</p><p>expanding into global markets.</p><p>By 2018, UiPath was considered the fastest-growing enterprise software</p><p>company—ever. The annual recurring revenue soared from $1 million to $100</p><p>million, with over 1,800 customers.3 The company had the most widely used</p><p>RPA system in the world.</p><p>UiPath attracted a total of $448 million in venture capital from marque firms</p><p>like CapitalG, Sequoia Capital, and Accel. The valuation was at $3 billion.</p><p>In light of all this, more RPA startups snagged significant funding as well.</p><p>Then again, the market is forecasted to see tremendous growth. Grand</p><p>View Research predicts that spending will hit $3.97 billion in the United</p><p>States by 2025.4</p><p>Interestingly enough, Forrester had this to say about the RPA trend:</p><p>Today’s most successful companies generally operate</p><p>with fewer employees than those of the past. Consider</p><p>that Kodak at its peak in 1973 employed 120,000, but</p><p>when Facebook bought Instagram in 2012, the photo-</p><p>sharing site employed only 13 workers. In 2019, we</p><p>predict that one in 10 startups—operating in a more</p><p>agile, lean, and scalable fashion— will look at the</p><p>world through the lens of tasks, not jobs, and will</p><p>build business models around automation-first</p><p>principles.5</p><p>2 http://business-review.eu/news/the-story-of-uipath-how-it-became-</p><p>romanias-first-unicorn-164248</p><p>3 www.uipath.com/newsroom/uipath-raises-225-million-series-c-led-by-</p><p>capitalg-and-sequoia</p><p>4 www.grandviewresearch.com/press-release/global-robotic-process-</p><p>automation-rpa-market</p><p>5 https://go.forrester.com/blogs/predictions-2019-automation-will-become-</p><p>central-to-business-strategy-and-operations/</p><p>Chapter 5 | Robotic Process Automation (RPA)</p><p>http://business-review.eu/news/the-story-of-uipath-how-it-became-romanias-first-unicorn-164248</p><p>http://business-review.eu/news/the-story-of-uipath-how-it-became-romanias-first-unicorn-164248</p><p>http://www.uipath.com/newsroom/uipath-raises-225-million-series-c-led-by-capitalg-and-sequoia</p><p>http://www.uipath.com/newsroom/uipath-raises-225-million-series-c-led-by-capitalg-and-sequoia</p><p>http://www.grandviewresearch.com/press-release/global-robotic-process-automation-rpa-market</p><p>http://www.grandviewresearch.com/press-release/global-robotic-process-automation-rpa-market</p><p>https://go.forrester.com/blogs/predictions-2019-automation-will-become-central-to-business-strategy-and-operations/</p><p>https://go.forrester.com/blogs/predictions-2019-automation-will-become-central-to-business-strategy-and-operations/</p><p>93</p><p>RPA is yet another area that has been supercharged with AI. If anything, it</p><p>could be the gateway for many companies because the implementation usually</p><p>does not take long or require heavy costs.</p><p>In this chapter, we’ll take a look at RPA and see how it could be a critical</p><p>driver for many companies.</p><p>What Is RPA?</p><p>The term Robotic Process Automation can be a bit confusing. The word</p><p>“robotic” does not mean physical robots (we’ll cover these in Chapter 7);</p><p>rather, it is about software-based robots or bots.</p><p>RPA allows you to use low-code visual drag-and-drop systems to automate</p><p>the workflow of a process. Some examples include the following:</p><p>• Inputting, changing, and tracking Human Resources (HR)</p><p>documents, contracts, and employee information</p><p>• Detecting issues with customer service and taking actions</p><p>to resolve the problems</p><p>• Processing an insurance claim</p><p>• Sending invoices</p><p>• Issuing refunds to customers</p><p>• Reconciling financial records</p><p>• Transferring data from one system to another</p><p>• Providing standard replies to customers</p><p>This is all done by having a bot replicate the workflows for an application, say</p><p>for an ERP (Enterprise Resource Planning) or CRM (Customer Relationship</p><p>Management) system. This may even be done with the RPA program recording</p><p>the steps from employees or with the use of OCR (optical character</p><p>recognition) technology to translate handwritten notes. Think of RPA as a</p><p>digital employee.</p><p>There are two flavors of this type of technology:</p><p>• Unattended RPA: This is a process that is completely</p><p>autonomous as the bot will run in the background. Now</p><p>this does not mean there is no human intervention.</p><p>There will still be intervention for exception management.</p><p>This is when the bot encounters something it does not</p><p>understand.</p><p>Artificial Intelligence Basics</p><p>94</p><p>• RDA (Robotic Desktop Automation): This is where RPA</p><p>helps an employee with a job or task. A common use</p><p>case is with a contact center. That is, when a call comes</p><p>in, the rep can use RDA to help find answers, send</p><p>messages, pull customer profile information, and get</p><p>insight on what to do next. The technology helps improve</p><p>or augment the efficiency of the worker.</p><p>Pros andCons ofRPA</p><p>Of course, quite a bit of time for a typical employee—in the back office—is</p><p>spent on routine tasks. But with RPA, companies can often get a strong ROI</p><p>(Return on Investment)—so long as the implementation is done right.</p><p>Here are some other advantages:</p><p>• Customer Satisfaction: RPA means minimal errors as well</p><p>as high speed. A bot also works 24/7. This means that</p><p>customer satisfaction scores—like the NPS (Net</p><p>Promoter Score)—should improve. Note that increasingly</p><p>more customers, such as from the Millennial generation,</p><p>prefer to deal with apps/web sites, not people! RPA also</p><p>means that reps will have more time to spend on value-</p><p>added tasks, instead of dealing with the tedious matters</p><p>that waste time.</p><p>• Scalability: Once a bot is created, it can be quickly</p><p>expanded to meet spikes in activity. This can be critical</p><p>for seasonal businesses like retailers.</p><p>• Compliance: For people, it’s tough to keep track of rules,</p><p>regulations, and laws. Even worse, they often change. But</p><p>with RPA, compliance is built into the process—and is</p><p>always followed. This can be a major benefit in terms of</p><p>avoiding legal problems and fines.</p><p>• Insights and Analytics: Next-generation RPA platforms</p><p>come equipped with sophisticated dashboards, which</p><p>focus on KPIs for your business. You can also set up alerts</p><p>if there are any problems.</p><p>• Legacy Systems: Older companies are often bogged down</p><p>with old IT systems, which makes it extremely tough to</p><p>pull off a digital transformation. But RPA software is able</p><p>to work fairly well with legacy IT environments.</p><p>Chapter 5 | Robotic Process Automation (RPA)</p><p>95</p><p>• Data: Because of the automation, the data is much</p><p>cleaner as there are minimal input errors. This means</p><p>that organizations will—over time—have more</p><p>accurate understandings of their businesses. The data</p><p>quality will also increase the likelihood of success of AI</p><p>implementations.</p><p>While all this is great, RPA still has its drawbacks. For example, if you have</p><p>current processes that are inefficient and you rush to implement the RPA</p><p>system, you will be essentially replicating a bad approach! This is why it is</p><p>critical to evaluate your workflows before implementing a system.</p><p>But there are certainly other potential landmines to note, such as the</p><p>following:</p><p>• Brittleness: RPA can easily break if there are changes in the</p><p>underlying applications. This could also be the case if there</p><p>are changes in procedures and regulations. It’s true that</p><p>newer systems are getting better at adapting and may also</p><p>leverage APIs. But RPA is not about hands-off activity.</p><p>• Virtualized Apps: This type of software, such as from</p><p>Citrix, can be difficult with RPA systems because they</p><p>cannot effectively capture the processes. The reason is</p><p>that the data is stored on an outside server and the</p><p>output is a snapshot on a monitor. But some companies</p><p>are using AI to solve the problem, such as UiPath. The</p><p>company has created a system, called “Pragmatic AI,”</p><p>which uses computer vision to interpret the screen</p><p>snapshots to record the processes.</p><p>• Specialization: Many RPA tools are for general-purpose</p><p>activities. But there may be areas that require</p><p>specialization, such as with finance. In this case, you may</p><p>look at a niche software app that can handle it.</p><p>• Testing: This is absolutely critical. You want to first sample</p><p>some transactions to make sure the system is working</p><p>correctly. After this, you can do a more extensive rollout</p><p>of the RPA system.</p><p>• Ownership: The temptation is to have IT own the RPA</p><p>implementation and management. But this is probably</p><p>not advisable. The reason? RPA systems are fairly low</p><p>tech. After all, they can be developed by nonprogrammers.</p><p>Because of this, the business managers are ideal for</p><p>owning the process since they can generally handle the</p><p>technical issues and also have a firmer grasp of the</p><p>employee workflows.</p><p>Artificial Intelligence Basics</p><p>96</p><p>• Resistance: Change is always difficult. With RPA, there</p><p>may be fears that the technology will displace jobs. This</p><p>means you need to have a clear set of messages, which</p><p>focus on the benefits of the technology. For example,</p><p>RPA will mean more time to focus on important matters,</p><p>which should make a person’s job more interesting and</p><p>meaningful.</p><p>What Can YouExpect fromRPA?</p><p>When it comes to RPA, the industry is still in the early phases. Yet there are</p><p>clear signs that the technology is making a big difference for many companies.</p><p>Take a look at the research report from Computer Economics Technology,</p><p>which included roughly 250 companies (they were across many industries and</p><p>had revenues that were from $20 million to over $1 billion). Of those that</p><p>implemented an RPA system, about half reported a positive return within 18</p><p>months of deployment. This is definitely standout for enterprise software,</p><p>which can be challenging in getting adoption.6</p><p>And to get a sense of the strategic importance of this technology, look to see</p><p>what the US Defense Department—which is engaged in over 500 AI</p><p>projects—is doing. Here’s what the agency’s Joint Artificial Intelligence Center</p><p>director, Air Force Lt. Gen. Jack Shanahan, had to say during a Congressional</p><p>hearing:</p><p>When you talk about smart automation, or in the</p><p>vernacular of the industry, Robotic Process Automation,</p><p>it’s not headline grabbing in terms of big AI projects,</p><p>but it may be where the most efficiencies can be found.</p><p>That’s the case if you read some of the dailies in</p><p>industry, whether it’s in medicine or finance, this is</p><p>where early gains are being realized in AI.Some of the</p><p>other projects we take on in the department are</p><p>probably years in the making in return on investment.7</p><p>Despite all this, there are still many failed RPA implementations as well. Ernst</p><p>& Young, for example, has received large amount of consulting business</p><p>because of this. Based on this experience, the failure rate for initial RPA</p><p>projects ranges from 30% to 50%.8</p><p>6 www.computereconomics.com/article.cfm?id=2633</p><p>7 https://federalnewsnetwork.com/artificial-intelligence/2019/03/</p><p>dod-laying-groundwork-for-multi-generational-effort-on-ai/</p><p>8 www.cmswire.com/information-management/why-rpa-implementation-</p><p>projects-fail/</p><p>Chapter 5 | Robotic Process Automation (RPA)</p><p>http://www.computereconomics.com/article.cfm?id=2633</p><p>https://federalnewsnetwork.com/artificial-intelligence/2019/03/dod-laying-groundwork-for-multi-generational-effort-on-ai/</p><p>https://federalnewsnetwork.com/artificial-intelligence/2019/03/dod-laying-groundwork-for-multi-generational-effort-on-ai/</p><p>http://www.cmswire.com/information-management/why-rpa-implementation-projects-fail/</p><p>http://www.cmswire.com/information-management/why-rpa-implementation-projects-fail/</p><p>97</p><p>But this is inevitable with any type of enterprise software system. Yet so far,</p><p>the problems appear mostly to be about planning, strategy, and expectations—</p><p>not the technology.</p><p>Another problem is that the hype surrounding RPA may be raising expectations</p><p>to excessive levels. This means that disappointment will be fairly common,</p><p>even if implementations are successful!</p><p>Of course, technologies are not cure-alls. And they require much time, effort,</p><p>and diligence to work.</p><p>How toImplement RPA</p><p>Then what are some steps to take for a successful RPA implementation?</p><p>There is no standard answer, but there are certainly some best practices emerging:</p><p>• Determine the right functions to automate.</p><p>• Assess the processes.</p><p>• Select the RPA vendor and deploy the software.</p><p>• Set in place a team to manage the RPA platform.</p><p>Let’s take a closer look at each of these.</p><p>Determine theRight Functions toAutomate</p><p>Yes, excessive automation at Tesla was a mistake. To be precise, my</p><p>mistake. Humans are underrated.</p><p>—Elon Musk, CEO of Tesla9</p><p>Even though RPA is powerful and can move the needle in a big way for a</p><p>company, the capabilities are still fairly limited. The technology essentially</p><p>makes the most sense for automating repetitive, structured, and routine</p><p>processes. This involves things like scheduling, inputting/transferring data,</p><p>following rules/workflows, cut and paste, filling out forms, and search. This</p><p>means that RPA can actually have a role in just about every department in an</p><p>organization.</p><p>Then where does this technology generally fail to deliver? Well, if a process</p><p>requires independent judgment, then RPA probably does not make sense. The</p><p>same goes for when the processes are subject to frequent change. In this situation,</p><p>you can spend lots of time with ongoing adjustments to the configurations.</p><p>9 https://twitter.com/elonmusk/status/984882630947753984?lang=en</p><p>Artificial Intelligence Basics</p><p>https://twitter.com/elonmusk/status/984882630947753984?lang=en</p><p>98</p><p>Once you establish a part of the business where the technology looks like a</p><p>good fit, there are a myriad of other considerations. In other words, you’ll</p><p>likely have more success with a project if you focus on the following:</p><p>• The areas of the business that have serious levels of</p><p>underperformance</p><p>• The processes that take up a high percentage of employee</p><p>time and involve high error rates</p><p>• The tasks that need more hiring when there are higher</p><p>volumes</p><p>• The areas that you are thinking of outsourcing</p><p>• A process that has a large number of steps and in which</p><p>there are various applications involved</p><p>Assess theProcesses</p><p>It’s common for a company to have many unwritten processes. And this is</p><p>fine. This approach allows for adaptability, which is what people are good at.</p><p>However, this is far from the case with a bot. To have a successful</p><p>implementation, you need to have a deep assessment of the processes. This</p><p>can actually take a while, and it may make sense to get outside consultants to</p><p>help out. They have the advantage of being more neutral and better able to</p><p>identify the weaknesses.</p><p>Some of the RPA vendors do have their own tools to help with analyzing</p><p>processes—which you should definitely use. There are also third-party</p><p>software providers that have their own offerings. One is Celonis, which</p><p>integrates with RPA platforms such as UiPath, Automation Anywhere, Blue</p><p>Prism, and others. The software performs essentially a digital MRI that</p><p>analyzes data, providing insights on how your processes really work. It will</p><p>also identify weakness and opportunities, such as to increase revenues,</p><p>improve customer satisfaction, and free up resources.</p><p>Regardless of what approach you take, it is critical that you formulate a clear-</p><p>cut plan that has the input from IT, higher management, and the departments</p><p>impacted. Also make sure to get analytics people involved, as there could be</p><p>opportunities for leveraging the data.</p><p>Select theRPA Vendor andDeploy theSoftware</p><p>By going through the first two steps, you’ll be in a very good position to</p><p>evaluate the different RPA systems. For example, if your main goal is to cut</p><p>staff, then you would look for software that is focused on unattended bots.</p><p>Chapter 5 | Robotic Process Automation (RPA)</p><p>99</p><p>Or, if you want to leverage data—such as for AI applications—then this will</p><p>lead to other types of RPA platforms.</p><p>Then, once you have selected one, you will begin the deployment. The good</p><p>news is that this can be relatively fast, say less than a month.</p><p>But as you go on to do further RPA projects, you may run into something</p><p>called automation fatigue. This is where the returns generally start to</p><p>deteriorate.</p><p>Think of it this way: When you start out, the focus will usually be on those</p><p>areas of the business that need automation the most, which means the ROI</p><p>will be significant. But over time, there will be a focus on tasks that are not as</p><p>amenable to automation, and it will likely take much more work to realize</p><p>even slight improvements.</p><p>Because of this, it is a good idea to temper expectations when engaging in a</p><p>widespread RPA transformation.</p><p>Set inPlace aTeam toManage theRPA Platform</p><p>Just because RPA provides a high degree of automation does not mean it</p><p>requires little management. Rather, the best approach is to put together a</p><p>team, which is often referred to as a center of excellence (CoE).</p><p>In order to make the best use of the CoE, you need to be clear on each</p><p>person’s responsibilities. For example, you should be able to answer the</p><p>following questions:</p><p>• What happens if there is a problem with a bot? At what</p><p>points should there be human intervention?</p><p>• Who is in charge of monitoring the RPA?</p><p>• Who is in charge of training?</p><p>• Who will have the role for the first line of support?</p><p>• Who is responsible for developing the bots?</p><p>For larger organizations, you might also want to expand the roles. You could</p><p>have an RPA champion, who would be the evangelist of the platform—for the</p><p>whole company. Or there could be an RPA change manager, who provides the</p><p>communication to help with adoption.</p><p>Finally, as the RPA implementation gets larger, a key goal should be to look at</p><p>how all the parts fit together. Like many other software systems, there is the</p><p>risk of sprawl across the organization—which can mean not getting higher</p><p>performance. This is where having a proactive CoE can make a</p><p>major positive</p><p>impact.</p><p>Artificial Intelligence Basics</p><p>100</p><p>RPA andAI</p><p>While still in the initial phases, AI is already making strides with RPA tools.</p><p>This is leading to the emergence of Cognitive Robotic Process Automation</p><p>(CRPA) software bots.</p><p>And this makes sense. After all, RPA is about optimizing processes and involves</p><p>large amounts of data. So vendors are starting to implement systems like</p><p>machine learning, deep learning, speech recognition, and Natural Language</p><p>Processing. Some of the leaders in the CRPA space include UiPath, Automation</p><p>Anywhere, Blue Prism, NICE Systems, and Kryon Systems.</p><p>For example, with Automation Anywhere, a bot can handle tasks such as</p><p>extracting invoices from emails, which involves sophisticated text processing.</p><p>The company also has prebuilt integrations with third-party AI services like</p><p>IBM Watson, AWS Machine Learning, and Google Cloud AI.10</p><p>“There has been a proliferation of AI-enabled services in recent years, but</p><p>businesses often struggle to operationalize them,” said Mukund Srigopal, who</p><p>is the director of Product Marketing at Automation Anywhere. “RPA is a</p><p>great way to infuse AI capabilities into business processes.”11</p><p>Here are some other ways CRPA can allow for AI functions:</p><p>• You can connect chatbots with your system, which will</p><p>allow for automated customer service (we’ll cover this</p><p>topic in Chapter 6).</p><p>• AI can find the right moment to send an email or alert.</p><p>• IVR (Interactive Voice Response) has gotten a bad</p><p>reputation over the years. Simply put, customers do not</p><p>like the hassle of going through multiple steps to solve a</p><p>problem. But with CRPA, you can use something called</p><p>Dynamic IVR. This personalizes the voice messages to</p><p>each customer, providing for a much better experience.</p><p>• NLP and text analysis can convert unstructured data into</p><p>structured data. This can make the CRPA more effective.</p><p>10 www.forbes.com/sites/tomtaulli/2019/02/02/what-you-need-to-know-</p><p>about-rpa-robotic-process-automation/</p><p>11 This is from the author’s interview with Mukund Srigopal, the director of Product</p><p>Marketing at Automation Anywhere.</p><p>Chapter 5 | Robotic Process Automation (RPA)</p><p>http://www.forbes.com/sites/tomtaulli/2019/02/02/what-you-need-to-know-about-rpa-robotic-process-automation/</p><p>http://www.forbes.com/sites/tomtaulli/2019/02/02/what-you-need-to-know-about-rpa-robotic-process-automation/</p><p>101</p><p>RPA intheReal World</p><p>To get a better sense of how RPA works and to understand the benefits,</p><p>here’s a look at a case study of Microsoft.12 Every year, the company pays</p><p>billions of dollars in royalties to game developers, partners, and content</p><p>creators. Yet the process was mostly manual, involving the sending of</p><p>thousands of statements—and yes, this was a big time waster for the company.</p><p>So the company selected Kyron for an RPA implementation. By doing an initial</p><p>process review, Microsoft realized that anywhere from 70% to 80% of the</p><p>statements were straightforward and could be easily automated. The rest</p><p>included exceptions that required human intervention, such as approvals.</p><p>With the RPA system, a visual detection algorithm could divvy up the</p><p>statements and find the exceptions. The setup was also fairly quick, taking</p><p>about 6 weeks.</p><p>As should be no surprise, the results had a material impact on the process.</p><p>For instance, a bot was able to take only 2.5 hours to complete 150 royalty</p><p>statements. By comparison, it would take 50 hours for employees to do this.</p><p>The bottom line: Microsoft achieved a 2,000% savings. There was also an</p><p>elimination of any rework from human error (which before was about 5% in</p><p>a given month).</p><p>Conclusion</p><p>As seen with the Microsoft case study, RPA can lead to major savings. But</p><p>there still needs to be diligent planning, so as to understand your processes.</p><p>For the most part, the focus should be on tasks that are manual and</p><p>repetitive—not those that rely heavily on judgment. Next, it is important to</p><p>setup a CoE to oversee the ongoing management of the automation, which</p><p>will help with handling exceptions, collecting data, and tracking KPIs.</p><p>RPA is also a great way to implement basic AI within an organization. Actually,</p><p>because there could be significant ROI, this may spur even more investment</p><p>in pursuing this technology.</p><p>12 www.kryonsystems.com/microsoft-case-study/</p><p>Artificial Intelligence Basics</p><p>http://www.kryonsystems.com/microsoft-case-study/</p><p>102</p><p>Key Takeaways</p><p>• Robotic Process Automation (RPA) allows you to use</p><p>low-code visual drag-and-drop systems to automate the</p><p>workflow of a process.</p><p>• Unattended RPA is when a process is completely</p><p>automated.</p><p>• RDA (Robotic Desktop Automation) is where RPA helps</p><p>an employee with a job or task.</p><p>• Some of the benefits of RPA include higher customer</p><p>satisfaction, lower error rates, improved compliance, and</p><p>easier integration with legacy systems.</p><p>• Some of the drawbacks of RPA include the difficulty with</p><p>adapting to changes in the underlying applications, issues</p><p>with virtualized apps, and resistance from employees.</p><p>• RPA tends to work best where you can automate</p><p>repetitive, structured, and routine processes, such as</p><p>scheduling, inputting/transferring data, and following</p><p>rules/workflows.</p><p>• When implementing an RPA solution, some of the steps</p><p>to consider include determining the functions to</p><p>automate, assessing the processes, selecting the RPA</p><p>vendor and deploying the software, and setting in place a</p><p>team to manage the platform.</p><p>• A center of excellence (CoE) is a team that manages an</p><p>RPA implementation.</p><p>• Cognitive Robotic Process Automation (CRPA) is an</p><p>emerging category of RPA that focuses on AI technologies.</p><p>Chapter 5 | Robotic Process Automation (RPA)</p><p>© Tom Taulli 2019</p><p>T. Taulli, Artif icial Intelligence Basics,</p><p>https://doi.org/10.1007/978-1-4842-5028-0_6</p><p>C H A P T E R</p><p>6</p><p>Natural</p><p>Language</p><p>Processing (NLP)</p><p>How Computers Talk</p><p>In 2014, Microsoft launched a chatbot—an AI system that communicates with</p><p>people—called Xiaoice. It was integrated into Tencent’s WeChat, the largest</p><p>social messaging service in China. Xiaoice performed quite well, getting to</p><p>40 million users within a few years.</p><p>In light of the success, Microsoft wanted to see if it could do something</p><p>similar in the US market. The company’s Bing and the Technology and Research</p><p>Group leveraged AI technologies to build a new chatbot: Tay. The developers</p><p>even enlisted the help of improvisational comedians to make the conversion</p><p>engaging and fun.</p><p>104</p><p>On March 23, 2016, Microsoft launched Tay on Twitter—and it was an</p><p>unmitigated disaster. The chatbot quickly spewed racist and sexist messages!</p><p>Here’s just one of the thousands of examples:</p><p>@TheBigBrebowski ricky gervais learned totalitarianism from adolf hitler, the</p><p>inventor of atheism1</p><p>Tay was a vivid illustration of Godwin’s Law. It reads as follows:the more an</p><p>online discussion continues, the higher are the odds that someone will bring</p><p>up Adolf Hitler or the Nazis.</p><p>So yes, Microsoft took down Tay within 24 hours and blogged an apology. In</p><p>it, the corporate vice president of Microsoft Healthcare, Peter Lee, wrote:</p><p>Looking ahead, we face some difficult—and yet exciting—</p><p>research challenges in AI design. AI systems feed off of</p><p>both positive and negative interactions with people. In</p><p>that sense, the challenges are just as much social as they</p><p>are technical. We will do everything possible to limit</p><p>technical exploits but also know we cannot fully predict</p><p>all possible human interactive misuses without learning</p><p>from mistakes. To do AI right, one needs to iterate with</p><p>many people and often in public forums. We must enter</p><p>each one with great caution and ultimately learn and</p><p>improve, step by step, and to do this without offending</p><p>people in the process. We will remain steadfast in our</p><p>efforts to learn from this and other experiences as we</p><p>work toward contributing to an Internet that represents</p><p>the best, not the worst, of humanity.2</p><p>A key to Tay was to repeat some of the content of the people asking</p><p>xi</p><p>As for this book, the goal is to provide actionable advice that can make a big</p><p>difference in your organization and career. Now you will not find deeply tech-</p><p>nical explanations, code snippets, or equations. Instead, Artif icial Intelligence</p><p>Basics is about answering the top-of-mind questions that managers have:</p><p>Where does AI make sense? What are the gotchas? How do you evaluate the</p><p>technology? What about starting an AI pilot?</p><p>This book also takes a real-world view of the technology. A big advantage I</p><p>have as a writer for Forbes.com and an advisor in the tech world is that I get</p><p>to talk to many talented people in the AI field—and this helps me to identify</p><p>what is really important in the industry. I also get to learn about case studies</p><p>and examples of what works.</p><p>This book is organized in a way to cover the main topics in AI—and you do</p><p>not have to read each chapter in order. Artif icial Intelligence Basics is meant to</p><p>be a handbook.</p><p>Here are brief descriptions of the chapters:</p><p>• Chapter 1—AI Foundations: This is an overview of the rich</p><p>history of AI, which goes back to the 1950s. You will</p><p>learn about brilliant researchers and computer scientists</p><p>like Alan Turing, John McCarthy, Marvin Minsky, and</p><p>Geoffrey Hinton. There will also be coverage of key con-</p><p>cepts like the Turing Test, which gauges if a machine has</p><p>achieved true AI.</p><p>• Chapter 2—Data: Data is the lifeblood of AI. It’s how</p><p>algorithms can find patterns and correlations to provide</p><p>insights. But there are landmines with data, such as qual-</p><p>ity and bias. This chapter provides a framework to work</p><p>with data in an AI project.</p><p>• Chapter 3—Machine Learning: This is a subset of AI and</p><p>involves traditional statistical techniques like regressions.</p><p>But in this chapter, we’ll also cover the advanced algo-</p><p>rithms, such as k-Nearest Neighbor (k-NN) and the</p><p>Naive Bayes Classifier. Besides this, there will be a look at</p><p>how to put together a machine learning model.</p><p>• Chapter 4—Deep Learning: This is another subset of AI</p><p>and is clearly the one that has seen much of the innova-</p><p>tion during the past decade. Deep learning is about using</p><p>neural networks to find patterns that mimic the brain. In</p><p>the chapter, we’ll take a look at the main algorithms like</p><p>recurrent neural networks (RNNs), convolutional neural</p><p>networks (CNNs), and generative adversarial networks</p><p>(GANs). There will also be explanations of key concepts</p><p>like backpropagation.</p><p>Introduction</p><p>xii</p><p>• Chapter 5—Robotic Process Automation: This uses systems</p><p>to automate repetitive processes, such as inputting data</p><p>in a Customer Relationship Management (CRM) system.</p><p>Robotic Process Automation (RPA) has seen tremendous</p><p>growth during the past few years because of the high ROI</p><p>(Return on Investment). The technology has also been an</p><p>introductory way for companies to implement AI.</p><p>• Chapter 6—Natural Language Processing (NLP): This form</p><p>of AI, which involves understanding conversations, is the</p><p>most ubiquitous as seen with Siri, Cortana, and Alexa.</p><p>But NLP systems, such as chatbots, have also become</p><p>critical in the corporate world. This chapter will show</p><p>ways to use this technology effectively and how to avoid</p><p>the tricky issues.</p><p>• Chapter 7—Physical Robots: AI is starting to have a major</p><p>impact on this industry. With deep learning, it is getting</p><p>easier for robots to understand their environments. In</p><p>this chapter, we’ll take a look at both consumer and</p><p>industrial robots, such as with a myriad of use cases.</p><p>• Chapter 8—Implementation of AI: We’ll take a step-by-step</p><p>approach to putting together an AI project, from the ini-</p><p>tial concept to the deployment. This chapter will also</p><p>cover the various tools like Python, TensorFlow, and</p><p>PyTorch.</p><p>• Chapter 9—The Future of AI: This chapter will cover some</p><p>of the biggest trends in AI like autonomous driving,</p><p>weaponization of AI, technological unemployment, drug</p><p>discovery, and regulation.</p><p>At the back of the book, you’ll also find an appendix of resources for further</p><p>study and a glossary of common terms related to AI.</p><p>Accompanying Material</p><p>Any updates will be provided on my site at www.Taulli.com.</p><p>Introduction</p><p>http://www.taulli.com/</p><p>© Tom Taulli 2019</p><p>T. Taulli, Artif icial Intelligence Basics,</p><p>https://doi.org/10.1007/978-1-4842-5028-0_1</p><p>C H A P T E R</p><p>1</p><p>AI Foundations</p><p>History Lessons</p><p>Artif icial intelligence would be the ultimate version of Google. The ultimate</p><p>search engine that would understand everything on the web. It would</p><p>understand exactly what you wanted, and it would give you the right</p><p>thing. We’re nowhere near doing that now. However, we can get</p><p>incrementally closer to that, and that is basically what we work on.</p><p>—Larry Page, the co-founder of Google Inc. and</p><p>CEO of Alphabet1</p><p>In Fredric Brown’s 1954 short story, “Answer,” all of the computers across</p><p>the 96 billion planets in the universe were connected into one super</p><p>machine. It was then asked, “Is there a God?” to which it answered, “Yes,</p><p>now there is a God.”</p><p>No doubt, Brown’s story was certainly clever—as well as a bit comical and</p><p>chilling! Science fiction has been a way for us to understand the implications</p><p>of new technologies, and artificial intelligence (AI) has been a major theme.</p><p>Some of the most memorable characters in science fiction involve androids or</p><p>computers that become self-aware, such as in Terminator, Blade Runner, 2001:</p><p>A Space Odyssey, and even Frankenstein.</p><p>But with the relentless pace of new technologies and innovation nowadays,</p><p>science fiction is starting to become real. We can now talk to our smartphones</p><p>and get answers; our social media accounts provide us with the content we’re</p><p>1 Founding CEO of Google Inc. The Academy of Achievement interview,</p><p>www.achievement.org, October 28, 2000.</p><p>http://www.achievement.org</p><p>2</p><p>interested in; our banking apps provide us with reminders; and on and on.</p><p>This personalized content creation almost seems magical but is quickly</p><p>becoming normal in our everyday lives.</p><p>To understand AI, it’s important to have a grounding in its rich history. You’ll</p><p>see how the development of this industry has been full of breakthroughs and</p><p>setbacks. There is also a cast of brilliant researchers and academics, like Alan</p><p>Turing, John McCarthy, Marvin Minsky, and Geoffrey Hinton, who pushed the</p><p>boundaries of the technology. But through it all, there was constant progress.</p><p>Let’s get started.</p><p>Alan Turing andtheTuring Test</p><p>Alan Turing is a towering figure in computer science and AI.He is often called</p><p>the “father of AI.”</p><p>In 1936, he wrote a paper called “On Computable Numbers.” In it, he set</p><p>forth the core concepts of a computer, which became known as the Turing</p><p>machine. Keep in mind that real computers would not be developed until</p><p>more than a decade later.</p><p>Yet it was his paper, called “Computing Machinery and Intelligence,” that</p><p>would become historic for AI.He focused on the concept of a machine that</p><p>was intelligent. But in order to do this, there had to be a way to measure it.</p><p>What is intelligence—at least for a machine?</p><p>This is where he came up with the famous “Turing Test.” It is essentially a</p><p>game with three players: two that are human and one that is a computer. The</p><p>evaluator, a human, asks open-ended questions of the other two (one human,</p><p>one computer) with the goal of determining which one is the human. If the</p><p>evaluator cannot make a determination, then it is presumed that the computer</p><p>is intelligent. Figure1-1 shows the basic workflow of the Turing Test.</p><p>Figure 1-1. The basic workflow of the Turing Test</p><p>Chapter 1 | AI Foundations</p><p>3</p><p>The genius of this concept is that there is no need to see if the machine</p><p>actually knows something, is self-aware, or even if it is correct. Rather, the</p><p>Turing Test indicates that a machine can process large amounts of information,</p><p>interpret speech, and communicate with humans.</p><p>Turing believed that it would actually not be until about the turn of the century</p><p>that a machine would pass his test. Yes, this was one</p><p>questions.</p><p>For the most part, this is a valid approach. As we saw in Chapter 1, this was</p><p>at the heart of the first chatbot, ELIZA.</p><p>But there also must be effective filters in place. This is especially the case</p><p>when a chatbot is used in a free-form platform like Twitter (or, for that matter,</p><p>in any real-world scenario).</p><p>However, failures like Tay are important. They allow us to learn and to evolve</p><p>the technology.</p><p>In this chapter, we’ll take a look at chatbots as well as Natural Language</p><p>Processing (NLP), which is a key part of how computers understand and</p><p>manipulate language. This is a subset of AI.</p><p>Let’s get started.</p><p>1 www.theverge.com/2016/3/24/11297050/tay-microsoft-chatbot-racist</p><p>2 https://blogs.microsoft.com/blog/2016/03/25/learning-tays-introduction/</p><p>Chapter 6 | Natural Language Processing (NLP)</p><p>http://www.theverge.com/2016/3/24/11297050/tay-microsoft-chatbot-racist</p><p>https://blogs.microsoft.com/blog/2016/03/25/learning-tays-introduction/</p><p>105</p><p>The Challenges ofNLP</p><p>As we saw in Chapter 1, language is the key to the Turing Test, which is meant</p><p>to validate AI.Language is also something that sets us apart from animals.</p><p>But this area of study is exceedingly complex. Here are just some of the</p><p>challenges with NLP:</p><p>• Language can often be ambiguous. We learn to speak in a</p><p>quick fashion and accentuate our meaning with nonverbal</p><p>cues, our tone, or reactions to the environment. For</p><p>example, if a golf ball is heading toward someone, you’ll</p><p>yell “Fore!” But an NLP system would likely not</p><p>understand this because it cannot process the context of</p><p>the situation.</p><p>• Language changes frequently as the world changes.</p><p>According to the Oxford English Dictionary, there were</p><p>more than 1,100 words, senses, and subentries in 2018</p><p>(in all, there are over 829,000)3. Some of the new entries</p><p>included mansplain and hangry.</p><p>• When we talk, we make grammar mistakes. But this is</p><p>usually not a problem as people have a great ability for</p><p>inference. But this is a major challenge for NLP as words</p><p>and phrases may have multiple meanings (this is called</p><p>polysemy). For example, noted AI researcher Geoffrey</p><p>Hinton likes to compare “recognize speech” and “wreck</p><p>a nice beach.”4</p><p>• Language has accents and dialects.</p><p>• The meaning of words can change based on, say, the use</p><p>of sarcasm or other emotional responses.</p><p>• Words can be vague. After all, what does it really mean</p><p>to be “late”?</p><p>• Many words have essentially the same meaning but</p><p>involve degrees of nuances.</p><p>• Conversations can be non-linear and have interruptions.</p><p>Despite all this, there have been great strides with NLP, as seen with apps like</p><p>Siri, Alexa, and Cortana. Much of the progress has also happened within the</p><p>last decade, driven by the power of deep learning.</p><p>3 https://wordcounter.io/blog/newest-words-added-to-the-dictionary-</p><p>in-2018/</p><p>4 www.deepinstinct.com/2019/04/16/applications-of-deep-learning/</p><p>Artificial Intelligence Basics</p><p>https://wordcounter.io/blog/newest-words-added-to-the-dictionary-in-2018/</p><p>https://wordcounter.io/blog/newest-words-added-to-the-dictionary-in-2018/</p><p>http://www.deepinstinct.com/2019/04/16/applications-of-deep-learning/</p><p>106</p><p>Now there can be confusion about human languages and computer languages.</p><p>Haven’t computers been able to understand languages like BASIC, C, and</p><p>C++ for years? This is definitely true. It’s also true that computer languages</p><p>have English words like if, then, let, and print.</p><p>But this type of language is very different from human language. Consider that</p><p>a computer language has a limited set of commands and strict logic. If you use</p><p>something incorrectly, this will result in a bug in the code—leading to a crash.</p><p>Yes, computer languages are very literal!</p><p>Understanding How AI Translates Language</p><p>Now as we saw in Chapter 1, NLP was an early focus for AI researchers. But</p><p>because of the limited computer power, the capabilities were quite weak. The</p><p>goal was to create rules to interpret words and sentences—which turned out</p><p>to be complex and not very scalable. In a way, NLP in the early years was</p><p>mostly like a computer language!</p><p>But over time, there evolved a general structure for it. This was critical since</p><p>NLP deals with unstructured data, which can be unpredictable and difficult to</p><p>interpret.</p><p>Here’s a general high-level look at the two key steps:</p><p>• Cleaning and Preprocessing the Text: This involves using</p><p>techniques like tokenization, stemming, and lemmatization</p><p>to parse the text.</p><p>• Language Understanding and Generation: This is definitely</p><p>the most intensive part of the process, which often uses</p><p>deep learning algorithms.</p><p>In the next few sections, we’ll look at the different steps in more detail.</p><p>Step #1—Cleaning andPreprocessing</p><p>Three things need to be done during the cleaning and preprocessing step:</p><p>tokenization, stemming, and lemmatization.</p><p>Tokenization</p><p>Before there can be NLP, the text must be parsed and segmented into various</p><p>parts—a process known as tokenization. For example, let’s say we have the</p><p>following sentence: “John ate four cupcakes.” You would then separate and</p><p>categorize each element. Figure6-1 illustrates this tokenization.</p><p>Chapter 6 | Natural Language Processing (NLP)</p><p>107</p><p>All in all, kind of easy? Kind of.</p><p>After tokenization, there will be normalization of the text. This will entail</p><p>converting some of the text so as to make it easier for analysis, such as by</p><p>changing the case to upper or lower, removing punctuation, and eliminating</p><p>contractions.</p><p>But this can easily lead to some problems. Suppose we have a sentence that</p><p>has “A.I.” Should we get rid of the periods? And if so, will the computer know</p><p>what “A I” means?</p><p>Probably not.</p><p>Interestingly enough, even the case of words can have a major impact on the</p><p>meaning. Just look at the difference between “fed” and the “Fed.” The Fed is</p><p>often another name for the Federal Reserve. Or, in another case, let’s suppose</p><p>we have “us” and “US.” Are we talking about the United States here?</p><p>Here are some of the other issues:</p><p>• White Space Problem: This is where two or more words</p><p>should be one token because the words form a compound</p><p>phrase. Some examples include “New York” and “Silicon</p><p>Valley.”</p><p>• Scientif ic Words and Phrases: It’s common for such words</p><p>to have hyphens, parentheses, and Greek letters. If you</p><p>strip out these characters, the system may not be able to</p><p>understand the meanings of the words and phrases.</p><p>Figure 6-1. Example of sentence tokenization</p><p>Artificial Intelligence Basics</p><p>108</p><p>• Messy Text: Let’s face it, many documents have grammar</p><p>and spelling errors.</p><p>• Sentence Splitting: Words like “Mr.” or “Mrs.” can</p><p>prematurely end a sentence because of the period.</p><p>• Non-important Words: There are ones that really add little</p><p>or no meaning to a sentence, like “the,” “a,” and “an.” To</p><p>remove these, you can use a simple Stop Words filter.</p><p>As you can see, it can be easy to mis-parse sentences (and in some languages,</p><p>like Chinese and Japanese, things can get even more difficult with the syntax).</p><p>But this can have far-ranging consequences. Since tokenization is generally the</p><p>first step, a couple errors can cascade through the whole NLP process.</p><p>Stemming</p><p>Stemming describes the process of reducing a word to its root (or lemma),</p><p>such as by removing affixes and suffixes. This has actually been effective for</p><p>search engines, which involve the use of clustering to come up with more</p><p>relevant results. With stemming, it’s possible to find more matches as the</p><p>word has a broader meaning and even to handle such things as spelling errors.</p><p>And when using an AI application, it can help improve the overall understanding.</p><p>Figure 6-2 shows an example of stemming.</p><p>Figure 6-2. Example of stemming</p><p>Chapter 6 | Natural Language Processing (NLP)</p><p>109</p><p>There are a variety of algorithms to stem words, many of which are fairly</p><p>simple. But they have mixed results. According to IBM:</p><p>The Porter algorithm, for example, will state that ‘universal’ has the</p><p>same stem as ‘university’ and ‘universities,’</p><p>an observation that may have</p><p>historical basis but is no longer semantically relevant. The Porter</p><p>stemmer also does not recognize that ‘theater’ and ‘theatre’ should</p><p>belong to the same stem class. For reasons such as these, Watson</p><p>Explorer Engine does not use the Porter stemmer as its English stemmer.5</p><p>In fact, IBM has created its own proprietary stemmer, and it allows for</p><p>significant customization.</p><p>Lemmatization</p><p>Lemmatization is similar to stemming. But instead of removing affixes or</p><p>prefixes, there is a focus on finding similar root words. An example is “better,”</p><p>which we could lemmatize to “good.” This works so long as the meaning</p><p>remains mostly the same. In our example, both are roughly similar, but “good”</p><p>has a clearer meaning. Lemmatization also may work with providing better</p><p>searches or language understanding, especially with translations.</p><p>Figure 6-3 shows an example of lemmatization.</p><p>5 www.ibm.com/support/knowledgecenter/SS8NLW_11.0.1/com.ibm.swg.im.</p><p>infosphere.dataexpl.engine.doc/c_correcting_stemming_errors.html</p><p>Figure 6-3. Example of lemmatization</p><p>Artificial Intelligence Basics</p><p>http://www.ibm.com/support/knowledgecenter/SS8NLW_11.0.1/com.ibm.swg.im.infosphere.dataexpl.engine.doc/c_correcting_stemming_errors.html</p><p>http://www.ibm.com/support/knowledgecenter/SS8NLW_11.0.1/com.ibm.swg.im.infosphere.dataexpl.engine.doc/c_correcting_stemming_errors.html</p><p>110</p><p>To effectively use lemmatization, the NLP system must understand the</p><p>meanings of the words and the context. In other words, this process usually</p><p>has better performance than stemming. On the other hand, it also means that</p><p>the algorithms are more complicated and there are higher levels of computing</p><p>power required.</p><p>Step #2—Understanding andGenerating Language</p><p>Once the text has been put into a format that computers can process, then</p><p>the NLP system must understand the overall meaning. For the most part, this</p><p>is the hardest part.</p><p>But over the years, researchers have developed a myriad of techniques to help</p><p>out, such as the following:</p><p>• Tagging Parts of Speech (POS): This goes through the text</p><p>and designates each word into its proper grammatical</p><p>form, say nouns, verbs, adverbs, etc. Think of it like an</p><p>automated version of your grade school English class!</p><p>What’s more, some POS systems have variations. Note</p><p>that a noun has singular nouns (NN), singular proper</p><p>nouns (NNP), and plural nouns (NNS).</p><p>• Chunking: The words will then be analyzed in terms of</p><p>phrases. For example, a noun phrase (NP) is a noun that</p><p>acts as the subject or object to a verb.</p><p>• Named Entity Recognition: This is identifying words that</p><p>represent locations, persons, and organizations.</p><p>• Topic Modelling: This looks for hidden patterns and clusters</p><p>in the text. One of the algorithms, called Latent Dirichlet</p><p>Allocation (LDA), is based on unsupervised learning</p><p>approaches. That is, there will be random topics assigned,</p><p>and then the computer will iterate to find matches.</p><p>For many of these processes, we can use deep learning models. They can be</p><p>extended to more areas of analysis—to allow for seamless language</p><p>understanding and generation. This is a process known as distributional</p><p>semantics.</p><p>With a convolutional neural network (CNN), which we learned about in</p><p>Chapter 4, you can find clusters of words that are translated into a feature</p><p>map. This has allowed for applications like language translation, speech</p><p>recognition, sentiment analysis, and Q&A. In fact, the model can even do</p><p>things like detect sarcasm!</p><p>Chapter 6 | Natural Language Processing (NLP)</p><p>111</p><p>Yet there are some problems with CNNs. For example, the model has</p><p>difficulties with text that has dependencies across large distances. But there</p><p>are some ways to handle this, such as with time-delayed neural networks</p><p>(TDNN) and dynamic convolutional neural networks (DCNN). These</p><p>methods have shown high performance in handling sequenced data. Although,</p><p>the model that has shown more success with this is the recurrent neural</p><p>network (RNN), as it memorizes data.</p><p>So far, we have been focused mostly on text analysis. But for there to be</p><p>sophisticated NLP, we also must build voice recognition systems. We’ll take a</p><p>look at this in the next section.</p><p>Voice Recognition</p><p>In 1952, Bell Labs created the first voice recognition system, called Audrey</p><p>(for Automatic Digit Recognition). It was able to recognize phonemes, which</p><p>are the most basic units of sounds in a language. English, for example, has 44.</p><p>Audrey could recognize the sound of a digit, from zero to nine. It was accurate</p><p>for the voice of the machine’s creator, HK Davis, about 90% of the time.6 And</p><p>for anyone else, it was 70% to 80% or so.</p><p>Audrey was a major feat, especially in light of the limited computing power</p><p>and memory available at the time. But the program also highlighted the major</p><p>challenges with voice recognition. When we speak, our sentences can be</p><p>complex and somewhat jumbled. We also generally talk fast—an average of</p><p>150 words per minute.</p><p>As a result, voice recognition systems improved at a glacially slow pace. In</p><p>1962, IBM’s Shoebox system could recognize only 16 words, 10 digits, and 6</p><p>mathematical commands.</p><p>It was not until the 1980s that there was significant progress in the technology.</p><p>The key breakthrough was the use of the hidden Markov model (HMM), which</p><p>was based on sophisticated statistics. For example, if you say the word “dog,”</p><p>there will be an analysis of the individual sounds d, o, and g. The HMM</p><p>algorithm will assign a score to each of these. Over time, the system will get</p><p>better at understanding the sounds and translate them into words.</p><p>While HMM was critical, it still was unable to effectively handle continuous</p><p>speech. For example, voice systems were based on template matching. This</p><p>involved translating sound waves into numbers, which was done by sampling. The</p><p>result was that the software would measure the frequency of the intervals and</p><p>store the results. But there had to be a close match. Because of this, the voice</p><p>input had to be quite clear and slow. There also had to be little background noise.</p><p>6 www.bbc.com/future/story/20170214-the-machines-that-learned-to-listen</p><p>Artificial Intelligence Basics</p><p>http://www.bbc.com/future/story/20170214-the-machines-that-learned-to-listen</p><p>112</p><p>But by the 1990s, software developers would make strides and come out with</p><p>commercial systems, such as Dragon Dictate, which could understand</p><p>thousands of words in continuous speech. However, adoption was still not</p><p>mainstream. Many people still found it easier to type into their computers and</p><p>use the mouse. Yet there were some professions, like medicine (a popular use</p><p>case with transcribing diagnosis of patients), where speech recognition found</p><p>high levels of usage.</p><p>With the emergence of machine learning and deep learning, voice systems</p><p>have rapidly become much more sophisticated and accurate. Some of the key</p><p>algorithms involve the use of the long short-term memory (LSTM), recurrent</p><p>neural networks, and deep feed-forward neural networks. Google would go</p><p>on to implement these approaches in Google Voice, which was available to</p><p>hundreds of millions of smartphone users. And of course, we’ve seen great</p><p>progress with other offerings like Siri, Alexa, and Cortana.</p><p>NLP intheReal World</p><p>For the most part, we have gone through the main parts of the NLP workflow.</p><p>Next, let’s take a look at the powerful applications of this technology.</p><p>Use Case: Improving Sales</p><p>Roy Raanani, who has a career in working with tech startups, thought that the</p><p>countless conversions that occur every day in business are mostly ignored.</p><p>Perhaps AI could transform this into an opportunity?</p><p>In 2015, he founded Chorus to use NLP to divine insights from conversations</p><p>from sales people. Raanani called this the Conversation Cloud, which records,</p><p>organizes, and transcribes calls—which are entered in a CRM (Customer</p><p>Relationship Management) system. Over time, the algorithms will start to</p><p>learn about best practices</p><p>and indicate how things can be improved.</p><p>But pulling this off has not been easy. According to a Chorus blog:</p><p>There are billions of ways to ask questions, raise</p><p>objections, set action items, challenge hypotheses,</p><p>etc. all of which need to be identified if sales patterns</p><p>are to be codified. Second, signals and patterns evolve:</p><p>new competitors, product names and features, and</p><p>industry-related terminology change over time, and</p><p>machine-learned models quickly become obsolete.7</p><p>7 https://blog.chorus.ai/a-taste-of-chorus-s-secret-sauce-how-our-</p><p>system-teaches-itself</p><p>Chapter 6 | Natural Language Processing (NLP)</p><p>https://blog.chorus.ai/a-taste-of-chorus-s-secret-sauce-how-our-system-teaches-itself</p><p>https://blog.chorus.ai/a-taste-of-chorus-s-secret-sauce-how-our-system-teaches-itself</p><p>113</p><p>For example, one of the difficulties—which can be easily overlooked—is how</p><p>to identify the parties who are talking (there are often more than three on a</p><p>call). Known as “speaker separation,” it is considered even more difficult than</p><p>speech recognition. Chorus has created a deep learning model that essentially</p><p>creates a “voice fingerprint”—which is based on clustering—for each speaker.</p><p>So after several years of R&D, the company was able to develop a system that</p><p>could analyze large amounts of conversations.</p><p>As a testament to this, look at one of Chorus’ customers, Housecall Pro,</p><p>which is a startup that sells mobile technologies for field service management.</p><p>Before adopting the software, the company would often create personalized</p><p>sales pitches for each lead. But unfortunately, it was unscalable and had mixed</p><p>results.</p><p>But with Chorus, the company was able to create an approach that did not</p><p>have much variation. The software made it possible to measure every word</p><p>and the impact on the sales conversions. Chorus also measured whether a</p><p>sales rep was on-script or not.</p><p>The outcome? The company was able to increase the win rate of the sales</p><p>organization by 10%.8</p><p>Use Case: Fighting Depression</p><p>Across the world, about 300 million people suffer from depression, according</p><p>to data from the World Health Organization.9 About 15% of adults will</p><p>experience some type of depression during their life.</p><p>This may go undiagnosed because of lack of healthcare services, which can</p><p>mean that a person’s situation could get much worse. Unfortunately, the</p><p>depression can lead to other problems.</p><p>But NLP may be able to improve the situation. A recent study from Stanford</p><p>used a machine learning model that processed 3D facial expressions and the</p><p>spoken language. As a result, the system was able to diagnose depression with</p><p>an average error rate of 3.67 when using the Patient Health Questionnaire</p><p>(PHQ) scale. The accuracy was even higher for more aggravated forms of</p><p>depression.</p><p>In the study, the researchers noted: “This technology could be deployed to</p><p>cell phones worldwide and facilitate low-cost universal access to mental</p><p>health care.”10</p><p>8 www.chorus.ai/case-studies/housecall/</p><p>9 www.verywellmind.com/depression-statistics-everyone-should-know-4159056</p><p>10 “Measuring Depression Symptom Severity from Spoken Language and 3D Facial Expressions,”</p><p>A Haque, M Guo, AS Miner, L Fei-Fei, presented at the NeurIPS 2018 Workshop on</p><p>Machine Learning for Health (ML4H), https://arxiv.org/abs/1811.08592.</p><p>Artificial Intelligence Basics</p><p>http://www.chorus.ai/case-studies/housecall/</p><p>http://www.verywellmind.com/depression-statistics-everyone-should-know-4159056</p><p>https://arxiv.org/abs/1811.08592</p><p>114</p><p>Use Case: Content Creation</p><p>In 2015, several tech veterans like Elon Musk, Peter Thiel, Reid Hoffman, and</p><p>Sam Altman launched OpenAI, with the support of a whopping $1 billion in</p><p>funding. Structured as a nonprofit, the goal was to develop an organization</p><p>with the goal “to advance digital intelligence in the way that is most likely to</p><p>benefit humanity as a whole, unconstrained by a need to generate financial</p><p>return.”11</p><p>One of the areas of research has been on NLP.To this end, the company</p><p>launched a model called GPT-2 in 2019, which was based on a dataset of</p><p>roughly eight million web pages. The focus was to create a system that could</p><p>predict the next word based on a group of text.</p><p>To illustrate this, OpenAI provided an experiment with the following text as</p><p>the input: “In a shocking finding, scientist discovered a herd of unicorns living</p><p>in a remote, previously unexplored valley, in the Andes Mountains. Even more</p><p>surprising to the researchers was the fact that the unicorns spoke perfect</p><p>English.”</p><p>From this, the algorithms created a convincing story that was 377 words in</p><p>length!</p><p>Granted, the researchers admitted that the storytelling was better for topics</p><p>that related more to the underlying data, on topics like Lord of the Rings and</p><p>even Brexit. As should be no surprise, GPT-2 demonstrated poor performance</p><p>for technical domains.</p><p>But the model was able to score high on several well-known evaluations of</p><p>reading comprehension. See Table6-1.12</p><p>11 https://openai.com/blog/introducing-openai/</p><p>12 https://openai.com/blog/better-language-models/</p><p>Table 6-1. Reading comprehension results</p><p>DataSet Prior Record for Accuracy GPT-2’s Accuracy</p><p>Winograd Schema Challenge 63.7% 70.70%</p><p>LAMBADA 59.23% 63.24%</p><p>Children’s Book Test Common Nouns 85.7% 93.30%</p><p>Children’s Book Test Named Entities 82.3% 89.05%</p><p>Chapter 6 | Natural Language Processing (NLP)</p><p>https://openai.com/blog/introducing-openai/</p><p>https://openai.com/blog/better-language-models/</p><p>115</p><p>Even though a typical human would score 90%+ on these tests, the</p><p>performance of GPT-2 is still impressive. It’s important to note that the model</p><p>used Google’s neural network innovation, called a Transformer, and</p><p>unsupervised learning.</p><p>In keeping with OpenAI’s mission, the organization decided not to release the</p><p>complete model. The fear was that it could lead to adverse consequences like</p><p>fake news, spoofed Amazon.com reviews, spam, and phishing scams.</p><p>Use Case: Body Language</p><p>Just focusing on language itself can be limiting. Body language is also something</p><p>that should be included in a sophisticated AI model.</p><p>This is something that Rana el Kaliouby has been thinking about for some</p><p>time. While growing up in Egypt, she earned her master’s degree in science</p><p>from the American University in Cairo and then went on to get her PhD in</p><p>computer science at Newnham College of the University of Cambridge. But</p><p>there was something that was very compelling to her: How can computers</p><p>detect human emotions?</p><p>However, in her academic circles, there was little interest. The consensus</p><p>view in the computer science community was that this topic was really not</p><p>useful.</p><p>But Rana was undeterred and teamed up with noted professor Rosalind Picard</p><p>to create innovative machine learning models (she wrote a pivotal book, called</p><p>Affective Computing, which looked at emotions and machines).13 Yet there also</p><p>had to be the use of other domains like neuroscience and psychology. A big</p><p>part of this was leveraging the pioneering work of Paul Ekman, who did</p><p>extensive research on human emotions based on a person’s facial muscles. He</p><p>found that there were six universal human emotions (wrath, grossness,</p><p>scaredness, joy, loneliness, and shock) that could be coded by 46 movements</p><p>called action units—all becoming a part of the Facial Action Coding System,</p><p>or FACS.</p><p>While at the MIT Media Lab, Rana developed an “emotional hearing aid,”</p><p>which was a wearable that allowed those people with autism to better interact</p><p>in social environments.14 The system would detect the emotions of people</p><p>and provide appropriate ways to react.</p><p>13 Rosalind W.Picard, Affective Computing (MIT Press).</p><p>14 www.newyorker.com/magazine/2015/01/19/know-feel</p><p>Artificial Intelligence Basics</p><p>http://www.newyorker.com/magazine/2015/01/19/know-feel</p><p>116</p><p>It was groundbreaking as the New York Times named it as one of the most</p><p>consequential innovations in 2006. But Rana’s system also caught the attention</p><p>of Madison Avenue. Simply put, the technology</p><p>could be an effective tool to</p><p>gauge an audience’s mood about a television commercial.</p><p>Then a couple years later, Rana launched Affectiva. The company quickly</p><p>grew and attracted substantial amounts of venture capital (in all, it has raised</p><p>$54.2 million).</p><p>Rana, who was once ignored, had now become one of the leaders in a</p><p>trendcalled “emotion-tracking AI.”</p><p>The flagship product for Affectiva is Affdex, which is a cloud-based platform for</p><p>testing audiences for video. About a quarter of the Fortune Global 500 use it.</p><p>But the company has developed another product, called Affectiva Automotive</p><p>AI, which is an in-cabin sensing system for a vehicle. Some of the capabilities</p><p>include the following:</p><p>• Monitoring for driver fatigue or distraction, which will</p><p>trigger an alert (say a vibration of the seat belt).</p><p>• Providing for a handoff to a semi-autonomous system if</p><p>the driver is not waking or is angry. There is even an</p><p>ability to provide route alternatives to lessen the potential</p><p>for road rage!</p><p>• Personalizing the content—say music—based on the</p><p>passenger’s emotions.</p><p>For all of these offerings, there are advanced deep learning systems that</p><p>process enormous amounts of features of a database that has more than 7.5</p><p>million faces. These models also account for cultural influences and</p><p>demographic differences—which is all done in real-time.</p><p>Voice Commerce</p><p>NLP-driven technologies like virtual assistants, chatbots, and smart speakers</p><p>are poised to have powerful business models—and may even disrupt markets</p><p>like e-commerce and marketing. We have already seen an early version of this</p><p>with Tencent’s WeChat franchise. The company, which was founded during</p><p>the heyday of the Internet boom in the late 1990s, started with a simple</p><p>PC-based messenger product called OICQ.But it was the introduction of</p><p>WeChat that was a game changer, which has since become China’s largest</p><p>social media platform with over 1 billion monthly active users.15</p><p>15 www.wsj.com/articles/iphones-toughest-rival-in-china-is-wechat-</p><p>a-messaging-app-1501412406</p><p>Chapter 6 | Natural Language Processing (NLP)</p><p>http://www.wsj.com/articles/iphones-toughest-rival-in-china-is-wechat-a-messaging-app-1501412406</p><p>http://www.wsj.com/articles/iphones-toughest-rival-in-china-is-wechat-a-messaging-app-1501412406</p><p>117</p><p>But this app is more than for exchanging messages and posting content.</p><p>WeChat has quickly morphed into an all-purpose virtual assistant, where you</p><p>can easily hail a ride-sharing service, make a payment at a local retailer, place</p><p>a reservation for a flight, or play a game. For example, the app accounts for</p><p>close to 35% of the entire usage time on smartphones in China on a monthly</p><p>basis. WeChat is also a major reason that the country has become increasingly</p><p>a cash-less society.</p><p>All this points to the power of an emerging category called voice commerce</p><p>(or v-commerce), where you can make purchases via chat or voice. It’s such a</p><p>critical trend that Facebook’s Mark Zuckerberg wrote a blog post,16 in early</p><p>2019, where he said the company would become more like…WeChat.</p><p>According to research from Juniper, the market for voice commerce is</p><p>forecasted to hit a whopping $80 billion by 2023.17 But in terms of the winners</p><p>in this market, it seems like a good bet that it will be those companies that</p><p>have large install bases of smart devices like Amazon, Apple, and Google. But</p><p>there will still be room for providers of next-generation NLP technologies.</p><p>OK then, how might these AI systems impact the marketing industry? Well,</p><p>to see how, there was an article in the Harvard Business Review, called</p><p>“Marketing in the Age of Alexa” by Niraj Dawar and Neil Bendle. In it, the</p><p>authors note that “AI assistants will transform how companies connect with</p><p>their customers. They’ll become the primary channel through which people</p><p>get information, goods and services, and marketing will turn into the batter</p><p>for their attention.”18</p><p>Thus, the growth in chatbots, digital assistants, and smart speakers could be</p><p>much bigger than the initial web-based e-commerce revolution. These</p><p>technologies have significant benefits for customers, such as convenience. It’s</p><p>easy to tell a device to buy something, and the machine will also learn about</p><p>your habits. So the next time you say you want to have a soft drink, the</p><p>computer will know what you are referring to.</p><p>But this may lead to a winners-take-all scenario. Ultimately, it seems like</p><p>consumers will use only one smart device for their shopping. In addition, for</p><p>brands that want to sell their goods, there will be a need to deeply understand</p><p>what customers really want, so as to become the preferred vendor within the</p><p>recommendation engine.</p><p>16 www.facebook.com/notes/mark-zuckerberg/a-privacy-focused-vision-</p><p>for-social-networking/10156700570096634/</p><p>17 https://voicebot.ai/2019/02/19/juniper-forecasts-80-billion-in-voice-</p><p>commerce-in-2023-or-10-per-assistant/</p><p>18 https://hbr.org/2018/05/marketing-in-the-age-of-alexa</p><p>Artificial Intelligence Basics</p><p>http://www.facebook.com/notes/mark-zuckerberg/a-privacy-focused-vision-for-social-networking/10156700570096634/</p><p>http://www.facebook.com/notes/mark-zuckerberg/a-privacy-focused-vision-for-social-networking/10156700570096634/</p><p>https://voicebot.ai/2019/02/19/juniper-forecasts-80-billion-in-voice-commerce-in-2023-or-10-per-assistant/</p><p>https://voicebot.ai/2019/02/19/juniper-forecasts-80-billion-in-voice-commerce-in-2023-or-10-per-assistant/</p><p>https://hbr.org/2018/05/marketing-in-the-age-of-alexa</p><p>118</p><p>Virtual Assistants</p><p>In 2003, as the United States was embroiled in wars in the Middle East, the</p><p>Defense Department was looking to invest in next-generation technologies</p><p>for the battlefield. One of the key initiatives was to build a sophisticated</p><p>virtual assistant, which could recognize spoken instructions. The Defense</p><p>Department budgeted $150 million for this and tasked the SRI (Stanford</p><p>Research Institute) Lab—based in Silicon Valley—to develop the application.19</p><p>Even though the lab was a nonprofit, it still was allowed to license its</p><p>technologies (like the inkjet printer) to startups.</p><p>And this is what happened with the virtual assistant. Some of the members of</p><p>SRI—Dag Kittlaus, Tom Gruber, and Adam Cheyer—called it Siri and started</p><p>their own company to capitalize on the opportunity. They founded the</p><p>operation in 2007, which was when Apple’s iPhone was launched.</p><p>But there had to be much more R&D to get the product to the point where</p><p>it could be useful for consumers. The founders had to develop a system to</p><p>handle real-time data, build a search engine for geographic information, and</p><p>build security for credit cards and personal data. But it was NLP that was the</p><p>toughest challenge.</p><p>In an interview, Cheyer noted:</p><p>The hardest technical challenge with Siri was dealing with the massive</p><p>amounts of ambiguity present in human language. Consider the phrase</p><p>‘book 4-star restaurant in Boston’ — seems very straightforward to</p><p>understand. Our prototype system could handle this easily. However,</p><p>when we loaded in tens of millions of business names and hundreds of</p><p>thousands of cities into the system as vocabulary (just about every word</p><p>in the English language is a business name), the number of candidate</p><p>interpretations went through the roof.20</p><p>But the team was able to solve the problems and turn Siri into a powerful</p><p>system, which was launched on Apple’s App Store in February 2010. “It’s the</p><p>most sophisticated voice recognition to appear on a smartphone yet,”</p><p>according to a review in Wired.com.21</p><p>Steve Jobs took notice and called the founders. Within a few days, they would</p><p>meet, and the discussions quickly led to an acquisition, which happened in late</p><p>April for more than $200 million.</p><p>However, Jobs thought there needed to be improvements to Siri. Because of</p><p>this, there was a re-release in 2011. This actually happened a day before Jobs</p><p>died.</p><p>19 www.huffingtonpost.com/2013/01/22/siri-do-engine-apple-iphone_n_</p><p>2499165.html</p><p>20 https://medium.com/swlh/the-story-behind-siri-fbeb109938b0</p><p>21 www.wired.com/2010/02/siri-voice-recognition-iphone/</p><p>Chapter 6 | Natural Language Processing (NLP)</p><p>http://www.huffingtonpost.com/2013/01/22/siri-do-engine-apple-iphone_n_2499165.html</p><p>http://www.huffingtonpost.com/2013/01/22/siri-do-engine-apple-iphone_n_2499165.html</p><p>https://medium.com/swlh/the-story-behind-siri-fbeb109938b0</p><p>http://www.wired.com/2010/02/siri-voice-recognition-iphone/</p><p>119</p><p>Fast forward to today, Siri has the largest market share position in the virtual</p><p>assistant market, with 48.6%. Google Assistant is at 28.7%, and Amazon.com’s</p><p>Alexa has 13.2%.22</p><p>According to the “Voice Assistant Consumer Adoption Report,” about 146.6</p><p>million people in the United States have tried virtual assistants on their</p><p>smartphones and over 50 million with smart speakers. But this only covers</p><p>part of the story. Voice technology is also becoming embedded into wearables,</p><p>headphones, and appliances.23</p><p>Here are some other interesting findings:</p><p>• Using voice to search for products outranked searches</p><p>for various entertainment options.</p><p>• When it comes to productivity, the most common use</p><p>cases for voice include making calls, sending emails, and</p><p>setting alarms.</p><p>• The most common use of voice on smartphones occurs</p><p>when a person is driving.</p><p>• Regarding complaints with voice assistants on</p><p>smartphones, the one with the highest percentage was</p><p>inconsistency in understanding requests. Again, this</p><p>points to the continuing challenges of NLP.</p><p>The growth potential for virtual assistants remains bright, and the category is</p><p>likely to be a key for the AI industry. Juniper Research forecasts that the</p><p>number of virtual assistants in use on a global basis will more than triple to</p><p>2.5 billion by 2023.24 The fastest category is actually expected to be smart</p><p>TVs. Yes, I guess we’ll be holding conversations with these devices!</p><p>Chatbots</p><p>There is often confusion between the differences between virtual assistants</p><p>and chatbots. Keep in mind that there is much overlap between the two. Both</p><p>use NLP to interpret language and perform tasks.</p><p>But there are still critical distinctions. For the most part, chatbots are focused</p><p>primarily for businesses, such as for customer support or sales functions.</p><p>22 www.businessinsider.com/siri-google-assistant-voice-market-</p><p>share-charts-2018-6</p><p>23 https://voicebot.ai/wp-content/uploads/2018/11/voice-assistant-consumer-</p><p>adoption-report-2018-voicebot.pdf</p><p>24 https://techcrunch.com/2019/02/12/report-voice-assistants-in-use-to-</p><p>triple-to-8-billion-by-2023/</p><p>Artificial Intelligence Basics</p><p>http://www.businessinsider.com/siri-google-assistant-voice-market-share-charts-2018-6</p><p>http://www.businessinsider.com/siri-google-assistant-voice-market-share-charts-2018-6</p><p>https://voicebot.ai/wp-content/uploads/2018/11/voice-assistant-consumer-adoption-report-2018-voicebot.pdf</p><p>https://voicebot.ai/wp-content/uploads/2018/11/voice-assistant-consumer-adoption-report-2018-voicebot.pdf</p><p>https://techcrunch.com/2019/02/12/report-voice-assistants-in-use-to-triple-to-8-billion-by-2023/</p><p>https://techcrunch.com/2019/02/12/report-voice-assistants-in-use-to-triple-to-8-billion-by-2023/</p><p>120</p><p>Virtual assistants, on the other hand, are geared for essentially everyone to</p><p>help with their daily activities.</p><p>As we saw in Chapter 1, the origins of chatbots go back to the 1960s with the</p><p>development of ELIZA.But it was not until the past decade or so that this</p><p>technology became useable at scale.</p><p>Here’s a sampling of interesting chatbots:</p><p>• Ushur: This is integrated in the enterprise systems for</p><p>insurance companies, allowing for the automation of</p><p>claims/bill processing and sales enablement. The software</p><p>has shown, on average, a reduction of 30% in service</p><p>center call volumes and a 90% customer response rate.25</p><p>The company built its own state-of-the-art linguistics</p><p>engine called LISA (this stands for Language Intelligence</p><p>Services Architecture). LISA includes NLP, NLU,</p><p>sentiment analysis, sarcasm detection, topic detection,</p><p>data extraction, and language translations. The technology</p><p>currently supports 60 languages, making it a useful</p><p>platform for global organizations.</p><p>• Mya: This is a chatbot that can engage in conversations in</p><p>the recruiting process. Like Ushur, this is also based on a</p><p>home-grown NLP technology. Some of the reasons for</p><p>this include having better communications but also</p><p>handling specific topics for hiring.26 Mya greatly reduces</p><p>time to interview and time to hire by eliminating major</p><p>bottlenecks.</p><p>• Jane.ai: This is a platform that mines data across an</p><p>organization’s applications and databases—say Salesforce.</p><p>com, Office, Slack, and Gmail—in order to make it much</p><p>easier to get answers, which are personalized. Note that</p><p>about 35% of an employee’s time is spent just trying to</p><p>find information! For example, a use case of Jane.ai is</p><p>USA Mortgage. The company used the technology, which</p><p>was integrated into Slack, to help brokers to look up</p><p>information for mortgage processing. The result is that</p><p>USA Mortgage has saved about 1,000 human labor hours</p><p>per month.27</p><p>25 The information came from the author’s interview with the CEO and co-founder of</p><p>Ushur, Simha Sadasiva.</p><p>26 The information came from the author’s interview with the CEO and co-founder of</p><p>Mya, Eyal Grayevsky.</p><p>27 The information came from the author’s interview with the CEO and co-founder of</p><p>Jane.ai, David Karandish.</p><p>Chapter 6 | Natural Language Processing (NLP)</p><p>121</p><p>Despite all this, chatbots have still had mixed results. For example, just</p><p>one of the problems is that it is difficult to program systems for specialized</p><p>domains.</p><p>Take a look at a study from UserTesting, which was based on the responses</p><p>from 500 consumers of healthcare chatbots. Some of the main takeaways</p><p>included: there remains lots of anxiety with chatbots, especially when handling</p><p>personal information, and the technology has problems with understanding</p><p>complex topics.28</p><p>So before deploying a chatbot, there are some factors to consider:</p><p>• Set Expectations: Do not overpromise with the capabilities</p><p>with chatbots. This will only set up your organization for</p><p>disappointment. For example, you should not pretend</p><p>that the chatbot is a human. This is a surefire way to</p><p>create bad experiences. As a result, you might want to</p><p>start off a chatbot conversation with “Hi, I’m a chatbot</p><p>here to help you with…”</p><p>• Automation: In some cases, a chatbot can handle the</p><p>whole process with a customer. But you should still have</p><p>people in the loop. “The goal for chatbots is not to</p><p>replace humans entirely, but to be the first line of defense,</p><p>so to speak,” said Antonio Cangiano, who is an AI</p><p>evangelist at IBM. “This can mean not only saving</p><p>companies money but also freeing up human agents</p><p>who’ll be able to spend more time on complex inquiries</p><p>that are escalated to them.”29</p><p>• Friction: As much as possible, try to find ways for the</p><p>chatbot to solve problems as quickly as possible. And this</p><p>may not necessarily be using a conversation. Instead,</p><p>providing a simple form to fill out could be a better</p><p>alternative, say to schedule a demo.</p><p>• Repetitive Processes: These are often ideal for chatbots.</p><p>Examples include authentication, order status, scheduling,</p><p>and simple change requests.</p><p>• Centralization: Make sure you integrate the data with your</p><p>chatbots. This will allow for more seamless experiences.</p><p>No doubt, customers quickly get annoyed if they have to</p><p>repeat information.</p><p>28 www.forbes.com/sites/bernardmarr/2019/02/11/7-amazing-examples-of-online-</p><p>chatbots-and-virtual-digital-assistants-in-practice/#32bb1084533e</p><p>29 This is from the author’s interview with Antonio Cangiano, who is an AI evangelist at IBM.</p><p>Artificial Intelligence Basics</p><p>http://www.forbes.com/sites/bernardmarr/2019/02/11/7-amazing-examples-of-online-chatbots-and-virtual-digital-assistants-in-practice/#32bb1084533e</p><p>http://www.forbes.com/sites/bernardmarr/2019/02/11/7-amazing-examples-of-online-chatbots-and-virtual-digital-assistants-in-practice/#32bb1084533e</p><p>122</p><p>• Personalize the</p><p>Experience: This is not easy but can yield</p><p>major benefits. Jonathan Taylor, who is the CTO of</p><p>Zoovu, has this example: “Purchasing a camera lens will</p><p>be different for every shopper. There are many variations</p><p>of lenses that perhaps a slightly informed shopper</p><p>understands—but the average consumer may not be as</p><p>informed. Providing an assistive chatbot to guide a</p><p>customer to the right lens can help provide the same</p><p>level of customer service as an in-store employee. The</p><p>assistive chatbot can ask the right questions, understanding</p><p>the goal of the customer to provide a personalized</p><p>product recommendation including ‘what kind of camera</p><p>do you already have,’ ‘why are you buying a new camera,’</p><p>and ‘what are you primarily trying to capture in your</p><p>photographs?’”30</p><p>• Data Analytics: It’s critical to monitor the feedback with a</p><p>chatbot. What’s the satisfaction? What’s the accuracy rate?</p><p>• Conversational Design and User Experience (UX): It’s</p><p>different than creating a web site or even a mobile app.</p><p>With a chatbot, you need to think about the user’s</p><p>personality, gender, and even cultural context. Moreover,</p><p>you must consider the “voice” of your company. “Rather</p><p>than creating mockups of a visual interface, think about</p><p>writing scripts and playing them out before to build it,”</p><p>said Gillian McCann, who is head of Cloud Engineering</p><p>and Artificial Intelligence at Workgrid Software.31</p><p>Even with the issues with chatbots, the technology is continuing to improve.</p><p>More importantly, these systems are likely to become an increasingly important</p><p>part of the AI industry. According to IDC, about $4.5 billion will be spent on</p><p>chatbots in 2019—which compares to a total of $35.8 billion estimated for AI</p><p>systems.32</p><p>Something else: A study from Juniper Research indicates that the cost savings</p><p>from chatbots are likely to be substantial. The firm predicts they will reach</p><p>$7.3 billion by 2023, up from a mere $209 million in 2019.33</p><p>30 This is from the author’s interview with Jonathan Taylor, who is the CTO of Zoovu.</p><p>31 This is from the author’s interview with Gillian McCann, who is the head of Cloud</p><p>Engineering and Artificial Intelligence at Workgrid Software.</p><p>32 www.twice.com/retailing/artificial-intelligence-retail-chatbots-</p><p>idc-spending</p><p>33 www.juniperresearch.com/press/press-releases/bank-cost-</p><p>savings-via-chatbots-to-reach</p><p>Chapter 6 | Natural Language Processing (NLP)</p><p>http://www.twice.com/retailing/artificial-intelligence-retail-chatbots-idc-spending</p><p>http://www.twice.com/retailing/artificial-intelligence-retail-chatbots-idc-spending</p><p>http://www.juniperresearch.com/press/press-releases/bank-cost-savings-via-chatbots-to-reach</p><p>http://www.juniperresearch.com/press/press-releases/bank-cost-savings-via-chatbots-to-reach</p><p>123</p><p>Future ofNLP</p><p>In 1947, Boris Katz was born in Moldova, which was part of the Soviet Union.</p><p>He would go on to graduate from Moscow State University, where he learned</p><p>about computers, and then left the country to the United States (with the</p><p>assistance of Senator Edward Kennedy).</p><p>He wasted little time with the opportunity. Besides writing more than 80</p><p>technical publications and receiving two US patents, he created the START</p><p>system that allowed for sophisticated NLP capabilities. It was actually the</p><p>basis for the first Q&A site on the Web in 1993. Yes, this was the forerunner</p><p>to breakout companies like Yahoo! and Google.</p><p>Boris’s innovations were also critical for IBM’s Watson, which is now at the</p><p>core of the company’s AI efforts. This computer, in 2011, would shock the</p><p>world when it beat two of the all-time champions of the popular game show</p><p>Jeopardy!</p><p>Despite all the progress with NLP, Boris is not satisfied. He believes we are</p><p>still in the early stages and lots more must be done to get true value. In an</p><p>interview with the MIT Technology Review, he said: “But on the other hand,</p><p>these programs [like Siri and Alexa] are so incredibly stupid. So there’s a</p><p>feeling of being proud and being almost embarrassed. You launch something</p><p>that people feel is intelligent, but it’s not even close.”34</p><p>This is not to imply he’s a pessimist. However, he still thinks there needs to</p><p>be a rethinking of NLP if it is to get to the point of “real intelligence.” To this</p><p>end, he believes researchers must look beyond pure computer science to</p><p>broad areas like neuroscience, cognitive science, and psychology. He also</p><p>thinks NLP systems must do a much better job of understanding the actions</p><p>in the real world.</p><p>Conclusion</p><p>For many people, the first interaction with NLP is with virtual assistants. Even</p><p>while the technology is far from perfect, it still is quite useful—especially for</p><p>answering questions or getting information, say about a nearby restaurant.</p><p>But NLP is also having a major impact in the business world. In the years</p><p>ahead, the technology will become increasingly important for e-commerce</p><p>and customer service—providing significant cost savings and allowing</p><p>employees to focus on more value-added activities.</p><p>34 www.technologyreview.com/s/612826/virtual-assistants-thinks-theyre-</p><p>doomed-without-a-new-ai-approach/</p><p>Artificial Intelligence Basics</p><p>https://www.technologyreview.com/s/612826/virtual-assistants-thinks-theyre-doomed-without-a-new-ai-approach/</p><p>https://www.technologyreview.com/s/612826/virtual-assistants-thinks-theyre-doomed-without-a-new-ai-approach/</p><p>124</p><p>True, there is still a long way to go because of the complexities of language.</p><p>But the progress continues to be rapid, especially with the help of next-</p><p>generation AI approaches like deep learning.</p><p>Key Takeaways</p><p>• Natural Language Processing (NLP) is the use of AI to</p><p>allow computers to understand people.</p><p>• A chatbot is an AI system that communicates with</p><p>people, say by voice or online chat.</p><p>• While there have been great strides in NLP, there is much</p><p>work to be done. Just some of the challenges include</p><p>ambiguity of language, nonverbal cues, different dialects</p><p>and accents, and changes to the language.</p><p>• The two main steps with NLP include cleaning/</p><p>preprocessing the text and using AI to understand and</p><p>generate language.</p><p>• Tokenization is where text is parsed and segmented into</p><p>various parts.</p><p>• With normalization, text is converted into a form that</p><p>makes it easier for analysis, such as by removing</p><p>punctuation or contractions.</p><p>• Stemming describes the process of reducing a word to its</p><p>root (or lemma), such as by removing affixes and suffixes.</p><p>• Similar to stemming, lemmatization involves finding</p><p>similar root words.</p><p>• For NLP to understand language, there are a variety of</p><p>approaches like tagging parts of speech (putting the text</p><p>in the grammatical form), chunking (processing text in</p><p>phrases), and topic modelling (finding hidden patterns</p><p>and clusters).</p><p>• A phoneme is the most basic unit of sound in a language.</p><p>Chapter 6 | Natural Language Processing (NLP)</p><p>© Tom Taulli 2019</p><p>T. Taulli, Artif icial Intelligence Basics,</p><p>https://doi.org/10.1007/978-1-4842-5028-0_7</p><p>C H A P T E R</p><p>7</p><p>Physical Robots</p><p>The Ultimate Manifestation of AI</p><p>In the city of Pasadena, I went to CaliBurger for lunch and noticed a crowd of</p><p>people next to the area where the food was being cooked—which was behind</p><p>glass. The people were taking photos with their smartphones!</p><p>Why? The reason was Flippy, an AI-powered robot that can cook burgers.</p><p>I was there at the restaurant with David Zito, the CEO and co-founder of the</p><p>company Miso Robotics that built the system. “Flippy helps improve the</p><p>quality of the food because of the consistency and reduces production costs,”</p><p>he said. “We also built the robot to be in strict compliance with regulatory</p><p>standards.”1</p><p>After lunch, I walked over to the lab for Miso Robotics, which included a</p><p>testing center with sample robots. It was here that I saw the convergence of</p><p>software AI systems and physical robots. The engineers were building Flippy’s</p><p>brain, which was uploaded to the cloud. Just some of the capabilities included</p><p>washing down utensils and the grill, learning to adapt</p><p>to problems with</p><p>cooking, switching between a spatula for raw meat and one for cooked meat,</p><p>and placing baskets in the fryer. All this was being done in real-time.</p><p>But the food service industry is just one of the many areas that will be greatly</p><p>impacted by robotics and AI.</p><p>1 This is based on the author’s interview, in January 2019, with David Zito, who is the CEO</p><p>and co-founder of Miso Robotics.</p><p>126</p><p>According to International Data Corporation (IDC), the spending on robotics</p><p>and drones is forecasted to go from $115.7 billion in 2019 to $210.3 billion</p><p>by 2022.2 This represents a compound annual growth rate of 20.2%. About</p><p>two thirds of the spending will be for hardware systems.</p><p>In this chapter, we’ll take a look at physical robots and how AI will transform</p><p>the industry.</p><p>What Is aRobot?</p><p>The origins of the word “robot” go back to 1921in a play by Karel Capek</p><p>called Rossum’s Universal Robots. It’s about a factory that created robots from</p><p>organic matter, and yes, they were hostile! They would eventually join together</p><p>to rebel against their human masters (consider that “robot” comes from the</p><p>Czech word robata for forced labor).</p><p>But as of today, what is a good definition for this type of system? Keep in mind</p><p>that there are many variations, as robots can have a myriad of forms and</p><p>functions.</p><p>But we can boil them down into a few key parts:</p><p>• Physical: A robot can range in size, from tiny machines</p><p>that can explore our body to massive industrial systems</p><p>to flying machines to underwater vessels. There also</p><p>needs to be some type of energy source, like a battery,</p><p>electricity, or solar.</p><p>• Act: Simply enough, a robot must be able to take certain</p><p>actions. This could include moving an item or even talking.</p><p>• Sense: In order to act, a robot must understand its</p><p>environment. This is possible with sensors and feedback</p><p>systems.</p><p>• Intelligence: This does not mean full-on AI capabilities. Yet a</p><p>robot needs to be able to be programmed to take actions.</p><p>Nowadays it’s not too difficult to create a robot from scratch. For example,</p><p>RobotShop.com has hundreds of kits that range from under $10 to as much</p><p>as $35,750.00 (this is the Dr. Robot Jaguar V6 Tracked Mobile Platform).</p><p>A heart-warming story of the ingenuity of building robots concerns a 2-year</p><p>old, Cillian Jackson. He was born with a rare genetic condition that rendered</p><p>him immobile. His parents tried to get reimbursement for a special electric</p><p>wheelchair but were denied.</p><p>2 www.idc.com/getdoc.jsp?containerId=prUS44505618</p><p>Chapter 7 | Physical Robots</p><p>http://www.idc.com/getdoc.jsp?containerId=prUS44505618</p><p>127</p><p>Well, the students at Farmington High School took action and built a system</p><p>for Cillian.3 Essentially, it was a robot wheelchair, and it took only a month</p><p>to finish. Because of this, Cillian can now chase around his two corgis around</p><p>the house!</p><p>While above we looked at the features of robots, there are key interactions</p><p>to consider too:</p><p>• Sensors: The typical sensor is a camera or a Lidar (light</p><p>detection and ranging), which uses a laser scanner to</p><p>create 3D images. But robots might also have systems for</p><p>sound, touch, taste, and even smell. In fact, they could</p><p>also include sensors that go beyond human capabilities,</p><p>such as night vision or detecting chemicals. The</p><p>information from the sensors is sent to a controller that</p><p>can activate an arm or other parts of the robot.</p><p>• Actuators: These are electro-mechanical devices like</p><p>motors. For the most part, they help with the movement</p><p>of the arms, legs, head, and any other movable part.</p><p>• Computer: There are memory storage and processors to</p><p>help with the inputs from the sensors. In advanced robots,</p><p>there may also be AI chips or Internet connections to AI</p><p>cloud platforms.</p><p>Figure 7-1 shows the interactions of these functions.</p><p>3 www.nytimes.com/2019/04/03/us/robotics-wheelchair.html</p><p>Figure 7-1. The general system for a physical robot</p><p>Artificial Intelligence Basics</p><p>http://www.nytimes.com/2019/04/03/us/robotics-wheelchair.html</p><p>128</p><p>There are also two main ways to operate a robot. First of all, there is remote</p><p>control by a human operation. In this case, the robot is called a telerobot.</p><p>Then there is the autonomous robot, which uses its own abilities to navigate—</p><p>such as with AI.</p><p>So what was the first mobile, thinking robot? It was Shakey. The name was</p><p>apt, as the project manager of the system, Charles Rosen, noted: “We worked</p><p>for a month trying to find a good name for it, ranging from Greek names to</p><p>whatnot, and then one of us said, ‘Hey, it shakes like hell and moves around,</p><p>let’s just call it Shakey.’”4</p><p>The Stanford Research Institute (SRI), with funding from DARPA, worked on</p><p>Shakey from 1966 to 1972. And it was quite sophisticated for the era. Shakey</p><p>was large, at over five feet tall, and had wheels to move and sensors and</p><p>cameras to help with touching. It was also wirelessly connected to DEC PDP-</p><p>10 and PDP-15 computers. From here, a person could enter commands via</p><p>teletype. Although, Shakey used algorithms to navigate its environment, even</p><p>closing doors.</p><p>The development of the robot was the result of a myriad of AI breakthroughs.</p><p>For example, Nils Nilsson and Richard Fikes created STRIPS (Stanford</p><p>Research Institute Problem Solver), which allowed for automated planning as</p><p>well as the A∗ algorithm for finding the shortest path with the least amount</p><p>of computer resources.5</p><p>By the late 1960s, as America was focused on the space program, Shakey got</p><p>quite a bit of buzz. A flattering piece in Life declared that the robot was the</p><p>“first electronic person.”6</p><p>But unfortunately, in 1972, as the AI winter took hold, DARPA pulled the</p><p>funding on Shakey. Yet the robot would still remain a key part of tech history</p><p>and was inducted into the Robot Hall of Fame in 2004.7</p><p>Industrial andCommercial Robots</p><p>The first real-world use of robots had to do with manufacturing industries.</p><p>But these systems did take quite a while to get adoption.</p><p>4 www.computerhistory.org/revolution/artificial-intelligence-robotics/</p><p>13/289</p><p>5 https://spectrum.ieee.org/view-from-the-valley/tech-history/space-age/</p><p>sri-shakey-robot-honored-as-ieee-milestone</p><p>6 www.sri.com/work/timeline-innovation/timeline.php?timeline=computing-</p><p>digital#!&innovation=shakey-the-robot</p><p>7 www.wired.com/2013/09/tech-time-warp-shakey-robot/</p><p>Chapter 7 | Physical Robots</p><p>http://www.computerhistory.org/revolution/artificial-intelligence-robotics/13/289</p><p>http://www.computerhistory.org/revolution/artificial-intelligence-robotics/13/289</p><p>https://spectrum.ieee.org/view-from-the-valley/tech-history/space-age/sri-shakey-robot-honored-as-ieee-milestone</p><p>https://spectrum.ieee.org/view-from-the-valley/tech-history/space-age/sri-shakey-robot-honored-as-ieee-milestone</p><p>http://www.sri.com/work/timeline-innovation/timeline.php?timeline=computing-digital#!&innovation=shakey-the-robot</p><p>http://www.sri.com/work/timeline-innovation/timeline.php?timeline=computing-digital#!&innovation=shakey-the-robot</p><p>http://www.wired.com/2013/09/tech-time-warp-shakey-robot/</p><p>129</p><p>The story begins with George Devol, an inventor who did not finish high</p><p>school. But this was not a problem. Devol had a knack for engineering and</p><p>creativity, as he would go on to create some of the core systems for microwave</p><p>ovens, barcodes, and automatic doors (during his life, he would obtain over</p><p>40 patents).</p><p>It was during the early 1950s that he also received a patent on a programmable</p><p>robot called “Unimate.” He struggled to get interest in his idea as every</p><p>investor turned him down.</p><p>However, in 1957, his life would change forever when he met Joseph</p><p>Engelberger at a cocktail party. Think of it like when Steve Jobs met Steve</p><p>Wozniak to create the Apple computer.</p><p>Engelberger was an engineer but also a savvy businessman. He even had a love</p><p>for reading science fiction, such as Isaac Asimov’s stories. Because of this,</p><p>Engelberger wanted the Unimate to benefit society.</p><p>Yet there was still resistance—as many people thought the idea was unrealistic</p><p>and,</p><p>well, science fiction—and it took a year to get funding. But once</p><p>Engelberger did, he wasted little time in building the robot and was able to sell</p><p>it to General Motors (GM) in 1961. Unimate was bulky (weighing 2,700</p><p>pounds) and had one 7-foot arm, but it was still quite useful and also meant</p><p>that people would not have to do inherently dangerous activities. Some of its</p><p>core functions included welding, spraying, and gripping—all done accurately</p><p>and on a 24/7 basis.</p><p>Engelberger looked for creative ways to evangelize his robot. To this end, he</p><p>appeared on Johnny Carson’s The Tonight Show in 1966, in which Unimate</p><p>putted a golf ball perfectly and even poured beer. Johnny quipped that the</p><p>machine could “replace someone’s job.”8</p><p>But industrial robots did have their nagging issues. Interestingly enough, GM</p><p>learned this the hard way during the 1980s. At the time, CEO Roger Smith</p><p>promoted the vision of a “lights out” factory—that is, where robots could</p><p>build cars in the dark!</p><p>He went on to shell out a whopping $90 billion on the program and even</p><p>created a joint venture, with Fujitsu-Fanuc, called GMF Robotics. The</p><p>organization would become the world’s largest manufacturer of robots.</p><p>But unfortunately, the venture turned out to be a disaster. Besides aggravating</p><p>unions, the robots often failed to live up to expectations. Just some of the</p><p>fiascos included robots that welded doors shut or painted themselves—not</p><p>the cars!</p><p>8 www.theatlantic.com/technology/archive/2011/08/unimate-robot-on-johnny-</p><p>carsons-tonight-show-1966/469779/</p><p>Artificial Intelligence Basics</p><p>http://www.theatlantic.com/technology/archive/2011/08/unimate-robot-on-johnny-carsons-tonight-show-1966/469779/</p><p>http://www.theatlantic.com/technology/archive/2011/08/unimate-robot-on-johnny-carsons-tonight-show-1966/469779/</p><p>130</p><p>However, the situation of GMF is nothing really new—and it’s not necessarily</p><p>about misguided senior managers. Take alook at Tesla, which is one of the</p><p>world’s most innovative companies. But CEO Elon Musk still suffered major</p><p>issues with robots on his factory floors. The problems got so bad that Tesla’s</p><p>existence was jeopardized.</p><p>In an interview on CBS This Morning in April 2018, Musk said he used too many</p><p>robots when manufacturing the Model 3 and this actually slowed down the</p><p>process.9 He noted that he should have had more people involved.</p><p>All this points to what Hans Moravec once wrote: “It is comparatively easy to</p><p>make computers exhibit adult level performance on intelligence tests or</p><p>playing checkers, and difficult or impossible to give them the skills of a one-</p><p>year-old when it comes to perception and mobility.”10 This is often called the</p><p>Moravec paradox.</p><p>Regardless of all this, industrial robots have become a massive industry,</p><p>expanding across diverse segments like consumer goods, biotechnology/</p><p>healthcare, and plastics. As of 2018, there were 35,880 industrial and</p><p>commercial robots shipped in North America, according to data from the</p><p>Robotic Industries Association (RIA).11 For example, the auto industry</p><p>accounted for about 53%, but this has been declining.</p><p>Jeff Burnstein, president of the Association for Advancing Automation, had</p><p>this to say:</p><p>And as we’ve heard from our members and at shows such as Automate,</p><p>these sales and shipments aren’t just to large, multinational companies</p><p>anymore. Small and medium-sized companies are using robots to solve</p><p>real-world challenges, which is helping them be more competitive on a</p><p>global scale.12</p><p>At the same time, the costs of manufacturing industrial robots continue to</p><p>drop. Based on research from ARK, there will be a 65% reduction by 2025—</p><p>with devices averaging less than $11,000 each.13 The analysis is based on</p><p>Wright’s Law, which states that for every cumulative doubling in the number</p><p>of units produced, there is a consistent decline in costs in percentage terms.</p><p>OK then, what about AI and robots? Where is that status of the technology?</p><p>Even with the breakthroughs with deep learning, there has generally been</p><p>slow progress with using AI with robots. Part of this is due to the fact that</p><p>much of the research has been focused on software-based models, such as</p><p>9 www.theverge.com/2018/4/13/17234296/tesla-model-3-robots-production-</p><p>hell-elon-musk</p><p>10 www.graphcore.ai/posts/is-moravecs-paradox-still-relevant-for-ai-today</p><p>11 www.apnews.com/b399fa71204d47199fdf4c753102e6c7</p><p>12 www.apnews.com/b399fa71204d47199fdf4c753102e6c7</p><p>13 https://ark-invest.com/research/industrial-robot-costs</p><p>Chapter 7 | Physical Robots</p><p>http://www.theverge.com/2018/4/13/17234296/tesla-model-3-robots-production-hell-elon-musk</p><p>http://www.theverge.com/2018/4/13/17234296/tesla-model-3-robots-production-hell-elon-musk</p><p>http://www.graphcore.ai/posts/is-moravecs-paradox-still-relevant-for-ai-today</p><p>http://www.apnews.com/b399fa71204d47199fdf4c753102e6c7</p><p>http://www.apnews.com/b399fa71204d47199fdf4c753102e6c7</p><p>https://ark-invest.com/research/industrial-robot-costs</p><p>131</p><p>with image recognition. But another reason is that physical robots require</p><p>sophisticated technologies to understand the environment—which is often</p><p>noisy and distracting—in real-time. This involves enabling simultaneous</p><p>localization and mapping (SLAM) in unknown environments while</p><p>simultaneously tracking the robot’s location. To do this effectively, there may</p><p>even need to be new technologies created, such as better neural network</p><p>algorithms and quantum computers.</p><p>Despite all this, there is certainly progress being made, especially with the use</p><p>of reinforcement learning techniques. Consider some of the following</p><p>innovations:</p><p>• Osaro: The company develops systems that allow robots</p><p>to learn quickly. Osaro describes this as “the ability to</p><p>mimic behavior that requires learned sensor fusion as</p><p>well as high level planning and object manipulation. It will</p><p>also enable the ability to learn from one machine to</p><p>another and improve beyond a human programmer’s</p><p>insights.” 14 For example, one of its robots was able to</p><p>learn, within only five seconds, how to lift and place a</p><p>chicken (the system is expected to be used in poultry</p><p>factories).15 But the technology could have many</p><p>applications, such as for drones, autonomous vehicles,</p><p>and IoT (Internet of Things).</p><p>• OpenAI: They have created the Dactyl, which is a robot</p><p>hand that has human-like dexterity. This is based on</p><p>sophisticated training of simulations, not real-world</p><p>interactions. OpenAI calls this “domain randomization,”</p><p>which presents the robot many scenarios—even those</p><p>that have a very low probability of happening. With</p><p>Dactyl, the simulations were able to involve about 100</p><p>years of problem solving.16 One of the surprising results</p><p>was that the system learned human hand actions that</p><p>were not preprogrammed—such as sliding of the finger.</p><p>Dactyl also has been trained to deal with imperfect</p><p>information, say when the sensors have delayed readings,</p><p>or when there is a need to handle multiple objects.</p><p>14 www.osaro.com/technology</p><p>15 www.technologyreview.com/s/611424/this-is-how-the-robot-uprising-</p><p>finally-begins/</p><p>16 https://openai.com/blog/learning-dexterity/</p><p>Artificial Intelligence Basics</p><p>http://www.osaro.com/technology</p><p>http://www.technologyreview.com/s/611424/this-is-how-the-robot-uprising-finally-begins/</p><p>http://www.technologyreview.com/s/611424/this-is-how-the-robot-uprising-finally-begins/</p><p>https://openai.com/blog/learning-dexterity/</p><p>132</p><p>• MIT: It can easily take thousands of sample data for a</p><p>robot to understand its environment, such as to detect</p><p>something as simple as a mug. But according to a research</p><p>paper from professors at MIT, there may be a way to</p><p>reduce this. They used a neural network that focused on</p><p>only a few key features.17 The research is still in the early</p><p>stages, but it could prove very impactful for robots.</p><p>• Google: Beginning in 2013, the company went on an M&A</p><p>(mergers and acquisitions) binge for robotics companies.</p><p>But the results were disappointing. Despite this, it has</p><p>not given up on the business. Over the</p><p>past few years,</p><p>Google has focused on pursuing simpler robots that are</p><p>driven by AI and the company has created a new division,</p><p>called Robotics at Google. For example, one of the</p><p>robots can look at a bin of items and identify the one that</p><p>is requested—picking it up with a three-fingered hand—</p><p>about 85% of the time. A typical person, on the other</p><p>hand, was able to do this at about 80%.18</p><p>So does all this point to complete automation? Probably not—at least for the</p><p>foreseeable future. Keep in mind that a major trend is the development of</p><p>cobots. These are robots that work along with people. All in all, it is turning</p><p>into a much more powerful approach, as there can be leveraging of the</p><p>advantages of both machines and humans.</p><p>Note that one of the major leaders in this category is Amazon.com. Back in</p><p>2012, the company shelled out $775 million for Kiva, a top industrial robot</p><p>manufacturer. Since then, Amazon.com has rolled out about 100,000 systems</p><p>across more than 25 fulfillment centers (because of this, the company has</p><p>seen 40% improvement in inventory capacity).19 This is how the company</p><p>describes it:</p><p>Amazon Robotics automates fulfilment center operations using various</p><p>methods of robotic technology including autonomous mobile robots,</p><p>sophisticated control software, language perception, power management,</p><p>computer vision, depth sensing, machine learning, object recognition,</p><p>and semantic understanding of commands.20</p><p>Within the warehouses, robots quickly move across the floor helping to</p><p>locate and lift storage pods. But people are also critical as they are better able</p><p>to identify and pick individual products.</p><p>17 https://arxiv.org/abs/1903.06684</p><p>18 www.nytimes.com/2019/03/26/technology/google-robotics-lab.html</p><p>19 https://techcrunch.com/2019/03/29/built-robotics-massive-construction-</p><p>excavator-drives-itself/</p><p>20 www.amazonrobotics.com/#/vision</p><p>Chapter 7 | Physical Robots</p><p>https://arxiv.org/abs/1903.06684</p><p>http://www.nytimes.com/2019/03/26/technology/google-robotics-lab.html</p><p>https://techcrunch.com/2019/03/29/built-robotics-massive-construction-excavator-drives-itself/</p><p>https://techcrunch.com/2019/03/29/built-robotics-massive-construction-excavator-drives-itself/</p><p>http://www.amazonrobotics.com/#/vision</p><p>133</p><p>Yet the setup is very complicated. For example, warehouse employees wear</p><p>Robotic Tech Vests so as not to be run down by robots!21 This technology</p><p>makes it possible for a robot to identify a person.</p><p>But there are other issues with cobots. For example, there is the real fear that</p><p>employees will ultimately be replaced by the machines. What’s more, it’s</p><p>natural for people to feel like a proverbial cog in the wheel, which could mean</p><p>lower morale. Can people really bond with robots? Probably not, especially</p><p>industrial robots, which really do not have human qualities.</p><p>Robots intheReal World</p><p>OK then, let’s now take a look at some of the other interesting use cases with</p><p>industrial and commercial robots.</p><p>Use Case: Security</p><p>Both Erik Schluntz and Travis Deyle have extensive backgrounds in the robotics</p><p>industry, with stints at companies like Google and SpaceX. In 2016, they</p><p>wanted to start their own venture but first spent considerable time trying to</p><p>find a real-world application for the technology, which involved talking to</p><p>numerous companies. Schluntz and Deyle found one common theme: the need</p><p>for physical security of facilities. How could robots provide protection after</p><p>5 pm—without having to spend large amounts on security guards?</p><p>This resulted in the launch of Cobalt Robotics. The timing was spot-on</p><p>because of the convergence of technologies like computer vision, machine</p><p>learning, and, of course, the strides in robotics.</p><p>While using traditional security technology is effective—say with cameras and</p><p>sensors—they are static and not necessarily good for real-time response. But</p><p>with a robot, it’s possible to be much more proactive because of the mobility</p><p>and the underlying intelligence.</p><p>However, people are still in the loop. Robots can then do what they are good</p><p>at, such as 24/7 data processing and sensing, and people can focus on thinking</p><p>critically and weighing the alternatives.</p><p>Besides its technology, Cobalt has been innovative with its business model,</p><p>which it calls Robotics as a Service (RaaS). By charging a subscription, these</p><p>devices are much more affordable for customers.</p><p>21 www.theverge.com/2019/1/21/18191338/amazon-robot-warehouse-tech-</p><p>vest-utility-belt-safety</p><p>Artificial Intelligence Basics</p><p>http://www.theverge.com/2019/1/21/18191338/amazon-robot-warehouse-tech-vest-utility-belt-safety</p><p>http://www.theverge.com/2019/1/21/18191338/amazon-robot-warehouse-tech-vest-utility-belt-safety</p><p>134</p><p>Use Case: Floor-Scrubbing Robots</p><p>We are likely to see some of the most interesting applications for robots in</p><p>categories that are fairly mundane. Then again, these machines are really good</p><p>at handling repetitive processes.</p><p>Take a look at Brain Corp, which was founded in 2009 by Dr. Eugene Izhikevich</p><p>and Dr. Allen Gruber. They initially developed their technology for Qualcomm</p><p>and DARPA.But Brain has since gone on to leverage machine learning and</p><p>computer vision for self-driving robots. In all, the company has raised $125</p><p>million from investors like Qualcomm and SoftBank.</p><p>Brain’s flagship robot is Auto-C, which efficiently scrubs floors. Because of the</p><p>AI system, called BrainOS (which is connected to the cloud), the machine is</p><p>able to autonomously navigate complex environments. This is done by pressing</p><p>a button, and then Auto-C quickly maps the route.</p><p>In late 2018, Brain struck an agreement with Walmart to roll out 1,500</p><p>Auto-C robots across hundreds of store locations.22 The company has also</p><p>deployed robots at airports and malls.</p><p>But this is not the only robot in the works for Walmart. The company is also</p><p>installing machines that can scan shelves to help with inventory management.</p><p>With about 4,600 stores across the United States, robots will likely have a</p><p>major impact on the retailer.23</p><p>Use Case: Online Pharmacy</p><p>As a second-generation pharmacist, TJ Parker had first-hand experience with</p><p>the frustrations people felt when managing their prescriptions. So he</p><p>wondered: Might the solution be to create a digital pharmacy?</p><p>He was convinced that the answer was yes. But while he had a strong</p><p>background in the industry, he needed a solid tech co-founder, which he found</p><p>in Elliot Cohen, an MIT engineer. They would go on to create PillPack in 2013.</p><p>The focus was to reimagine the customer experience. By using an app or</p><p>going to the PillPack web site, a user could easily sign up—such as to input</p><p>insurance information, enter prescription needs, and schedule deliveries.</p><p>When the user received the package, it would have detailed information</p><p>about dose instructions and even images of each pill. Furthermore, each of</p><p>the pills included labels and were presorted into containers.</p><p>22 www.wsj.com/articles/walmart-is-rolling-out-the-robots-11554782460</p><p>23 https://techcrunch.com/2019/04/10/the-startup-behind-walmarts-</p><p>shelf-scanning-robots/</p><p>Chapter 7 | Physical Robots</p><p>http://www.wsj.com/articles/walmart-is-rolling-out-the-robots-11554782460</p><p>https://techcrunch.com/2019/04/10/the-startup-behind-walmarts-shelf-scanning-robots/</p><p>https://techcrunch.com/2019/04/10/the-startup-behind-walmarts-shelf-scanning-robots/</p><p>135</p><p>To make all this a reality required a sophisticated technology infrastructure,</p><p>called PharmacyOS.It also was based on a network of robots, which were</p><p>located in an 80,000-square-foot warehouse. Through this, the system could</p><p>efficiently sort and package the prescriptions. But the facility also had licensed</p><p>pharmacists to manage the process and make sure everything was in</p><p>compliance.</p><p>In June 2018, Amazon.com shelled out about $1 billion for PillPack. On the</p><p>news, the shares of companies like CVS and Walgreens dropped on the fears</p><p>that the e-commerce giant was preparing to make a big play for the healthcare</p><p>market.</p><p>Use Case: Robot</p><p>Scientists</p><p>Developing prescription drugs is enormously expensive. Based on research</p><p>from the Tufts Center for the Study of Drug Development, the average comes</p><p>to about $2.6 billion per approved compound.24 In addition, it can easily take</p><p>over a decade to get a new drug to market because of the onerous regulations.</p><p>But the use of sophisticated robots and deep learning could help. To see how,</p><p>look at what researchers at the Universities of Aberystwyth and Cambridge</p><p>have done. In 2009, they launched Adam, which was essentially a robot</p><p>scientist that helped with the drug discovery process. Then a few years later,</p><p>they launched Eve, which was the next-generation robot.</p><p>The system can come up with hypotheses and test them as well as run</p><p>experiments. But the process is not just about brute-force calculations (the</p><p>system can screen more than 10,000 compounds per day).25 With deep learning,</p><p>Eve is able to use intelligence to better identify those compounds with the most</p><p>potential. For example, it was able to show that triclosan—a common element</p><p>found in toothpaste to prevent the buildup of plaque—could be effective against</p><p>parasite growth in malaria. This is especially important since the disease has</p><p>been becoming more resistant to existing therapies.</p><p>Humanoid andConsumer Robots</p><p>The popular cartoon, The Jetsons, came out in the early 1960s and had a great</p><p>cast of characters. One was Rosie, which was a robot maid that always had a</p><p>vacuum cleaner in hand.</p><p>24 www.policymed.com/2014/12/a-tough-road-cost-to-develop-one-new-</p><p>drug-is-26-billion-approval-rate-for-drugs-entering-clinical-de.html</p><p>25 www.cam.ac.uk/research/news/artificially-intelligent-robot-scientist-</p><p>eve-could-boost-search-for-new-drugs</p><p>Artificial Intelligence Basics</p><p>http://www.policymed.com/2014/12/a-tough-road-cost-to-develop-one-new-drug-is-26-billion-approval-rate-for-drugs-entering-clinical-de.html</p><p>http://www.policymed.com/2014/12/a-tough-road-cost-to-develop-one-new-drug-is-26-billion-approval-rate-for-drugs-entering-clinical-de.html</p><p>http://www.cam.ac.uk/research/news/artificially-intelligent-robot-scientist-eve-could-boost-search-for-new-drugs</p><p>http://www.cam.ac.uk/research/news/artificially-intelligent-robot-scientist-eve-could-boost-search-for-new-drugs</p><p>136</p><p>Who wouldn’t want something like this? I would. But don’t expect something</p><p>like Rosie coming to a home anytime soon. When it comes to consumer</p><p>robots, we are still in the early days. In other words, we are instead seeing</p><p>robots that have only some human features.</p><p>Here are notable examples:</p><p>• Sophia: Developed by the Hong Kong–based company</p><p>Hanson Robotics, this is perhaps the most famous. In</p><p>fact, in late 2017 Saudi Arabia granted her citizenship!</p><p>Sophia, which has the likeness of Audrey Hepburn, can</p><p>walk and talk. But there are also subtleties with her</p><p>actions, such as sustaining eye contact.</p><p>• Atlas: The developer is Boston Dynamics, which launched</p><p>this in the summer of 2013. No doubt, Atlas has gotten</p><p>much better over the years. It can, for example, perform</p><p>backflips and pick itself up when it falls down.</p><p>• Pepper: This is a humanoid robot, created by SoftBank</p><p>Robotics, that is focused on providing customer service,</p><p>such as at retail locations. The machine can use gestures—</p><p>to help improve communication—and can also speak</p><p>multiple languages.</p><p>As humanoid technologies get more realistic and advanced, there will inevitably</p><p>be changes in society. Social norms about love and friendship will evolve. After</p><p>all, as seen with the pervasiveness with smartphones, we are already seeing</p><p>how technology can change the way we relate to people, say with texting and</p><p>engaging in social media. According to a survey of Millennials from Tappable,</p><p>close to 10% would rather sacrifice their pinky finger than forgo their</p><p>smartphone!26</p><p>As for robots, we may see something similar. It’s about social robots. Such a</p><p>machine—which is life-like with realistic features and AI—could ultimately</p><p>become like, well, a friend or…even a lover.</p><p>Granted, this is likely far in the future. But as of now, there are certainly some</p><p>interesting innovations with social robots. One example is ElliQ, which</p><p>involves a tablet and a small robot head. For the most part, it is for those who</p><p>live alone, such as the elderly. ElliQ can talk but also provide invaluable</p><p>assistance like give reminders for taking medicine. The system can allow for</p><p>video chats with family members as well.27</p><p>26 www.mediapost.com/publications/article/322677/one-in-10-millennials-</p><p>would-rather-lose-a-finger-t.html</p><p>27 www.wsj.com/articles/on-demand-grandkids-and-robot-pals-technology-</p><p>strives-to-cure-senior-loneliness-11550898010?mod=hp_lead_pos9</p><p>Chapter 7 | Physical Robots</p><p>http://www.mediapost.com/publications/article/322677/one-in-10-millennials-would-rather-lose-a-finger-t.html</p><p>http://www.mediapost.com/publications/article/322677/one-in-10-millennials-would-rather-lose-a-finger-t.html</p><p>http://www.wsj.com/articles/on-demand-grandkids-and-robot-pals-technology-strives-to-cure-senior-loneliness-11550898010?mod=hp_lead_pos9</p><p>http://www.wsj.com/articles/on-demand-grandkids-and-robot-pals-technology-strives-to-cure-senior-loneliness-11550898010?mod=hp_lead_pos9</p><p>137</p><p>Yet there are certainly downsides to social robots. Just look at the awful</p><p>situation of Jibo. The company, which had raised $72.7 million in venture</p><p>funding, created the first social robot for the home. But there were many</p><p>problems, such as product delays and the onslaught of knock-offs. Because of</p><p>all this, Jibo filed for bankruptcy in 2018, and by April the following year, the</p><p>servers were shut down.28</p><p>Needless to say, there were many disheartened owners of Jibo, evidenced by</p><p>the many posts on Reddit.</p><p>The Three Laws ofRobotics</p><p>Isaac Asimov, a prolific writer of many diverse subjects like science fiction,</p><p>history, chemistry, and Shakespeare, would also have a major impact on</p><p>robots. In a short story he wrote in 1942 (“Runaround”), he set forth his</p><p>Three Laws of Robotics:</p><p>1. A robot may not injure a human being or, through</p><p>inaction, allow a human being to come to harm.</p><p>2. A robot must obey the orders given to it by human</p><p>beings, except where such orders would conflict with the</p><p>First Law.</p><p>3. A robot must protect its own existence as long as such</p><p>protection does not conflict with the First or Second Law.</p><p>■ Note Asimov would later add another one, the zeroth law, which stated: “A robot may not</p><p>harm humanity, or, by inaction, allow humanity to come to harm.” He considered this law to be</p><p>the most important.</p><p>Asimov would write more short stories that reflected how the laws would</p><p>play out in complex situations, and they would be collected in a book called I,</p><p>Robot. All these took place in the world of the 21st century.</p><p>The Three Laws represented Asimov’s reaction to how science fiction portrayed</p><p>robots as malevolent. But he thought this was unrealistic. Asimov had the</p><p>foresight that there would emerge ethical rules to control the power of robots.</p><p>As of now, Asimov’s vision is starting to become more real—in other words,</p><p>it is a good idea to explore ethical principles. Granted, this may not necessarily</p><p>mean that his approach is the right way. But it is a good start, especially as</p><p>robots get smarter and more personal because of the power of AI.</p><p>28 https://techcrunch.com/2019/03/04/the-lonely-death-of-jibo-the-</p><p>social-robot/</p><p>Artificial Intelligence Basics</p><p>https://techcrunch.com/2019/03/04/the-lonely-death-of-jibo-the-social-robot/</p><p>https://techcrunch.com/2019/03/04/the-lonely-death-of-jibo-the-social-robot/</p><p>138</p><p>Cybersecurity andRobots</p><p>Cybersecurity has not been much of a problem with robots. But unfortunately,</p><p>this will not likely be the case for long. The main reason is that it is becoming</p><p>much more common for robots to be connected to the cloud. The same goes for</p><p>other systems, such as the Internet of Things or IoT, and autonomous cars. For</p><p>example, many of these systems are updated wirelessly, which</p><p>of many predictions of</p><p>AI that would come up short.</p><p>So how has the Turing Test held up over the years? Well, it has proven to be</p><p>difficult to crack. Keep in mind that there are contests, such as the Loebner</p><p>Prize and the Turing Test Competition, to encourage people to create</p><p>intelligent software systems.</p><p>In 2014, there was a case where it did look like the Turing Test was passed. It</p><p>involved a computer that said it was 13 years old.2 Interestingly enough, the</p><p>human judges likely were fooled because some of the answers had errors.</p><p>Then in May 2018 at Google’s I/O conference, CEO Sundar Pichai gave a</p><p>standout demo of Google Assistant.3 Before a live audience, he used the</p><p>device to call a local hairdresser to make an appointment. The person on the</p><p>other end of the line acted as if she was talking to a person!</p><p>Amazing, right? Definitely. Yet it still probably did not pass the Turing Test.</p><p>The reason is that the conversation was focused on one topic—not open</p><p>ended.</p><p>As should be no surprise, there has been ongoing controversy with the Turing</p><p>Test, as some people think it can be manipulated. In 1980, philosopher John</p><p>Searle wrote a famous paper, entitled “Minds, Brains, and Programs,” where</p><p>he set up his own thought experiment, called the “Chinese room argument”</p><p>to highlight the flaws.</p><p>Here’s how it worked: Let’s say John is in a room and does not understand the</p><p>Chinese language. However, he does have manuals that provide easy-to-use</p><p>rules to translate it. Outside the room is Jan, who does understand the</p><p>language and submits characters to John. After some time, she will then get</p><p>an accurate translation from John. As such, it’s reasonable to assume that Jan</p><p>believes that John can speak Chinese.</p><p>2 www.theguardian.com/technology/2014/jun/08/super-computer-</p><p>simulates-13-year-old-boy-passes-turing-test</p><p>3 www.theverge.com/2018/5/8/17332070/google-assistant-makes-</p><p>phone-call-demo-duplex-io-2018</p><p>Artificial Intelligence Basics</p><p>http://www.theguardian.com/technology/2014/jun/08/super-computer-simulates-13-year-old-boy-passes-turing-test</p><p>http://www.theguardian.com/technology/2014/jun/08/super-computer-simulates-13-year-old-boy-passes-turing-test</p><p>http://www.theverge.com/2018/5/8/17332070/google-assistant-makes-phone-call-demo-duplex-io-2018</p><p>http://www.theverge.com/2018/5/8/17332070/google-assistant-makes-phone-call-demo-duplex-io-2018</p><p>4</p><p>Searle’s conclusion:</p><p>The point of the argument is this: if the man in the room does not</p><p>understand Chinese on the basis of implementing the appropriate</p><p>program for understanding Chinese then neither does any other digital</p><p>computer solely on that basis because no computer, qua computer, has</p><p>anything the man does not have.4</p><p>It was a pretty good argument—and has been a hot topic of debate in AI</p><p>circles since.</p><p>Searle also believed there were two forms of AI:</p><p>• Strong AI: This is when a machine truly understands what</p><p>is happening. There may even be emotions and creativity.</p><p>For the most part, it is what we see in science fiction</p><p>movies. This type of AI is also known as Artificial General</p><p>Intelligence (AGI). Note that there are only a handful of</p><p>companies that focus on this category, such as Google’s</p><p>DeepMind.</p><p>• Weak AI: With this, a machine is pattern matching and</p><p>usually focused on narrow tasks. Examples of this include</p><p>Apple’s Siri and Amazon’s Alexa.</p><p>The reality is that AI is in the early phases of weak AI.Reaching the point of</p><p>strong AI could easily take decades. Some researchers think it may never</p><p>happen.</p><p>Given the limitations to the Turing Test, there have emerged alternatives,</p><p>such as the following:</p><p>• Kurzweil-Kapor Test: This is from futurologist Ray Kurzweil</p><p>and tech entrepreneur Mitch Kapor. Their test requires</p><p>that a computer carry on a conversation for two hours</p><p>and that two of three judges believe it is a human talking.</p><p>As for Kapor, he does not believe this will be achieved</p><p>until 2029.</p><p>• Coffee Test: This is from Apple co-founder Steve Wozniak.</p><p>According to the coffee test, a robot must be able to go</p><p>into a stranger’s home, locate the kitchen, and brew a</p><p>cup of coffee.</p><p>4 https://plato.stanford.edu/entries/chinese-room/</p><p>Chapter 1 | AI Foundations</p><p>https://plato.stanford.edu/entries/chinese-room/</p><p>5</p><p>The Brain Is a…Machine?</p><p>In 1943, Warren McCulloch and Walter Pitts met at the University of Chicago,</p><p>and they became fast friends even though their backgrounds were starkly</p><p>different as were their ages (McCulloch was 42 and Pitts was 18). McCulloch</p><p>grew up in a wealthy Eastern Establishment family, having gone to prestigious</p><p>schools. Pitts, on the other hand, grew up in a low-income neighborhood and</p><p>was even homeless as a teenager.</p><p>Despite all this, the partnership would turn into one of the most consequential</p><p>in the development of AI.McCulloch and Pitts developed new theories to</p><p>explain the brain, which often went against the conventional wisdom of</p><p>Freudian psychology. But both of them thought that logic could explain the</p><p>power of the brain and also looked at the insights from Alan Turing. From this,</p><p>they co-wrote a paper in 1943 called “A Logical Calculus of the Ideas Immanent</p><p>in Nervous Activity,” and it appeared in the Bulletin of Mathematical Biophysics.</p><p>The thesis was that the brain’s core functions like neurons and synapses could</p><p>be explained by logic and mathematics, say with logical operators like And,</p><p>Or, and Not. With these, you could construct a complex network that could</p><p>process information, learn, and think.</p><p>Ironically, the paper did not get much traction with neurologists. But it did get</p><p>the attention with those working on computers and AI.</p><p>Cybernetics</p><p>While Norbert Wiener created various theories, his most famous one was</p><p>about cybernetics. It was focused on understanding control and communications</p><p>with animals, people, and machines—showing the importance of feedback loops.</p><p>In 1948, Wiener published Cybernetics: Or Control and Communication in the</p><p>Animal and the Machine. Even though it was a scholarly work—filled with complex</p><p>equations—the book still became a bestseller, hitting the New York Times list.</p><p>It was definitely wide ranging. Some of the topics included Newtonian</p><p>mechanics, meteorology, statistics, astronomy, and thermodynamics. This</p><p>book would anticipate the development of chaos theory, digital</p><p>communications, and even computer memory.</p><p>But the book would also be influential for AI.Like McCulloch and Pitts, Wiener</p><p>compared the human brain to the computer. Furthermore, he speculated that</p><p>a computer would be able to play chess and eventually beat grand masters.</p><p>The main reason is that he believed that a machine could learn as it played</p><p>games. He even thought that computers would be able to replicate themselves.</p><p>But Cybernetics was not utopian either. Wiener was also prescient in</p><p>understanding the downsides of computers, such as the potential for</p><p>dehumanization. He even thought that machines would make people unnecessary.</p><p>Artificial Intelligence Basics</p><p>6</p><p>It was definitely a mixed message. But Wiener’s ideas were powerful and</p><p>spurred the development of AI.</p><p>The Origin Story</p><p>John McCarthy’s interest in computers was spurred in 1948, when he attended</p><p>a seminar, called “Cerebral Mechanisms in Behavior,” which covered the topic</p><p>of how machines would eventually be able to think. Some of the participants</p><p>included the leading pioneers in the field such as John von Neumann, Alan</p><p>Turing, and Claude Shannon.</p><p>McCarthy continued to immerse himself in the emerging computer industry—</p><p>including a stint at Bell Labs—and in 1956, he organized a ten-week research</p><p>project at Dartmouth University. He called it a “study of artificial intelligence.”</p><p>It was the first time the term had been used.</p><p>The attendees included academics like Marvin Minsky, Nathaniel Rochester,</p><p>Allen Newell, O.G. Selfridge, Raymond Solomonoff, and Claude Shannon. All</p><p>of them would go on to become major players in AI.</p><p>The goals for the study were definitely ambitious:</p><p>The study is</p><p>exposes them to</p><p>malware, viruses, and even ransoms. Furthermore, when it comes to electric</p><p>vehicles, there is also a vulnerability to attacks from the charging network.</p><p>In fact, your data could linger within a vehicle! So if it is wrecked or you sell</p><p>it, the information—say video, navigation details, and contacts from paired</p><p>smartphone connections—may become available to other people. A white</p><p>hat hacker, called GreenTheOnly, has been able to extract this data from a</p><p>variety of Tesla models at junkyards, according to CNBC.com.29 But it’s</p><p>important to note that the company does provide options to wipe the data</p><p>and you can opt out of data collection (but this means not having certain</p><p>advantages, like over-the-air (OTA) updates).</p><p>Now if there is a cybersecurity breach with a robot, the implications can</p><p>certainly be devastating. Just imagine if a hacker infiltrated a manufacturing</p><p>line or a supply chain or even a robotic surgery system. Lives could be in</p><p>jeopardy.</p><p>Regardless, there has not been much investment in cybersecurity for robots.</p><p>So far, there are just a handful of companies, like Karamba Security and</p><p>Cybereason, that are focused on this. But as the problems get worse, there</p><p>will inevitably be a ramping of investments from VCs and new initiatives from</p><p>legacy cybersecurity firms.</p><p>Programming Robots forAI</p><p>It is getting easier to create intelligent robots, as systems get cheaper and</p><p>there are new software platforms emerging. A big part of this has been due to</p><p>the Robot Operating System (ROS), which is becoming a standard in the</p><p>industry. The origins go back to 2007 when the platform began as an open</p><p>source project at the Stanford Artificial Intelligence Laboratory.</p><p>Despite its name, ROS is really not a true operating system. Instead, it is</p><p>middleware that helps to manage many of the critical parts of a robot: planning,</p><p>simulations, mapping, localization, perception, and prototypes. ROS is also</p><p>modular, as you can easily pick and choose the functions you need. The result</p><p>is that the system can easily cut down on development time.</p><p>29 www.cnbc.com/2019/03/29/tesla-model-3-keeps-data-like-crash-videos-</p><p>location-phone-contacts.html</p><p>Chapter 7 | Physical Robots</p><p>http://www.cnbc.com/2019/03/29/tesla-model-3-keeps-data-like-crash-videos-location-phone-contacts.html</p><p>http://www.cnbc.com/2019/03/29/tesla-model-3-keeps-data-like-crash-videos-location-phone-contacts.html</p><p>139</p><p>Another advantage: ROS has a global community of users. Consider that</p><p>there are over 3,000 packages for the platform.30</p><p>As a testament to the prowess of ROS, Microsoft announced in late 2018 that</p><p>it would release a version for the Windows operating system. According to</p><p>the blog post from Lou Amadio, the principal software engineer of Windows</p><p>IoT, “As robots have advanced, so have the development tools. We see</p><p>robotics with artificial intelligence as universally accessible technology to</p><p>augment human abilities.”31</p><p>The upshot is that ROS can be used with Visual Studio and there will be</p><p>connections to the Azure cloud, which includes AI Tools.</p><p>OK then, when it comes to developing intelligent robots, there is often a</p><p>different process than with the typical approach with software-based AI.That</p><p>is, there not only needs to be a physical device but also a way to test it. Often</p><p>this is done by using a simulation. Some developers will even start with</p><p>creating cardboard models, which can be a great way to get a sense of the</p><p>physical requirements.</p><p>But of course, there are also useful virtual simulators, such as MuJoCo,</p><p>Gazebo, MORSE, and V-REP.These systems use sophisticated 3D graphics to</p><p>deal with movements and the physics of the real world.</p><p>Then how do you create the AI models for robots? Actually, it is little different</p><p>from the approach with software-based algorithms (as we covered in Chapter</p><p>2). But with a robot, there is the advantage that it will continue to collect data</p><p>from its sensors, which can help evolve the AI.</p><p>The cloud is also becoming a critical factor in the development of intelligent</p><p>robots, as seen with Amazon.com. The company has leveraged its hugely</p><p>popular AWS platform with a new offering, called AWS RoboMaker. By using</p><p>this, you can build, test, and deploy robots without much configuration. AWS</p><p>RoboMaker operates on ROS and also allows the use of services for machine</p><p>learning, analytics, and monitoring. There are even prebuilt virtual 3D worlds</p><p>for retail stores, indoor rooms, and race tracks! Then once you are finished</p><p>with the robot, you can use AWS to develop an over-the-air (OTA) system for</p><p>secure deployment and periodic updates.</p><p>And as should be no surprise, Google is planning on releasing its own robot</p><p>cloud platform (it’s expected to launch in 2019).32</p><p>30 www.ros.org/is-ros-for-me/</p><p>31 https://blogs.windows.com/windowsexperience/2018/09/28/bringing-the-</p><p>power-of-windows-10-to-the-robot-operating-system/</p><p>32 www.therobotreport.com/google-cloud-robotics-platform/</p><p>Artificial Intelligence Basics</p><p>http://www.ros.org/is-ros-for-me/</p><p>https://blogs.windows.com/windowsexperience/2018/09/28/bringing-the-power-of-windows-10-to-the-robot-operating-system/</p><p>https://blogs.windows.com/windowsexperience/2018/09/28/bringing-the-power-of-windows-10-to-the-robot-operating-system/</p><p>http://www.therobotreport.com/google-cloud-robotics-platform/</p><p>140</p><p>The Future ofRobots</p><p>Rodney Brooks is one of the giants of the robotics industry. In 1990, he</p><p>co-founded iRobot to find ways to commercialize the technology. But it was</p><p>not easy. It was not until 2002 that the company launched its Roomba</p><p>vacuuming robot, which was a big hit with consumers. As of this writing,</p><p>iRobot has a market value of $3.2 billion and posted more than $1 billion in</p><p>revenues for 2018.</p><p>But iRobot was not the only startup for Brooks. He would also help to launch</p><p>Rethink Robotics—and his vision was ambitious. Here’s how he put it during</p><p>2010, when his company announced a $20 million funding:</p><p>Our robots will be intuitive to use, intelligent and highly flexible. They’ll</p><p>be easy to buy, train, and deploy and will be unbelievably inexpensive.</p><p>[Rethink Robotics] will change the definition of how and where robots</p><p>can be used, dramatically expanding the robot marketplace.33</p><p>But unfortunately, as with iRobot, there were many challenges. Even though</p><p>Brook’s idea for cobots was pioneering—and would ultimately prove to be a</p><p>lucrative market—he had to struggle with the complications of building an</p><p>effective system. The focus on safety meant that precision and accuracy was</p><p>not up to the standards of industrial customers. Because of this, the demand</p><p>for Rethink’s robots was tepid.</p><p>By October 2018, the company ran out of cash and had to close its doors. In</p><p>all, Rethink had raised close to $150 million from VCs and strategic investors</p><p>like Goldman Sachs, Sigma Partners, GE, and Bezos Expeditions. The company’s</p><p>intellectual property was sold off to a German automation firm, HAHN Group.</p><p>True, this is just one example. But then again, it does show that even the</p><p>smartest tech people can get things wrong. And more importantly, the</p><p>robotics market has unique complexities. When it comes to the evolution of</p><p>this category, progress may be choppy and volatile.</p><p>As Cobalt’s Schluntz has noted:</p><p>While the industry has made progress in the last decade, robotics hasn’t</p><p>yet realized its full potential. Any new technology will create a wave of</p><p>numerous new companies, but only a few will survive and turn into lasting</p><p>businesses. The Dot-Com bust killed the majority of internet companies,</p><p>but Google, Amazon, and Netflix all survived. What robotics companies</p><p>need to do is to be upfront about what their robots can do for customers</p><p>today, overcome Hollywood stereotypes of robots as the bad guys, and</p><p>demonstrate a clear ROI (Return On Investment) to customers.34</p><p>33 www.rethinkrobotics.com/news-item/heartland-robotics-raises-20-million-</p><p>in-series-b-financing/</p><p>34 From the author’s interview with Erik Schluntz, CTO</p><p>of Cobalt Robotics.</p><p>Chapter 7 | Physical Robots</p><p>http://www.rethinkrobotics.com/news-item/heartland-robotics-raises-20-million-in-series-b-financing/</p><p>http://www.rethinkrobotics.com/news-item/heartland-robotics-raises-20-million-in-series-b-financing/</p><p>141</p><p>Conclusion</p><p>Until the past few years, robots were mostly for high-end manufacturing, such</p><p>as for autos. But with the growth in AI and the lower costs for building</p><p>devices, robots are becoming more widespread across a range of industries.</p><p>As seen in this chapter, there are interesting use cases with robots that do</p><p>things like clean floors or provide security for facilities.</p><p>But the use of AI with robotics is still in the nascent stages. Programming</p><p>hardware systems is far from easy, and there is the need of sophisticated</p><p>systems to navigate environments. However, with AI approaches like</p><p>reinforcement learning, there has been accelerated progress.</p><p>But when thinking of using robots, it’s important to understand the limitations.</p><p>There also must be a clear-cut purpose. If not, a deployment can easily lead to</p><p>a costly failure. Even some of the world’s most innovative companies, like</p><p>Google and Tesla, have had challenges in working with robots.</p><p>Key Takeaways</p><p>• A robot can take actions, sense its environment, and have</p><p>some level of intelligence. There are also key functions</p><p>like sensors, actuators (such as motors), and computers.</p><p>• There are two main ways to operate a robot: the</p><p>telerobot (this is controlled by a human) and autonomous</p><p>robot (based on AI systems).</p><p>• Developing robots is incredibly complicated. Even some</p><p>of the world’s best technologists, like Tesla’s Elon Musk,</p><p>have had major troubles with the technology. A key</p><p>reason is the Moravec paradox. Basically, what’s easy for</p><p>humans is often difficult for robots and vice versa.</p><p>• While AI is making an impact on robots, the process has</p><p>been slow. One reason is that there has been more</p><p>emphasis on software-based technologies. But also</p><p>robots are extremely complicated when it comes to</p><p>moving and understanding the environment.</p><p>• Cobots are machines that work alongside humans. The</p><p>idea is that this will allow for the leveraging of the</p><p>advantages of both machines and people.</p><p>Artificial Intelligence Basics</p><p>142</p><p>• The costs of robots are a major reason for lack of</p><p>adoption. But innovative companies, like Cobalt Robotics,</p><p>are using new business models to help out, such as with</p><p>subscriptions.</p><p>• Consumer robots are still in the initial stages, especially</p><p>compared to industrial robots. But there are some</p><p>interesting use cases, such as with machines that can be</p><p>companions for people.</p><p>• During the 1950s, science fiction writer Isaac Asimov</p><p>created the Three Laws of robotics. For the most part,</p><p>they focused on making sure that the machines would</p><p>not harm people or society. Even though there are</p><p>criticisms of Asimov’s approach, they are still widely</p><p>accepted.</p><p>• Security has not generally been a problem with robots.</p><p>But this will likely change—and fast. After all, more</p><p>robots are connected to the cloud, which allows for the</p><p>intrusion of viruses and malware.</p><p>• The Robot Operating System (ROS) has become a</p><p>standard for the robotics industry. This middleware helps</p><p>with planning, simulations, mapping, localization, perception,</p><p>and prototypes.</p><p>• Developing intelligent robots has many challenges</p><p>because of the need to create physical systems. Although,</p><p>there are tools to help out, such as by allowing for</p><p>sophisticated simulations.</p><p>Chapter 7 | Physical Robots</p><p>© Tom Taulli 2019</p><p>T. Taulli, Artif icial Intelligence Basics,</p><p>https://doi.org/10.1007/978-1-4842-5028-0_8</p><p>C H A P T E R</p><p>8</p><p>Implementation</p><p>ofAI</p><p>Moving the Needle for Your Company</p><p>In March 2019, a shooter live-streamed on Facebook his brutal killing of 50</p><p>people in two mosques in New Zealand. It was viewed about 4,000 times and</p><p>was not shut off until 29 minutes after the attack.1 The video was then</p><p>uploaded to other platforms and was viewed millions of times.</p><p>Yes, this was a stark example of how AI can fail in a horrible way.</p><p>In a blog post, Facebook’s VP of Product Management, Guy Rosen, noted:</p><p>AI systems are based on ‘training data,’ which means</p><p>you need many thousands of examples of content in</p><p>order to train a system that can detect certain types</p><p>of text, imagery or video. This approach has worked</p><p>very well for areas such as nudity, terrorist propaganda</p><p>and also graphic violence where there is a large</p><p>number of examples we can use to train our systems.</p><p>However, this particular video did not trigger our</p><p>automatic detection systems. To achieve that we will</p><p>need to provide our systems with large volumes of</p><p>1 www.cnbc.com/2019/03/21/why-facebooks-ai-didnt-detect-the-new-zealand-</p><p>mosque-shooting-video.html</p><p>http://www.cnbc.com/2019/03/21/why-facebooks-ai-didnt-detect-the-new-zealand-mosque-shooting-video.html</p><p>http://www.cnbc.com/2019/03/21/why-facebooks-ai-didnt-detect-the-new-zealand-mosque-shooting-video.html</p><p>144</p><p>data of this specific kind of content, something which</p><p>is difficult as these events are thankfully rare. Another</p><p>challenge is to automatically discern this content from</p><p>visually similar, innocuous content—for example if</p><p>thousands of videos from live-streamed video games</p><p>are flagged by our systems, our reviewers could miss</p><p>the important real-world videos where we could alert</p><p>first responders to get help on the ground.2</p><p>It also did not help that there were various bad actors that re-uploaded edited</p><p>versions of the video in order to foil Facebook’s AI system.</p><p>Of course, this was a big lesson in the shortfalls of technology, and the</p><p>company says it is committed to keep improving its systems. But the Facebook</p><p>case study also highlights that even the most technologically sophisticated</p><p>companies have major challenges. This is why when it comes to implementing</p><p>AI, there needs to be solid planning as well as an understanding that there will</p><p>inevitably be problems. But it can be tough as senior managers at companies</p><p>are under pressure to get results from this technology.</p><p>In this chapter, we’ll take a look at some of the best practices for AI</p><p>implementations.</p><p>Approaches toImplementing AI</p><p>Using AI in a company generally involves two approaches: using vendor</p><p>software or creating in-house models. The first one is the most prevalent—</p><p>and may be enough for a large number of companies. The irony is that you</p><p>may already be using software, say from Salesforce.com, Microsoft, Google,</p><p>Workday, Adobe, or SAP, that already has powerful AI capabilities. In other</p><p>words, a good approach is to make sure you are taking advantage of these to</p><p>the fullest.</p><p>To see what’s available, take a look at Salesforce.com’s Einstein, which was</p><p>launched in September 2016. This AI system is seamlessly embedded into the</p><p>main CRM (Customer Relationship Management) platform, allowing for more</p><p>predictive and personalized actions for sales, service, marketing, and</p><p>commerce. Salesforce.com calls Einstein a “personal data scientist” as it is</p><p>fairly easy to use, such as with drag and drop to create the workflows. Some</p><p>of the capabilities include the following:</p><p>• Predictive Scoring: This shows the likelihood that a lead</p><p>will convert into an opportunity.</p><p>2 https://newsroom.fb.com/news/2019/03/technical-update-on-new-zealand/</p><p>Chapter 8 | Implementation ofAI</p><p>https://newsroom.fb.com/news/2019/03/technical-update-on-new-zealand/</p><p>145</p><p>• Sentiment Analysis: This provides a way to get a sense of</p><p>how people view your brand and products by analyzing</p><p>social media.</p><p>• Smart Recommendations: Einstein crunches data to show</p><p>what products are the most ideal for leads.</p><p>However, while these prebuilt features make it easier to use AI, there are still</p><p>potential issues. “We have been building AI functions into our applications</p><p>during the past few years and this has been a great learning experience,” said</p><p>Ricky Thakrar, who is Zoho’s customer experience evangelist. “But to make</p><p>the technology</p><p>work, the users must use the software right. If the sales people</p><p>are not inputting information correctly, then the results will likely be off. We</p><p>also found that there should be at least three months of usage for the models</p><p>to get trained. And besides, even if your employees are doing everything right,</p><p>this does not mean that the AI predictions will be perfect. Always take things</p><p>with a grain of salt.”3</p><p>Now as for building your own AI models, this is a significant commitment for</p><p>a company. And this is what we’ll be covering in this chapter.</p><p>But regardless of what approach you may take, the implementation and use of AI</p><p>should first begin with education and training. It does not matter whether the</p><p>employees are non-technical people or software engineers. For AI to be</p><p>successful in an organization, everyone must have a core understanding of the</p><p>technology. Yes, this book will be helpful but there are many online resources to</p><p>help out as well, such as from training platforms like Lynda, Udacity, and Udemy.</p><p>They provide hundreds of high-quality courses on many topics about AI.</p><p>To give a sense of what a corporate training program looks like, consider Adobe.</p><p>Even though the company has incredibly talented engineers, there are still a</p><p>large number who do not have a background in AI.Some of them may not have</p><p>specialized in this in school or their work. Yet Adobe wanted to ensure that all</p><p>the engineers had a solid grasp of the core principles of AI.To this end, the</p><p>company has a six-month certification program, which trained 5,000 engineers</p><p>in 2018. The goal is to unleash the data scientist in each engineer.</p><p>The program includes both online courses and in-person sessions, which not</p><p>only cover technical topics but also areas like strategy and even ethics. Adobe</p><p>also provides help from senior computer scientists to assist students to</p><p>master the topics.</p><p>Next, early on in the implementation process, it’s essential to think about the</p><p>potential risks. Perhaps one of the most threatening is bias since it can easily</p><p>seep into an AI model.</p><p>3 This is based on the author’s interview, in April 2019, with Ricky Thakrar, who is Zoho’s</p><p>customer experience evangelist.</p><p>Artificial Intelligence Basics</p><p>146</p><p>An example of this is Amazon.com, which shut down its AI-powered recruiting</p><p>software in 2017. The main issue was that it was biased for hiring males.</p><p>Interestingly enough, this was a classic case of a training problem for the</p><p>model. Consider that a majority of the resume submissions were from men—</p><p>so the data was skewed. Amazon.com even tried to tweak the model, but still</p><p>the results were far from being gender neutral.4</p><p>In this case, the issue was not just about making decisions that were based on</p><p>faulty premises. Amazon.com was also probably exposing itself to potential</p><p>legal liability, such as with discrimination claims.</p><p>Given the tricky issues with AI, more companies are putting together ethics</p><p>boards. But even this can be fraught with problems. Hey, what may be ethical</p><p>for one person may not be a big deal for someone else, right? Definitely.</p><p>For example, Google closed down its own ethics board in about a week of its</p><p>launch. It appears the main reason was the backlash that came from including</p><p>a member from the Heritage Foundation, which is a conservative think tank.5</p><p>The Steps forAI Implementation</p><p>If you plan to implement your own AI models, what are the main steps to</p><p>consider? What are the best practices? Well, first of all, it’s critically important</p><p>that your data is fairly clean and structured in a way to allow for modelling</p><p>(see Chapter 2).</p><p>Here are some other steps to look at:</p><p>• Identify a problem to solve.</p><p>• Put together a strong team.</p><p>• Select the right tools and platforms.</p><p>• Create the AI model (we went through this process in</p><p>Chapter 3).</p><p>• Deploy and monitor the AI model.</p><p>Let’s take a look at each.</p><p>4 www.reuters.com/article/us-amazon-com-jobs-automation-insight/</p><p>amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-women-</p><p>idUSKCN1MK08G</p><p>5 www.theverge.com/2019/4/4/18296113/google-ai-ethics-board-ends-</p><p>controversy-kay-coles-james-heritage-foundation</p><p>Chapter 8 | Implementation ofAI</p><p>http://www.reuters.com/article/us-amazon-com-jobs-automation-insight/amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-women-idUSKCN1MK08G</p><p>http://www.reuters.com/article/us-amazon-com-jobs-automation-insight/amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-women-idUSKCN1MK08G</p><p>http://www.reuters.com/article/us-amazon-com-jobs-automation-insight/amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-women-idUSKCN1MK08G</p><p>http://www.theverge.com/2019/4/4/18296113/google-ai-ethics-board-ends-controversy-kay-coles-james-heritage-foundation</p><p>http://www.theverge.com/2019/4/4/18296113/google-ai-ethics-board-ends-controversy-kay-coles-james-heritage-foundation</p><p>147</p><p>Identify aProblem toSolve</p><p>Founded in 1976, HCL Technologies is one of the largest IT consulting firms,</p><p>with 132,000 employees across 44 countries, and has half the Fortune 500 as</p><p>customers. The company also has implemented a large number of AI systems.</p><p>Here’s what Kalyan Kumar, who is the corporate vice president and global</p><p>CTO of HCL Technologies, has to say:</p><p>Business leaders need to understand and realize that</p><p>the adoption of Artificial Intelligence is a journey and</p><p>not a sprint. It is critical that the people driving AI</p><p>adoption within an enterprise remain realistic about</p><p>the timeframe and what AI is capable of doing. The</p><p>relationship between humans and AI is mutually</p><p>empowering, and any AI implementation may take</p><p>some time before it starts to make a positive and</p><p>significant impact.6</p><p>It’s great advice. This is why—especially for companies that are starting in the</p><p>AI journey—it’s essential to take an experimental approach. Think of it as</p><p>putting together a pilot program—that is, you are in the “crawl and walk phase.”</p><p>But when it comes to the AI implementation process, it’s common to get too</p><p>focused on the different technologies, which are certainly fascinating and</p><p>powerful. Yet success is far more than just technology; in other words, there</p><p>must first be a clear business case. So here are some areas to think about</p><p>when starting out:</p><p>• No doubt, decisions in companies are often ad hoc and,</p><p>well, a matter of guessing! But with AI, you have an</p><p>opportunity to use data-driven decision-making, which</p><p>should have more accuracy. Then where in your</p><p>organization can this have the biggest benefit?</p><p>• As seen with Robotic Process Automation (RPA), which</p><p>we covered in Chapter 5, AI can be extremely effective</p><p>when handling repetitive and mundane tasks.</p><p>• Chatbots can be another way to start out with AI.They</p><p>are relatively easy to set up and can serve specific use</p><p>cases, such as customer service. You can learn more</p><p>about this in Chapter 6.</p><p>6 This is based on the author’s interview, in March 2019, with Kalyan Kumar, who is the</p><p>corporate vice president and global CTO of HCL Technologies.</p><p>Artificial Intelligence Basics</p><p>148</p><p>Andrew Ng, who is the CEO of Landing AI and the former head of Google</p><p>Brain, has come up with various approaches to think about when identifying</p><p>what to focus on with your initial AI project:7</p><p>• Quick Win: A project should take anywhere from 6 to 12</p><p>months and must have a high probability of success,</p><p>which should help provide momentum for more</p><p>initiatives. Andrew suggests having a couple projects as it</p><p>increases the odds of getting a win.</p><p>• Meaningful: A project does not have to be transformative.</p><p>But it should have results that help improve the company</p><p>in a notable way, creating more buy-in for additional AI</p><p>investments. The value usually comes from lower costs,</p><p>higher revenues, finding new extensions of the business,</p><p>or mitigating risks.</p><p>• Industry-Specific Focus: This is critical since a successful</p><p>project will be another factor in boosting buy-in. Thus, if</p><p>you have a company that sells a subscription service,</p><p>then</p><p>an AI system to lessen churn would be a good place to</p><p>start.</p><p>• Data: Do not limit your options based on the amount of</p><p>data you have. Andrew notes that a successful AI project</p><p>may have as little as 100 data points. But the data must</p><p>still be high quality and fairly clean, which are key topics</p><p>covered in Chapter 2.</p><p>When looking at this phase, it is also worth evaluating the “tango” between</p><p>employees and machines. Keep in mind that this is often missed—and it can</p><p>have adverse consequences on an AI project. As we’ve seen in this book, AI is</p><p>great at processing huge amounts of data with little error at great speed. The</p><p>technology is also excellent with predictions and detecting anomalies. But</p><p>there are tasks that humans do much better, such as being creative, engaging</p><p>in abstraction, and understanding concepts.</p><p>Note the following example of this from Erik Schluntz, who is the co-founder</p><p>and CTO at Cobalt Robotics:</p><p>Our security robots are excellent at detecting unusual</p><p>events in workplace and campus settings, like spotting</p><p>a person in a dark office with AI-powered thermal-</p><p>imaging. But one of our human operators then steps</p><p>7 https://hbr.org/2019/02/how-to-choose-your-first-ai-project</p><p>Chapter 8 | Implementation ofAI</p><p>https://hbr.org/2019/02/how-to-choose-your-first-ai-project</p><p>149</p><p>in and makes the call of how to respond. Even with all</p><p>of AI’s potential, it’s still not the best mission-critical</p><p>option when pitted against constantly changing</p><p>environmental variables and human unpredictability.</p><p>Consider the gravity of AI making a mistake in different</p><p>situations—failing to detect a malicious intruder is</p><p>much worse than accidentally sounding a false alarm</p><p>to one of our operators.8</p><p>Next, make sure you are clear-cut about the KPIs and measure them diligently.</p><p>For example, if you are developing a custom chatbot for customer service,</p><p>you might want to measure against metrics like the resolution rate and</p><p>customer satisfaction.</p><p>And finally, you will need to do an IT assessment. If you have mostly legacy</p><p>systems, then it could be more difficult and expensive to implement AI, even</p><p>if vendors have APIs and integrations. This means you will need to temper</p><p>your expectations.</p><p>Despite all this, the investments can truly move the needle, even for old-line</p><p>companies. To see an example of this, consider Symrise, whose roots go back</p><p>more than 200 years in Germany. As of this writing, the company is a global</p><p>producer of flavors and fragrances, with over 30,000 products.</p><p>A few years ago, Symrise embarked on a major initiative, with the help of IBM,</p><p>to leverage AI to create new perfumes. The company not only had to retool</p><p>its existing IT infrastructure but also had to spend considerable time fine-</p><p>tuning the models. But a big help was that it already had an extensive dataset,</p><p>which allowed for more precision. Note that even a slight deviation in the</p><p>mixture of a compound can make a perfume fail.</p><p>According to Symrise’s president of Scent and Care, Achim Daub:</p><p>Now our perfumers can work with an AI apprentice</p><p>by their side, that can analyze thousands of formulas</p><p>and historical data to identify patterns and predict</p><p>novel combinations, helping to make them more</p><p>productive, and accelerate the design process by</p><p>guiding them toward formulas that have never been</p><p>seen before.9</p><p>8 This is based on the author’s interview, in April 2019, with Erik Schluntz, who is the</p><p>co-founder and CTO at Cobalt Robotics.</p><p>9 www.symrise.com/newsroom/article/breaking-new-fragrance-ground-</p><p>with-artificial-intelligence-ai-ibm-research-and-symrise-are-workin/</p><p>Artificial Intelligence Basics</p><p>http://www.symrise.com/newsroom/article/breaking-new-fragrance-ground-with-artificial-intelligence-ai-ibm-research-and-symrise-are-workin/</p><p>http://www.symrise.com/newsroom/article/breaking-new-fragrance-ground-with-artificial-intelligence-ai-ibm-research-and-symrise-are-workin/</p><p>150</p><p>Forming theTeam</p><p>How large should the initial team be for an AI project? Perhaps a good guide</p><p>is to use Jeff Bezos’ “two pizza rule.”10 In other words, is this enough to feed</p><p>the people who are participating?</p><p>Oh, and there should be no rush to build the team. Everyone must be highly</p><p>focused on success and understand the importance of the project. If there is</p><p>little to show from the AI project, the prospects for future initiatives could be</p><p>in jeopardy.</p><p>The team will need a leader who generally has a business or operational</p><p>background but also has some technical skills. Such a person should be able to</p><p>identify the business case for the AI project but also communicate the vision</p><p>to multiple stakeholders in the company, such as the IT department and</p><p>senior management.</p><p>In terms of the technical people, there will probably not be a need for a PhD</p><p>in AI.While such people are brilliant, they are often focused primarily on</p><p>innovations in the field, such as by refining models or creating new ones.</p><p>These skillsets are usually not essential for an AI pilot.</p><p>Rather, look for those people who have a background in software engineering</p><p>or data science. However, as noted earlier in the chapter, these people may</p><p>not have a strong background in AI.Because of this, there may be a need to</p><p>have them spend a few months of training on learning the core principles of</p><p>machine learning and deep learning. There should also be a focus on</p><p>understanding how to use AI platforms, such as TensorFlow.</p><p>Given the challenges, it may be a good idea to seek the help of consultants,</p><p>who can help identify the AI opportunities but also provide advice on data</p><p>preparation and the development of the models.</p><p>Since an AI pilot will be experimental, the team should have people who are</p><p>willing to take risks and are open minded. If not, progress could be extremely</p><p>difficult.</p><p>The Right Tools andPlatforms</p><p>There are many tools for helping create AI models, and most of them are</p><p>open source. Even though it’s good to test them out, it is still advisable to first</p><p>conduct your IT assessment. By doing this, you should be in a better position</p><p>to evaluate the AI Tools.</p><p>10 www.geekwire.com/2018/amazon-tops-600k-worldwide-employees-</p><p>1st-time-13-jump-year-ago/</p><p>Chapter 8 | Implementation ofAI</p><p>http://www.geekwire.com/2018/amazon-tops-600k-worldwide-employees-1st-time-13-jump-year-ago/</p><p>http://www.geekwire.com/2018/amazon-tops-600k-worldwide-employees-1st-time-13-jump-year-ago/</p><p>151</p><p>Something else: You may realize that your company is already using multiple</p><p>AI Tools and platforms! This may cause issues with integration and the</p><p>management of the process with AI projects. In light of this, a company should</p><p>develop a strategy for the tools. Think of it as your AI Tools stack.</p><p>OK then, let’s take a look at some of the more common languages, platforms,</p><p>and tools for AI.</p><p>Python Language</p><p>Guido van Rossum, who got his master’s degree in mathematics and computer</p><p>science from the University of Amsterdam in 1982, would go on to work at</p><p>various research institutes in Europe like the Corporation for National</p><p>Research Initiatives (CNRI). But it was in the late 1980s that he would create</p><p>his own computer language, called Python. The name actually came from the</p><p>popular British comedy series Monty Python.</p><p>So the language was kind of offbeat—but this made it so powerful. Python</p><p>would soon become the standard for AI development.</p><p>Part of this was due to the simplicity. With just a few scripts of code, you can</p><p>create sophisticated models, say with functions like filter, map, and reduce.</p><p>But of course, the language allows for much sophisticated coding as well.</p><p>Van Rossum developed Python with a clear philosophy:11</p><p>• Beautiful is better than ugly.</p><p>• Explicit is better than implicit.</p><p>• Simple is better than complex.</p><p>• Complex is better than complicated.</p><p>• Flat is better than nested.</p><p>• Sparse is better than dense.</p><p>These are just some of the principles.</p><p>What’s more, Python had the advantage of growing in the academic community,</p><p>which had access to the Internet</p><p>that helped accelerate the distribution. But</p><p>it also made it possible for the emergence of a global ecosystem with thousands</p><p>of different AI packages and libraries. Here are just some:</p><p>11 www.python.org/dev/peps/pep-0020/</p><p>Artificial Intelligence Basics</p><p>http://www.python.org/dev/peps/pep-0020/</p><p>152</p><p>• NumPy: This allows for scientific computing applications.</p><p>At the heart of this is the ability to create a sophisticated</p><p>array of objects at high performance. This is critical for</p><p>high-end data processing in AI models.</p><p>• Matplotlib: With this, you can plot datasets. Often</p><p>Matplotlib is used in conjunction with NumPy/Pandas</p><p>(Pandas refers to “Python Data Analysis Library”). This</p><p>library makes it relatively easy to create data structures</p><p>for developing AI models.</p><p>• SimpleAI: This is an implementation of the AI algorithms</p><p>from the book Artif icial Intelligence: A Modern Approach, by</p><p>Stuart Russel and Peter Norvig. The library not only has</p><p>rich functionality but also provides helpful resources to</p><p>navigate the process.</p><p>• PyBrain: This is a modular machine learning library that</p><p>makes it possible to create sophisticated models—neural</p><p>networks and reinforcement learning systems—without</p><p>much coding.</p><p>• Scikit-Learn: Launched in 2007, this library has a deep</p><p>source of capabilities, allowing for regression, clustering,</p><p>and classification of data.</p><p>Another benefit for Python is that there are many resources for learning.</p><p>A quick search on YouTube will show thousands of free courses.</p><p>Now there are other solid languages you can use for AI like C++, C#, and</p><p>Java. While they are generally more powerful than Python, they are also</p><p>complex. Besides, when it comes to building models, there is often little need</p><p>to create full-fledged applications. And finally, there are Python libraries built</p><p>for high-speed AI machines—with GPUs—like CUDA Python.</p><p>AI Frameworks</p><p>There are a myriad of AI frameworks, which provide end-to-end systems to</p><p>build models, train them, and deploy them. By far the most popular is</p><p>TensorFlow, which is backed by Google. The company started development</p><p>of this framework in 2011, through its Google Brain division. The goal was to</p><p>find a way to create neural networks faster so as to embed the technology</p><p>across many Google applications</p><p>By 2015, Google decided to open source TensorFlow, primarily because the</p><p>company wanted to accelerate the progress of AI.And no doubt, this is what</p><p>happened. By open sourcing TensorFlow, Google made its technology an</p><p>industry standard for development. The software has been downloaded over</p><p>Chapter 8 | Implementation ofAI</p><p>153</p><p>41 million times, and there are more than 1,800 contributors. In fact,</p><p>TensorFlow Lite (which is for embedded systems) is running on more than</p><p>2 billion mobile devices.12</p><p>The ubiquity of the platform has resulted in a large ecosystem. This means</p><p>there are many add-ons like TensorFlow Federated (for decentralized data),</p><p>TensorFlow Privacy, TensorFlow Probability, TensorFlow Agents (for</p><p>reinforcement learning), and Mesh TensorFlow (for massive datasets).</p><p>To use TensorFlow, you have the option of a variety of languages to create</p><p>your models, such as Swift, JavaScript, and R.Although, for the most part, the</p><p>most common one is Python.</p><p>In terms of the basic structure, TensorFlow takes in input data as a</p><p>multidimensional array, which is also known as a tensor. There is a flow to it,</p><p>represented by a chart, as the data courses through the system.</p><p>When you enter commands into TensorFlow, they are processed using a</p><p>sophisticated C++ kernel. This allows for much higher performance, which</p><p>can be essential as some models can be massive.</p><p>TensorFlow can be used for just about anything when it comes to AI.Here are</p><p>some of the models that it has powered:</p><p>• Researchers from NERSC (National Energy Research</p><p>Scientific Computing Center) at the Lawrence Berkeley</p><p>National Laboratory created a deep learning system to</p><p>better predict extreme weather. It was the first such</p><p>model that broke the expo (1 billion billion calculations)</p><p>computing barrier. Because of this, the researchers won</p><p>the Gordon Bell Prize.13</p><p>• Airbnb used TensorFlow to build a model that categorized</p><p>millions of listing photos, which increased the guest</p><p>experience and led to higher conversions.14</p><p>• Google used TensorFlow to analyze data from NASA’s</p><p>Kepler space telescope. The result? By training a neural</p><p>network, the model discovered two exoplanets. Google</p><p>also made available the code to the public.15</p><p>Google has been working on TensorFlow 2.0, and a key focus is to make the</p><p>API process simpler. There is also something called Datasets, which helps to</p><p>streamline the preparation of data for AI models.</p><p>12 https://medium.com/tensorflow/recap-of-the-2019-tensorflow-dev-</p><p>summit-1b5ede42da8d</p><p>13 www.youtube.com/watch?v=p45kQklIsd4&feature=youtu.be</p><p>14 www.youtube.com/watch?v=tPb2u9kwh2w&feature=youtu.be</p><p>15 https://ai.googleblog.com/2018/03/open-sourcing-hunt-for-exoplanets.</p><p>html</p><p>Artificial Intelligence Basics</p><p>https://medium.com/tensorflow/recap-of-the-2019-tensorflow-dev-summit-1b5ede42da8d</p><p>https://medium.com/tensorflow/recap-of-the-2019-tensorflow-dev-summit-1b5ede42da8d</p><p>http://www.youtube.com/watch?v=p45kQklIsd4&feature=youtu.be</p><p>http://www.youtube.com/watch?v=tPb2u9kwh2w&feature=youtu.be</p><p>https://ai.googleblog.com/2018/03/open-sourcing-hunt-for-exoplanets.html</p><p>https://ai.googleblog.com/2018/03/open-sourcing-hunt-for-exoplanets.html</p><p>154</p><p>Then what are some of the other AI frameworks? Let’s take a look:</p><p>• PyTorch: Facebook is the developer of this platform, which</p><p>was released in 2016. Like TensorFlow, the main language</p><p>to program the system is Python. While PyTorch is still in</p><p>the early phases, it is already considered the runner-up to</p><p>TensorFlow in terms of usage. So what is different with</p><p>this platform? PyTorch has a more intuitive interface. The</p><p>platform also allows for dynamic computation of graphs.</p><p>This means you can easily make changes to your models</p><p>in runtime, which helps speed up development. PyTorch</p><p>also makes it possible for having different types of back-</p><p>end CPUs and GPUs.</p><p>• Keras: Even though TensorFlow and PyTorch are for</p><p>experienced AI experts, Keras is for beginners. With a</p><p>small amount of code—in Python—you can create neural</p><p>networks. In the documentation, it notes: “Keras is an</p><p>API designed for human beings, not machines. It puts user</p><p>experience front and center. Keras follows best practices</p><p>for reducing cognitive load: it offers consistent and simple</p><p>APIs, it minimizes the number of user actions required for</p><p>common use cases, and it provides clear and actionable</p><p>feedback upon user error.”16 There is a “Getting Started”</p><p>guide that takes only 30 seconds! Yet the simplicity does</p><p>not mean that it is not powerful. The fact is that you can</p><p>create sophisticated models with Keras. For example,</p><p>TensorFlow has integrated Keras on its own platform.</p><p>Even for those who are pros at AI, the system can be quite</p><p>useful for doing initial experimentations with models.</p><p>With AI development, there is another common tool: Jupyter Notebook. It’s</p><p>not a platform or development tool. Instead, Jupyter Notebook is a web app</p><p>that makes it easy to code in Python and R to create visualizations and import</p><p>AI systems. You can also easily share your work with other people, similar to</p><p>what GitHub does.</p><p>During the past few years, there has also emerged a new category of AI Tools</p><p>called automated machine learning or autoML. These systems help to deal</p><p>with processes like data prep and feature selection. For the most part, the</p><p>goal is to provide help for those organizations that do not have experienced</p><p>data scientists and AI engineers. This is all about the fast-growing trend of the</p><p>“citizen data scientist”—that is, a person who does not have a strong technical</p><p>background who can still create useful models.</p><p>16 https://keras.io/</p><p>Chapter 8 | Implementation ofAI</p><p>https://keras.io/</p><p>155</p><p>Some of the players in the</p><p>autoML space include H2O.ai, DataRobot, and</p><p>SaaS. The systems are intuitive and use drag-and-drop ease with the</p><p>development of models. As should be no surprise, mega tech operators like</p><p>Facebook and Google have created autoML systems for their own teams. In</p><p>the case of Facebook, it has Asimo, which helps manage the training and</p><p>testing of 300,000 models every month.17</p><p>For a use case of autoML, take a look at Lenovo Brazil. The company was</p><p>having difficulty creating machine learning models to help predict and manage</p><p>the supply chain. It had two people who coded 1,500 lines of R code each</p><p>week—but this was not enough. The fact is that it would not be cost-effective</p><p>to hire more data scientists.</p><p>Hence the company implemented DataRobot. By automating various</p><p>processes, Lenovo Brazil was able to create models with more variables,</p><p>which led to better results. Within a few months, the number of users of</p><p>DataRobot went from two to ten.</p><p>Table 8-1 shows some other results.18</p><p>Pretty good, right? Absolutely. But there are still come caveats. With Lenovo</p><p>Brazil, the company had the benefit of skilled data scientists, who understood</p><p>the nuances of creating models.</p><p>However, if you use an autoML tool without such expertise, you could easily</p><p>run into serious trouble. There’s a good chance that you may create models</p><p>that have faulty assumptions or data. If anything, the results may ultimately</p><p>prove far worse than not using AI! Because of this, DataRobot actually requires</p><p>that a new customer have a dedicated field engineer and data scientist work</p><p>with the company for the first year.19</p><p>Now there are also low-code platforms that have proven to be useful in</p><p>accelerating the development of AI projects. One of the leaders in the space</p><p>is Appian, which has the bold guarantee of “Idea to app in eight weeks.”</p><p>17 www.aimlmarketplace.com/technology/machine-learning/the-rise-of-</p><p>automated-machine-learning</p><p>18 https://3gp10c1vpy442j63me73gy3s-wpengine.netdna-ssl.com/wp-content/</p><p>uploads/2018/08/Lenovo-Case-Study.pdf</p><p>19 www.wsj.com/articles/yes-you-too-can-be-an-ai-expert-11554168513</p><p>Table 8-1. The results of implementing an autoML system</p><p>Tasks Before After</p><p>Model creation 4 weeks 3 days</p><p>Production models 2 days 5 minutes</p><p>Accuracy of predictions</p><p>technology. But there should also be a focus on data</p><p>quality. If not, the results will likely be off the mark.</p><p>The second path is to do an AI project, which is based on your company’s</p><p>own data. To be successful, there must be a strong team that has a blend of</p><p>technical, business, and domain expertise. There will also likely be a need for</p><p>some AI training. This is the case even for those with backgrounds in data</p><p>science and engineering.</p><p>From here, there should be no rush in the steps of the project: assessing the</p><p>IT environment, setting up a clear business objective, cleaning the data,</p><p>selecting the right tools and platforms, creating the AI model, and deploying</p><p>the system. With early projects, there will inevitably be challenges so it’s</p><p>critical to be flexible. But the effort should be well worth it.</p><p>Key Takeaways</p><p>• Even the best companies have difficulties with</p><p>implementing AI. Because of this, there must be great</p><p>care, diligence, and planning. It’s also important to realize</p><p>that failure is common.</p><p>• There are two main ways to use AI in a company: through</p><p>a vendor’s software application or an in-house model.</p><p>The latter is much more difficult and requires a major</p><p>commitment from the organization.</p><p>• When using off-the-shelf AI applications, there is still</p><p>much work to be done. For example, if the employees</p><p>are not correctly inputting the data, then the results will</p><p>likely be off.</p><p>• Education is critical with an AI implementation, even for</p><p>experienced engineers. There are excellent online</p><p>training resources to help out with this.</p><p>• Be mindful of the risks of AI implementations, such as</p><p>bias, security, and privacy.</p><p>• Some of the key parts of the AI implementation process</p><p>include the following: identify a problem to solve; put</p><p>together a strong team; select the right tools and</p><p>platforms; create the AI model; and deploy and monitor</p><p>the AI model.</p><p>Chapter 8 | Implementation ofAI</p><p>159</p><p>• When developing a model, look at how the technology</p><p>relates to people. The fact is that people can be much</p><p>better at certain tasks.</p><p>• Forming the team is not easy, so do not rush the process.</p><p>Have a leader who has a good business or operational</p><p>background, with a mix of technical skills.</p><p>• It’s good to experiment with the various AI Tools.</p><p>However, before doing this, make sure you do an IT</p><p>assessment.</p><p>• Some of the popular AI Tools include TensorFlow,</p><p>PyTorch, Python, Keras, and the Jupyter Notebook.</p><p>• Automated machine learning or autoML tools help to</p><p>deal with processes like data prep and feature selection</p><p>for AI models. The focus is on those who do not have</p><p>technical skills.</p><p>• Deployment of the AI model is more than just scaling. It’s</p><p>also critical to have the system easy to use, so as to allow</p><p>for much more adoption.</p><p>Artificial Intelligence Basics</p><p>© Tom Taulli 2019</p><p>T. Taulli, Artif icial Intelligence Basics,</p><p>https://doi.org/10.1007/978-1-4842-5028-0_9</p><p>C H A P T E R</p><p>9</p><p>The Future ofAI</p><p>The Pros and Cons</p><p>At the Web Summit conference in late 2017, the legendary physicist Stephen</p><p>Hawking offered his opinion about the future of AI.On the one hand, he was</p><p>hopeful that the technology could outpace human intelligence. This would</p><p>likely mean that many horrible diseases will be cured and perhaps there will</p><p>be ways to deal with environmental problems, including climate change.</p><p>But there was the dark side as well. Hawking talked about how the technology</p><p>had the potential to be the “worst event in the history of our civilization.”1</p><p>Just some of the problems include mass unemployment and even killer robots.</p><p>Because of this, he urged for ways to control AI.</p><p>Hawking’s ideas are certainly not on the fringe. Prominent tech entrepreneurs</p><p>like Elon Musk and Bill Gates also have expressed deep worries about AI.</p><p>Yet there are many who are decidedly optimistic, if not exuberant. Masayoshi</p><p>Son, who is the CEO of SoftBank and the manager of the $100 billion Vision</p><p>venture fund, is one of them. In an interview with CNBC, he proclaimed that</p><p>within 30 years, we’ll have flying cars, people will be living much longer, and</p><p>we’ll have cured many diseases.2 He also noted that the main focus of his fund</p><p>is on AI.</p><p>1 www.cnbc.com/2017/11/06/stephen-hawking-ai-could-be-worst-event-in-</p><p>civilization.html</p><p>2 www.cnbc.com/2019/03/08/softbank-ceo-ai-will-completely-change-the-way-</p><p>humans-live-within-30-years.html</p><p>http://www.cnbc.com/2017/11/06/stephen-hawking-ai-could-be-worst-event-in-civilization.html</p><p>http://www.cnbc.com/2017/11/06/stephen-hawking-ai-could-be-worst-event-in-civilization.html</p><p>http://www.cnbc.com/2019/03/08/softbank-ceo-ai-will-completely-change-the-way-humans-live-within-30-years.html</p><p>http://www.cnbc.com/2019/03/08/softbank-ceo-ai-will-completely-change-the-way-humans-live-within-30-years.html</p><p>162</p><p>OK then, who is right? Will the future be dystopian or utopian? Or will it be</p><p>somewhere in the middle? Well, predicting new technologies is exceedingly</p><p>difficult, almost impossible. Here are some examples of forecasts that have</p><p>been wide off the mark:</p><p>• Thomas Edison declared that AC (alternating current)</p><p>would fail.3</p><p>• In his book The Road Ahead (published in late 1995), Bill</p><p>Gates did not mention the Internet.</p><p>• In 2007, Jim Balsillie, the co-CEO of Research in Motion</p><p>(the creator of the BlackBerry device), said that the</p><p>iPhone would get little traction.4</p><p>• In the iconic science fiction movie Blade Runner—</p><p>released in 1982 and was set in 2019—there were many</p><p>predictions that were wrong like phone booths with</p><p>video phones and androids (or “replicants”) that were</p><p>nearly indistinguishable from humans.</p><p>Despite all this, there is one thing that is certain: In the coming years, we’ll see</p><p>lots of innovation and change from AI.This seems inevitable, especially since</p><p>there continues to be huge amounts invested in the industry.</p><p>So then, let’s take a look at some of the areas that are likely to have an</p><p>outsized impact on society.</p><p>Autonomous Cars</p><p>When it comes to AI, one of the most far-reaching areas is autonomous cars.</p><p>Interestingly enough, this category is not really new. Yes, it’s been a hallmark</p><p>of lots of science fiction stories for many decades! But for some time, there</p><p>have been many real-life examples of innovation, like the following:</p><p>• Stanford Cart: Its development started in the early 1960s,</p><p>and the original goal was to create a remote-controlled</p><p>vehicle for moon missions. But the researchers eventually</p><p>changed their focus and developed a basic autonomous</p><p>vehicle, which used cameras and AI for navigation. While</p><p>it was a standout achievement for the era, it was not</p><p>practical as it required more than 10 minutes to plan for</p><p>any move!</p><p>3 www.msn.com/en-us/news/technology/the-best-and-worst-technology-predictions-</p><p>of-all-time/ss-BBIMwm3#image=5</p><p>4 www.recode.net/2017/1/9/14215942/iphone-steve-jobs-apple-ballmer-</p><p>nokia-anniversary</p><p>Chapter 9 | The Future ofAI</p><p>http://www.msn.com/en-us/news/technology/the-best-and-worst-technology-predictions-of-all-time/ss-BBIMwm3#image=5</p><p>http://www.msn.com/en-us/news/technology/the-best-and-worst-technology-predictions-of-all-time/ss-BBIMwm3#image=5</p><p>http://www.recode.net/2017/1/9/14215942/iphone-steve-jobs-apple-ballmer-nokia-anniversary</p><p>http://www.recode.net/2017/1/9/14215942/iphone-steve-jobs-apple-ballmer-nokia-anniversary</p><p>163</p><p>• Ernst Dickmanns: A brilliant German aerospace engineer,</p><p>he would turn his attention to the idea of converting</p><p>a Mercedes van into an autonomous vehicle…in the</p><p>mid-1980s. He wired together cameras, sensors,</p><p>and computers. He also was creative in how he used</p><p>software, such as by only focusing the graphics processing</p><p>on important visual details to save on power. By doing all</p><p>this, he was able to develop a system that would control</p><p>a car’s steering, gas pedal, and brakes. He tested the</p><p>Mercedes on a Paris highway—in 1994—and it went over</p><p>600 miles, with a speed up to 81 MPH.5 Nevertheless, the</p><p>research funding was pulled because it was far from clear</p><p>if</p><p>there could be commercialization in a timely manner.</p><p>It also did not help that AI was entering another winter.</p><p>But the inflection point for autonomous cars came in 2004. The main catalyst</p><p>was the Iraq War, which was taking a horrible toll on American soldiers. For</p><p>DARPA, the belief was that autonomous vehicles could be a solution.</p><p>But the agency faced many tough challenges. This is why it set up a contest,</p><p>dubbed the DARPA Grand Challenge, in 2004, which had a $1 million grand</p><p>prize to encourage wider innovation. The event involved a 150-mile race in</p><p>the Mojave Desert, and unfortunately, it was not encouraging as the cars</p><p>performed miserably. None of them finished the race!</p><p>But this only spurred even more innovation. By the next year, five cars finished</p><p>the race. Then in 2007, the cars were so advanced that they were able to take</p><p>actions like U-turns and merging.</p><p>Through this process, DARPA was able to allow for the creation of the key</p><p>components for autonomous vehicles:</p><p>• Sensors: These include radar and ultrasonic systems that</p><p>can detect vehicles and other obstacles, such as curbs.</p><p>• Video Cameras: These can detect road signs, traffic lights,</p><p>and pedestrians.</p><p>• Lidar (Light Detection and Ranging): This device—which is</p><p>usually at the top of an autonomous car—shoots laser</p><p>beams to measure the surroundings. The data is then</p><p>integrated into existing maps.</p><p>• Computer: This helps with the control of the car, including</p><p>the steering, acceleration, and braking. The system</p><p>leverages AI to learn but also has built-in rules for avoiding</p><p>objects, obeying the laws, and so on.</p><p>5 www.politico.eu/article/delf-driving-car-born-1986-ernst-dickmanns-</p><p>mercedes/</p><p>Artificial Intelligence Basics</p><p>http://www.politico.eu/article/delf-driving-car-born-1986-ernst-dickmanns-mercedes/</p><p>http://www.politico.eu/article/delf-driving-car-born-1986-ernst-dickmanns-mercedes/</p><p>164</p><p>Now when it comes to autonomous cars, there is lots of confusion of what</p><p>“autonomous” really means. Is it when a car drives itself completely alone—</p><p>or must there be a human driver?</p><p>To understand the nuances, there are five levels of autonomy:</p><p>• Level 0: This is where a human controls all the systems.</p><p>• Level 1: With this, computers control limited functions</p><p>like cruise control or braking—but only one at a time.</p><p>• Level 2: This type of car can automate two functions.</p><p>• Level 3: This is where a car automates all the safety functions.</p><p>But the driver can intervene if something goes wrong.</p><p>• Level 4: The car can generally drive itself. But there are</p><p>cases in which a human must participate.</p><p>• Level 5: This is the Holy Grail, in which the car is completely</p><p>autonomous.</p><p>The auto industry is one of the biggest markets, and AI is likely to unleash</p><p>wrenching changes. Consider that transportation is the second largest</p><p>household expenditure, behind housing, and twice as large as healthcare.</p><p>Something else to keep in mind: The typical car is used only about 5% of the</p><p>time as it is usually parked somewhere.6</p><p>In light ofthe enormous opportunity for improvement, it should be no surprise</p><p>that the autonomous car industry has seen massive amounts of investment. This</p><p>has not only been about venture capitalists investing in a myriad of startups but</p><p>also innovation from traditional automakers like Ford, GM, and BMW.</p><p>Then when might we see this industry become mainstream? The estimates</p><p>vary widely. But according to a study from Allied Market Research, the market</p><p>is forecasted to hit $556.67 billion by 2026, which would represent a compound</p><p>annual growth rate of 39.47%.7</p><p>But there is still much to work out. “At best, we are still years away from a</p><p>car that doesn’t require a steering wheel,” said Scott Painter, who is the CEO</p><p>and founder of Fair. “Cars will still need to be insured, repaired, and maintained,</p><p>even if you came back from the future in a Delorean and brought the manual</p><p>for how to make these cars fully autonomous. We make 100 million cars-per-</p><p>year, of which 16 million-a-year are in the U.S.And, supposing you wanted the</p><p>whole supply to have these artificial intelligence features, it would still take 20</p><p>years until we had more cars on the road including all the different levels of</p><p>A.I. versus the number of cars that didn’t have those technologies.”8</p><p>6 www.sec.gov/Archives/edgar/data/1759509/000119312519077391/d633517ds1a.htm</p><p>7 www.alliedmarketresearch.com/autonomous-vehicle-market</p><p>8 From the author’s interview, in May 2019, with Scott Painter, who is the CEO and founder</p><p>of Fair.</p><p>Chapter 9 | The Future ofAI</p><p>http://www.sec.gov/Archives/edgar/data/1759509/000119312519077391/d633517ds1a.htm</p><p>http://www.alliedmarketresearch.com/autonomous-vehicle-market</p><p>165</p><p>But there are many other factors to keep in mind. After all, the fact remains that</p><p>driving is complex, especially in urban and suburban areas. What if a traffic sign</p><p>is changed or even manipulated? How about if an autonomous car must deal</p><p>with a dilemma like having to decide to crash into an oncoming car or plunging</p><p>into a curb, which may have pedestrians? All these are extremely difficult.</p><p>Evening seemingly simple tasks can be tough to pull off. John Krafcik, who is</p><p>the CEO of Google’s Waymo, points out that parking lots are a prime</p><p>example.9 They require finding available spots, avoiding other cars and</p><p>pedestrians (that can be unpredictable), and moving into the space.</p><p>But technology is just one of the challenges with autonomous vehicles. Here</p><p>are some others to consider:</p><p>• Infrastructure: Our cities and towns are built for traditional</p><p>cars. But by mixing autonomous vehicles, there will</p><p>probably be many logistical issues. How does a car</p><p>anticipate the actions of human drivers? Actually, there</p><p>may be a need to install sensors alongside roads. Or</p><p>another option is to have separate roads for autonomous</p><p>vehicles. Governments also will probably need to change</p><p>driver’s ed, providing guidance on how to interact with</p><p>autonomous vehicles while on the road.</p><p>• Regulation: This is a big wild card. For the most part, this</p><p>may be the biggest impediment as governments tend to</p><p>work slowly and are resistant to change. The United</p><p>States is also a highly litigious country—which may be</p><p>another factor that could curb development.</p><p>• Adoption: Autonomous vehicles will probably not be cheap,</p><p>as systems like Lidar are costly. This will certainly be a</p><p>limiting factor. But at the same time, there are indications</p><p>of skepticism from the general public. According to a</p><p>survey from AAA, about 71% of the respondents said</p><p>they are afraid of riding in an autonomous vehicle.10</p><p>Given all this, the initial phase of autonomous vehicles will probably be for</p><p>controlled situations, say for trucking, mining, or shuttles. A case of this is</p><p>Suncor Energy, which uses autonomous trucks for excavating various sites in</p><p>Canada.</p><p>Ride-sharing networks—like Uber and Lyft—may be another starting point.</p><p>These services are fairly structured and understandable to the public.</p><p>9 www.businessinsider.com/waymo-ceo-john-krafcik-explains-big-challenge-</p><p>for-self-driving-cars-2019-4</p><p>10 https://newsroom.aaa.com/2019/03/americans-fear-self-driving-cars-survey/</p><p>Artificial Intelligence Basics</p><p>http://www.businessinsider.com/waymo-ceo-john-krafcik-explains-big-challenge-for-self-driving-cars-2019-4</p><p>http://www.businessinsider.com/waymo-ceo-john-krafcik-explains-big-challenge-for-self-driving-cars-2019-4</p><p>https://newsroom.aaa.com/2019/03/americans-fear-self-driving-cars-survey/</p><p>166</p><p>Keep in mind that Waymo has been testing a self-driving taxi service in</p><p>Phoenix (this is similar to a ride-sharing system like Uber, but the cars have</p><p>autonomous systems). Here’s how a blog post from the company explains it:</p><p>We’ll start by giving riders access to our app. They can use it to call our</p><p>self-driving vehicles 24 hours a day, 7 days a week. They can ride across</p><p>several cities in the Metro Phoenix area, including Chandler, Tempe,</p><p>Mesa, and Gilbert. Whether it’s for a fun night out or just</p><p>to get a break</p><p>from driving, our riders get the same clean vehicles every time and our</p><p>Waymo driver with over 10 million miles of experience on public roads.</p><p>Riders will see price estimates before they accept the trip based on</p><p>factors like the time and distance to their destination.11</p><p>Waymo has found that a key is education because the riders have lots of</p><p>questions. To deal with this, the company has built in a chat system in the app</p><p>to contact a support person. The dashboard of the car also has a screen that</p><p>provides details of the ride.</p><p>According to the blog post, “Feedback from riders will continue to be vital</p><p>every step of the way.”12</p><p>US vs. China</p><p>The rapid ascent of China has been astonishing. Within a few years, the</p><p>economy may be larger than the United States, and a key part of the growth</p><p>will be AI. The Chinese government has set forth the ambitious goal of</p><p>spending $150 billion on this technology through 2030.13 In the meantime,</p><p>there will continue to be major investments from companies like Baidu,</p><p>Alibaba, and Tencent.</p><p>Even though China is often considered to not be as creative or innovative as</p><p>Silicon Valley—often tagged as “copycats”—this perception may prove to be</p><p>a myth. A study from the Allen Institute for Artificial Intelligence highlights</p><p>that China is expected to outrank the United States in the most cited technical</p><p>papers on AI.14</p><p>The country has some other advantages, which AI expert and venture</p><p>capitalist Kai-Fu Lee has pointed out in his provocative book, AI Superpowers:</p><p>China, Silicon Valley, and the New World Order15:</p><p>11 https://medium.com/waymo/riding-with-waymo-one-today-9ac8164c5c0e</p><p>12 Ibid.</p><p>13 www.diamandis.com/blog/rise-of-ai-in-china</p><p>14 www.theverge.com/2019/3/14/18265230/china-is-about-to-overtake-america-</p><p>in-ai-research</p><p>15 New York: Houghton Mifflin Harcourt, 2018.</p><p>Chapter 9 | The Future ofAI</p><p>https://medium.com/waymo/riding-with-waymo-one-today-9ac8164c5c0e</p><p>http://www.diamandis.com/blog/rise-of-ai-in-china</p><p>http://www.theverge.com/2019/3/14/18265230/china-is-about-to-overtake-america-in-ai-research</p><p>http://www.theverge.com/2019/3/14/18265230/china-is-about-to-overtake-america-in-ai-research</p><p>167</p><p>• Enthusiasm: Back in the 1950s, Russia’s launch of Sputnik</p><p>sparked interest in people in the United States to become</p><p>engineers for the space program. Something similar has</p><p>actually happened in China. When the country’s top Go</p><p>player, Ke Jie, lost to the AlphaGo AI system, this was a</p><p>wake-up call. The result is that this has inspired many</p><p>young people to pursue a career in AI.</p><p>• Data: With a population of over 1.3 billion, China is rich</p><p>with data (there are more than 700 million Internet</p><p>users). But the country’s authoritarian government is</p><p>also critical as privacy is not considered particularly</p><p>important, which means there is much more leeway</p><p>when developing AI models. For example, in a paper</p><p>published in Nature Medicine, the Chinese researchers</p><p>had access to data on 600,000 patients to conduct a</p><p>healthcare study.16 While still in the early stages, it</p><p>showed that an AI model was able to effectively diagnose</p><p>childhood conditions like the flu and meningitis.</p><p>• Infrastructure: As a part of the Chinese government’s</p><p>investment plans, there has been a focus on creating</p><p>next-generation cities that allow for autonomous cars</p><p>and other AI systems. There has also been an aggressive</p><p>rollout of 5G networks.</p><p>As for the United States, the government has been much more tentative with</p><p>AI.President Trump has signed an executive order—called the “American AI</p><p>Initiative”—to encourage development of the technology, but the terms are</p><p>vague and it is far from clear how much money will be committed to it.</p><p>Technological Unemployment</p><p>The concept of technological unemployment, which gained notoriety from</p><p>famed economist John Maynard Keynes during the Great Depression, explains</p><p>how innovations can lead to long-term job loss. However, evidence of this has</p><p>been elusive. Notwithstanding the fact that automation has severely impacted</p><p>industries like manufacturing, there is often a transition of the workforce as</p><p>people adapt.</p><p>But could the AI revolution be different? It very well could. For example,</p><p>California Governor Gavin Newsom fears that his state could see massive</p><p>unemployment in areas like trucking and warehousing—and soon.17</p><p>16 www.nature.com/articles/s41591-018-0335-9</p><p>17 www.mercurynews.com/2019/03/18/were-not-prepared-for-the-promise-of-</p><p>artificial-intelligence-experts-warn/</p><p>Artificial Intelligence Basics</p><p>http://www.nature.com/articles/s41591-018-0335-9</p><p>http://www.mercurynews.com/2019/03/18/were-not-prepared-for-the-promise-of-artificial-intelligence-experts-warn/</p><p>http://www.mercurynews.com/2019/03/18/were-not-prepared-for-the-promise-of-artificial-intelligence-experts-warn/</p><p>168</p><p>Here’s another example: Harvest CROO Robotics has built a robot, called</p><p>Harv, that can pick strawberries and other plants without causing bruises.</p><p>Granted, it is still in the experimental phase, but the system is quickly</p><p>improving. The expectation is that one robot will do the work of 30 people.18</p><p>And of course, there will be no wages to pay or labor liability exposure.</p><p>But AI may mean more than replacing low-skilled jobs. There are already signs</p><p>that the technology could have a major impact on white-collar professions.</p><p>Let’s face it, there is even more incentive to automate these jobs because they</p><p>fetch higher compensation.</p><p>Just one category that could face AI job loss is the legal field, as a variety of</p><p>startups are gunning for the market like Lawgood, NexLP, and RAVN ACE.The</p><p>solutions are focused on automating areas such as legal research and contract</p><p>review.19 Even though the systems are far from perfect, they can certainly</p><p>process much more volume than people—and can also get smarter as they</p><p>are used more and more.</p><p>True, the overall job market is dynamic, and there will be new types of careers</p><p>that will be created. There will also likely be AI innovations that are assistive</p><p>for employees—making their job easier to do. For example, software startup</p><p>Measure Square has been able to use sophisticated algorithms to convert</p><p>paper-based floorplans into digitally interactive floorplans. Because of this, it</p><p>has been easier to get projects started and completed on time.</p><p>However, in light of the potential transformative impact of AI, it does seem</p><p>reasonable that there will be an adverse impact on a broad range of industries.</p><p>Perhaps a foreshadowing of this is what happened with job losses from</p><p>manufacturing in the 1960s to 1990s. According to the Pew Research Center,</p><p>there has been virtually no real wage growth in the last 40 years.20 During this</p><p>period, the United States has also experienced a widening gap in wealth.</p><p>Berkeley economist Gabriel Zucman estimates that 0.1% of the population</p><p>controls nearly 20% of the wealth.21</p><p>18 www.washingtonpost.com/news/national/wp/2019/02/17/feature/inside-the-</p><p>race-to-replace-farmworkers-with-robots/</p><p>19 www.cnbc.com/2017/02/17/lawyers-could-be-replaced-by-artificial-</p><p>intelligence.html</p><p>20 www.pewresearch.org/fact-tank/2018/08/07/for-most-us-workers-real-</p><p>wages-have-barely-budged-for-decades/</p><p>21 http://fortune.com/2019/02/08/growing-wealth-inequality-us-study/</p><p>Chapter 9 | The Future ofAI</p><p>https://www.washingtonpost.com/news/national/wp/2019/02/17/feature/inside-the-race-to-replace-farmworkers-with-robots/</p><p>https://www.washingtonpost.com/news/national/wp/2019/02/17/feature/inside-the-race-to-replace-farmworkers-with-robots/</p><p>https://www.cnbc.com/2017/02/17/lawyers-could-be-replaced-by-artificial-intelligence.html</p><p>https://www.cnbc.com/2017/02/17/lawyers-could-be-replaced-by-artificial-intelligence.html</p><p>http://www.pewresearch.org/fact-tank/2018/08/07/for-most-us-workers-real-wages-have-barely-budged-for-decades/</p><p>http://www.pewresearch.org/fact-tank/2018/08/07/for-most-us-workers-real-wages-have-barely-budged-for-decades/</p><p>http://fortune.com/2019/02/08/growing-wealth-inequality-us-study/</p><p>169</p><p>Yet there are actions that can be taken. First of all, governments can look to</p><p>provide education and transition assistance. With the pace of change in</p><p>today’s world, there will need to be ongoing renewal of skills for most people.</p><p>IBM CEO Ginni Rometty has noted that AI will change all jobs within the next</p><p>5–10 years. By the way, her company has seen a 30% reduction of headcount</p><p>in the HR department because of automation.22</p><p>Next, there are some people who advocate basic income, which provides a</p><p>minimum amount of compensation to everyone. This would certainly soften</p><p>some of the inequality, but it also has drawbacks. People definitely get pride</p><p>and satisfaction from their careers. So what might a person’s morale be if he</p><p>or she cannot find a job? It could have a profound impact.</p><p>Finally, there is even talk of some type of AI tax. This would essentially claw</p><p>back the large gains from those companies that benefit from the technology.</p><p>Although, given their power, it probably would be tough to pass this type of</p><p>legislation.</p><p>The Weaponization ofAI</p><p>The Air Force Research Lab is working on prototypes for something called</p><p>Skyborg. It’s right out of Star Wars. Think of Skyborg as R2-D2 that serves as an</p><p>AI wingman for a fighter jet, helping to identify targets and threats.23 The AI</p><p>robot may also be able to take control if the pilot is incapacitated or distracted.</p><p>The Air Force is even looking at using the technology to operate drones.</p><p>Cool, huh? Certainly. But there is a major issue: By using AI, might humans</p><p>ultimately be taken out of the loop when making life-and-death decisions on</p><p>the battlefield? Could this ultimately lead to more bloodshed? Perhaps the</p><p>machines will make the wrong decisions—causing even more problems?</p><p>Many AI researchers and entrepreneurs are concerned. To this end, more</p><p>than 2,400 have signed a statement that calls for a ban of so-called robot</p><p>killers.24</p><p>Even the United Nations is exploring some type of ban. But the United States,</p><p>along with Australia, Israel, the United Kingdom, and Russia, have resisted this</p><p>move.25 As a result, there may be a true AI arms race emerging.</p><p>22 www.cnbc.com/2019/04/03/ibm-ai-can-predict-with-95-percent-accuracy-</p><p>which-employees-will-quit.html</p><p>23 www.popularmechanics.com/military/aviation/a26871027/air-force-ai-</p><p>fighter-plane-skyborg/</p><p>24 www.theguardian.com/science/2018/jul/18/thousands-of-scientists-pledge-</p><p>not-to-help-build-killer-ai-robots</p><p>25 www.theguardian.com/science/2019/mar/29/uk-us-russia-opposing-killer-</p><p>robot-ban-un-ai</p><p>Artificial Intelligence Basics</p><p>http://www.cnbc.com/2019/04/03/ibm-ai-can-predict-with-95-percent-accuracy-which-employees-will-quit.html</p><p>http://www.cnbc.com/2019/04/03/ibm-ai-can-predict-with-95-percent-accuracy-which-employees-will-quit.html</p><p>http://www.popularmechanics.com/military/aviation/a26871027/air-force-ai-fighter-plane-skyborg/</p><p>http://www.popularmechanics.com/military/aviation/a26871027/air-force-ai-fighter-plane-skyborg/</p><p>http://www.theguardian.com/science/2018/jul/18/thousands-of-scientists-pledge-not-to-help-build-killer-ai-robots</p><p>http://www.theguardian.com/science/2018/jul/18/thousands-of-scientists-pledge-not-to-help-build-killer-ai-robots</p><p>http://www.theguardian.com/science/2019/mar/29/uk-us-russia-opposing-killer-robot-ban-un-ai</p><p>http://www.theguardian.com/science/2019/mar/29/uk-us-russia-opposing-killer-robot-ban-un-ai</p><p>170</p><p>According to a paper from the RAND Corporation, there is even the potential</p><p>that the technology could lead to nuclear war, say by the year 2040. How?</p><p>The authors note that AI may make it easier to target submarines and mobile</p><p>missile systems. According to the report:</p><p>Nations may be tempted to pursue first-strike</p><p>capabilities as a means of gaining bargaining leverage</p><p>over their rivals even if they have no intention of</p><p>carrying out an attack, researchers say. This undermines</p><p>strategic stability because even if the state possessing</p><p>these capabilities has no intention of using them, the</p><p>adversary cannot be sure of that.26</p><p>But in the near term, AI will probably have the most impact on information</p><p>warfare, which could still be highly destructive. We got a glimpse of this when</p><p>the Russian government interfered with the 2016 presidential election. The</p><p>approach was fairly low-tech as it used social media troll farms to disseminate</p><p>fake news—but the consequences were significant.</p><p>But as AI gets more powerful and becomes more affordable, we’ll likely see it</p><p>supercharge these kinds of campaigns. For example, deepfake systems can</p><p>easily create life-like photos and videos of people that could be used to quickly</p><p>spread messages.</p><p>Drug Discovery</p><p>The advances in drug discovery have been almost miraculous as we now have</p><p>cures for such intractable diseases like hepatitis C and have continued to</p><p>make strides with a myriad of cancers. But of course, there is certainly much</p><p>that needs to be done. The fact is that drug companies are having more</p><p>troubles coming up with treatments. Here’s just one example: In March 2019,</p><p>Biogen announced that one of its drugs for Alzheimer’s, which was in Phase III</p><p>trials, failed to show meaningful results. On the news, the company’s shares</p><p>plunged by 29%, wiping out $18 billion of market value.27</p><p>Consider that traditional drug development often involves much trial and</p><p>error, which can be time consuming. Then might there be a better way?</p><p>Increasingly, researchers are looking to AI for help. We are seeing a variety of</p><p>startups spring up that are focusing on the opportunity.</p><p>26 www.rand.org/news/press/2018/04/24.html</p><p>27 www.wsj.com/articles/biogen-shares-drop-28-after-ending-</p><p>alzheimers-phase-3-trials-11553170765</p><p>Chapter 9 | The Future ofAI</p><p>http://www.rand.org/news/press/2018/04/24.html</p><p>http://www.wsj.com/articles/biogen-shares-drop-28-after-ending-alzheimers-phase-3-trials-11553170765</p><p>http://www.wsj.com/articles/biogen-shares-drop-28-after-ending-alzheimers-phase-3-trials-11553170765</p><p>171</p><p>One is Insitro. The company, which got its start in 2019, had little trouble</p><p>raising a staggering $100 million in its Series A round. Some of the investors</p><p>included Alexandria Venture Investments, Bezos Expeditions (which is the</p><p>investment firm of Amazon.com’s Jeff Bezos), Mubadala Investment Company,</p><p>Two Sigma Ventures, and Verily.</p><p>Even though the team is relatively small—with about 30 employees—they all</p><p>are brilliant researchers who span areas like data science, deep learning,</p><p>software engineering, bioengineering, and chemistry. The CEO and founder,</p><p>Daphne Koller, has the rare blend of experience in advanced computer science</p><p>and health sciences, having led Google’s healthcare business, Calico.</p><p>As a testament to Insitro’s prowess, the company has already struck a partnership</p><p>with mega drug operator Gilead. It involves potential payments of over $1 billion</p><p>for research on nonalcoholic steatohepatitis (NASH), which is a serious liver</p><p>disease.28 A key is that Gilead has been able to assemble a large amount of data,</p><p>which can train the models. This will be done using cells outside of a person’s</p><p>body—that is, with an invitro system. Gilead has some urgency for looking at</p><p>alternative approaches since one of its NASH treatments, selonsertib, failed in</p><p>its clinical trials (it was for those who had the disease in the later stages).</p><p>The promise of AI is that it will speed up drug discovery because deep learning</p><p>should be able to identify complex patterns. But the technology could also turn</p><p>out to be helpful in developing personalized treatments—such as geared to a</p><p>person’s genetic make-up—which is likely to be critical for curing certain diseases.</p><p>Regardless, it is probably best to temper expectations. There will be major</p><p>hurdles to deal with as the healthcare industry will need to undergo changes</p><p>because there will be increased education for AI. This will take time, and</p><p>there will likely be resistance.</p><p>Next, deep learning is generally a “black box” when it comes to understanding</p><p>how the algorithms really</p><p>to proceed on the basis of the conjecture that every aspect</p><p>of learning or any other feature of intelligence can in principle be so</p><p>precisely described that a machine can be made to simulate it. An</p><p>attempt will be made to find how to make machines use language, form</p><p>abstractions and concepts, solve kinds of problems now reserved for</p><p>humans, and improve themselves. We think that a significant advance</p><p>can be made in one or more of these problems if a carefully selected</p><p>group of scientists work on it together for a summer.5</p><p>At the conference, Allen Newell, Cliff Shaw, and Herbert Simon demoed a</p><p>computer program called the Logic Theorist, which they developed at the</p><p>Research and Development (RAND) Corporation. The main inspiration came</p><p>from Simon (who would win the Nobel Prize in Economics in 1978). When</p><p>he saw how computers printed out words on a map for air defense systems,</p><p>he realized that these machines could be more than just about processing</p><p>numbers. It could also help with images, characters, and symbols—all of which</p><p>could lead to a thinking machine.</p><p>Regarding Logic Theorist, the focus was on solving various math theorems</p><p>from Principia Mathematica. One of the solutions from the software turned</p><p>out to be more elegant—and the co-author of the book, Bertrand Russell,</p><p>was delighted.</p><p>5 www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html</p><p>Chapter 1 | AI Foundations</p><p>https://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html/</p><p>7</p><p>Creating the Logic Theorist was no easy feat. Newell, Shaw, and Simon used</p><p>an IBM 701, which used machine language. So they created a high-level</p><p>language, called IPL (Information Processing Language), that sped up the</p><p>programming. For several years, it was the language of choice for AI.</p><p>The IBM 701 also did not have enough memory for the Logic Theorist. This</p><p>led to another innovation: list processing. It allowed for dynamically allocating</p><p>and deallocating memory as the program ran.</p><p>Bottom line: The Logic Theorist is considered the first AI program ever</p><p>developed.</p><p>Despite this, it did not garner much interest! The Dartmouth conference was</p><p>mostly a disappointment. Even the phrase “artificial intelligence” was criticized.</p><p>Researchers tried to come up with alternatives, such as “complex information</p><p>processing.” But they were not catchy like AI was—and the term stuck.</p><p>As for McCarthy, he continued on his mission to push innovation in</p><p>AI.Consider the following:</p><p>• During the late 1950s, he developed the Lisp</p><p>programming language, which was often used for AI</p><p>projects because of the ease of using nonnumerical</p><p>data. He also created programming concepts like</p><p>recursion, dynamic typing, and garbage collection. Lisp</p><p>continues to be used today, such as with robotics and</p><p>business applications. While McCarthy was developing</p><p>the language, he also co-founded the MIT Artificial</p><p>Intelligence Laboratory.</p><p>• In 1961, he formulated the concept of time-sharing of</p><p>computers, which had a transformative impact on the</p><p>industry. This also led to the development of the Internet</p><p>and cloud computing.</p><p>• A few years later, he founded Stanford’s Artificial</p><p>Intelligence Laboratory.</p><p>• In 1969, he wrote a paper called “Computer-Controlled</p><p>Cars,” in which he described how a person could enter</p><p>directions with a keyboard and a television camera would</p><p>navigate the vehicle.</p><p>• He won the Turing Award in 1971. This prize is considered</p><p>the Nobel Prize for Computer Science.</p><p>Artificial Intelligence Basics</p><p>8</p><p>In a speech in 2006, McCarthy noted that he was too optimistic about the</p><p>progress of strong AI. According to him, “we humans are not very good at</p><p>identifying the heuristics we ourselves use.”6</p><p>Golden Age ofAI</p><p>From 1956 to 1974, the AI field was one of the hottest spots in the tech world.</p><p>A major catalyst was the rapid development in computer technologies. They</p><p>went from being massive systems—based on vacuum tubes—to smaller systems</p><p>run on integrated circuits that were much quicker and had more storage capacity.</p><p>The federal government was also investing heavily in new technologies. Part</p><p>of this was due to the ambitious goals of the Apollo space program and the</p><p>heavy demands of the Cold War.</p><p>As for AI, the main funding source was the Advanced Research Projects</p><p>Agency (ARPA), which was launched in the late 1950s after the shock of</p><p>Russia’s Sputnik. The spending on projects usually came with few requirements.</p><p>The goal was to inspire breakthrough innovation. One of the leaders of ARPA,</p><p>J.C. R.Licklider, had a motto of “fund people, not projects.” For the most</p><p>part, the majority of the funding was from Stanford, MIT, Lincoln Laboratories,</p><p>and Carnegie Mellon University.</p><p>Other than IBM, the private sector had little involvement in AI development.</p><p>Keep in mind that—by the mid-1950s—IBM would pull back and focus on the</p><p>commercialization of its computers. There was actually fear from customers</p><p>that this technology would lead to significant job losses. So IBM did not want</p><p>to be blamed.</p><p>In other words, much of the innovation in AI spun out from academia. For</p><p>example, in 1959, Newell, Shaw, and Simon continued to push the boundaries</p><p>in the AI field with the development of a program called “General Problem</p><p>Solver.” As the name implied, it was about solving math problems, such as the</p><p>Tower of Hanoi.</p><p>But there were many other programs that attempted to achieve some level of</p><p>strong AI. Examples included the following:</p><p>• SAINT or Symbolic Automatic INTegrator (1961): This</p><p>program, created by MIT researcher James Slagle, helped</p><p>to solve freshman calculus problems. It would be updated</p><p>into other programs, called SIN and MACSYMA, that did</p><p>much more advanced math. SAINT was actually the first</p><p>example of an expert system, a category of AI we’ll cover</p><p>later in this chapter.</p><p>6 www.technologyreview.com/s/425913/computing-pioneer-dies/</p><p>Chapter 1 | AI Foundations</p><p>http://www.technologyreview.com/s/425913/computing-pioneer-dies/</p><p>9</p><p>• ANALOGY (1963): This program was the creation of MIT</p><p>professor Thomas Evans. The application demonstrated</p><p>that a computer could solve analogy problems of an IQ</p><p>test.</p><p>• STUDENT (1964): Under the supervision of Minsky at</p><p>MIT, Daniel Bobrow created this AI application for his</p><p>PhD thesis. The system used Natural Language Processing</p><p>(NLP) to solve algebra problems for high school students.</p><p>• ELIZA (1965): MIT professor Joseph Weizenbaum</p><p>designed this program, which instantly became a big hit.</p><p>It even got buzz in the mainstream press. It was named</p><p>after Eliza (based on George Bernard Shaw’s play</p><p>Pygmalion) and served as a psychoanalyst. A user could</p><p>type in questions, and ELIZA would provide counsel (this</p><p>was the first example of a chatbot). Some people who</p><p>used it thought the program was a real person, which</p><p>deeply concerned Weizenbaum since the underlying</p><p>technology was fairly basic. You can find examples of</p><p>ELIZA on the web, such as at http://psych.fullerton.</p><p>edu/mbirnbaum/psych101/Eliza.htm.</p><p>• Computer Vision (1966): In a legendary story, MIT’s Marvin</p><p>Minsky said to a student, Gerald Jay Sussman, to spend</p><p>the summer linking a camera to a computer and getting</p><p>the computer to describe what it saw. He did just that</p><p>and built a system that detected basic patterns. It was the</p><p>first use of computer vision.</p><p>• Mac Hack (1968): MIT professor Richard D.Greenblatt</p><p>created this program that played chess. It was the first to</p><p>play in real tournaments and got a C-rating.</p><p>• Hearsay I (Late 1960s): Professor Raj Reddy developed</p><p>a continuous speech recognition system. Some of his</p><p>students would then go on to create Dragon Systems,</p><p>which became a major tech company.</p><p>During this period, there was a proliferation of AI academic papers and books.</p><p>Some of the topics included Bayesian methods, machine learning, and vision.</p><p>But there were generally two major theories about AI.One was led by Minsky,</p><p>who said that there needed to be symbolic systems. This meant that AI should</p><p>be based on</p><p>work. This could prove difficult in getting regulatory</p><p>approval for new drugs as the FDA focuses on causal relationships.</p><p>Finally, the human body is highly sophisticated, and we still are learning about</p><p>how it works. And besides, as we have seen with innovations like the decoding</p><p>of the Human Genome, it usually takes considerable time to understand new</p><p>approaches.</p><p>As a sign of the complexities, consider the situation of IBM’s Watson. Even</p><p>though the company has some of the most talented AI researchers and has</p><p>spent billions on the technology, it recently announced that it would no longer</p><p>sell Watson for drug discovery purposes.29</p><p>28 www.fiercebiotech.com/biotech/stealthy-insitro-opens-up-starting-</p><p>gilead-deal-worth-up-to-1-05b</p><p>29 https://khn.org/morning-breakout/ups-and-downs-of-artificial-</p><p>intelligence-ibm-stops-sales-development-of-watson-for-drug-discovery-</p><p>hospitals-learn-from-ehrs/</p><p>Artificial Intelligence Basics</p><p>http://www.fiercebiotech.com/biotech/stealthy-insitro-opens-up-starting-gilead-deal-worth-up-to-1-05b</p><p>http://www.fiercebiotech.com/biotech/stealthy-insitro-opens-up-starting-gilead-deal-worth-up-to-1-05b</p><p>https://khn.org/morning-breakout/ups-and-downs-of-artificial-intelligence-ibm-stops-sales-development-of-watson-for-drug-discovery-hospitals-learn-from-ehrs/</p><p>https://khn.org/morning-breakout/ups-and-downs-of-artificial-intelligence-ibm-stops-sales-development-of-watson-for-drug-discovery-hospitals-learn-from-ehrs/</p><p>https://khn.org/morning-breakout/ups-and-downs-of-artificial-intelligence-ibm-stops-sales-development-of-watson-for-drug-discovery-hospitals-learn-from-ehrs/</p><p>172</p><p>Government</p><p>An article from Bloomberg.com in April 2019 caused a big stir. It described a</p><p>behind-the-scenes look at how Amazon.com manages its Alexa speaker AI</p><p>system.30 While much of it is based on algorithms, there are also thousands of</p><p>people who analyze voice clips in order to help make the results better. Often</p><p>the focus is on dealing with the nuances of slang and regional dialects, which</p><p>have been difficult for deep learning algorithms.</p><p>But of course, it’s natural for people to wonder: Is my smart speaker really</p><p>listening to me? Are my conversations private?</p><p>Amazon.com was quick to point out that it has strict rules and requirements.</p><p>But even this ginned up even more concern! According to the Bloomberg.</p><p>com post, the AI reviewers would sometimes hear clips that involved</p><p>potentially criminal activity, such as sexual assault. But Amazon apparently has</p><p>a policy to not interfere.</p><p>As AI becomes more pervasive, we’ll have more of these kinds of stories; and</p><p>for the most part, there will not be clear-cut answers. Some people may</p><p>ultimately decide not to buy AI products. Yet this will probably be a small</p><p>group. Hey, even with the myriad of privacy issues with Facebook, there has</p><p>not been a decline in the user growth.</p><p>More likely, governments will start to wade in with AI issues. A group of</p><p>congresspersons have sponsored a bill, called the Algorithmic Accountability</p><p>Act, which aims to mandate that companies audit their AI systems (it would</p><p>be for larger companies, with revenues over $50 million and more than 1</p><p>million users).31 The law, if enacted, would be enforced by the Federal Trade</p><p>Commission.</p><p>There are also legislative moves from states and cities. In 2019, NewYork</p><p>City passed its own law to require more transparency with AI.32 There are</p><p>also efforts in Washington state, Illinois, and Massachusetts.</p><p>With all this activity, some companies are getting proactive, such as by</p><p>adopting their own ethics boards. Just look at Microsoft. The company’s ethics</p><p>board, called Aether (AI and Ethics in Engineering and Research), decided to</p><p>not allow the use of its facial recognition system for traffic stops in California.33</p><p>30 www.bloomberg.com/news/articles/2019-04-10/is-anyone-listening-to-</p><p>you-on-alexa-a-global-team-reviews-audio</p><p>31 www.theverge.com/2019/4/10/18304960/congress-algorithmic-</p><p>accountability-act-wyden-clarke-booker-bill-introduced-house-senate</p><p>32 www.wsj.com/articles/our-software-is-biased-like-we-are-can-</p><p>new-laws-change-that-11553313609?mod=hp_lead_pos8</p><p>33 www.geekwire.com/2019/policing-ai-task-industry-government-customers/</p><p>Chapter 9 | The Future ofAI</p><p>http://www.bloomberg.com/news/articles/2019-04-10/is-anyone-listening-to-you-on-alexa-a-global-team-reviews-audio</p><p>http://www.bloomberg.com/news/articles/2019-04-10/is-anyone-listening-to-you-on-alexa-a-global-team-reviews-audio</p><p>http://www.theverge.com/2019/4/10/18304960/congress-algorithmic-accountability-act-wyden-clarke-booker-bill-introduced-house-senate</p><p>http://www.theverge.com/2019/4/10/18304960/congress-algorithmic-accountability-act-wyden-clarke-booker-bill-introduced-house-senate</p><p>http://www.wsj.com/articles/our-software-is-biased-like-we-are-can-new-laws-change-that-11553313609?mod=hp_lead_pos8</p><p>http://www.wsj.com/articles/our-software-is-biased-like-we-are-can-new-laws-change-that-11553313609?mod=hp_lead_pos8</p><p>http://www.geekwire.com/2019/policing-ai-task-industry-government-customers/</p><p>173</p><p>In the meantime, we may see AI activism as well, in which people organize to</p><p>protest the use of certain applications. Again, Amazon.com has been the</p><p>target of this, with its Rekognition software that uses facial recognition to</p><p>help law enforcement identify suspects. The ACLU has raised concerns of</p><p>accuracy of the system, especially regarding women and minorities. In one of</p><p>its experiments, it found that Rekognition identified 28 members of the</p><p>Congress as having prior criminal records!34 As for Amazon.com, it has</p><p>disputed the claims.</p><p>Rekognition is only one among various AI applications in law enforcement</p><p>that are leading to controversy. Perhaps the most notable example is COMPAS</p><p>(Correctional Offender Management Profiling for Alternative Sanctions),</p><p>which uses analytics to gauge the probability of someone who may commit a</p><p>crime. The system is often used for sentencing. But the big issue is: Might this</p><p>violate a person’s constitutional right to due process since there is the real</p><p>risk that the AI will be incorrect or discriminatory? Actually, for now, there</p><p>are few good answers. But given the importance AI algorithms will play in our</p><p>justice system, it seems like a good bet that the Supreme Court will be making</p><p>new law.</p><p>AGI (Artificial General Intelligence)</p><p>In Chapter 1, we learned about the difference between strong and weak</p><p>AI.And for the most part, we are in the weak AI phase, in which the technology</p><p>is used for narrow categories.</p><p>As for strong AI, it’s about the ultimate: the ability for a machine to rival a</p><p>human. This is also known as Artificial General Intelligence or AGI.Achieving</p><p>this is likely many years away, perhaps something we may not see until the</p><p>next century or ever.</p><p>But of course, there are some brilliant researchers who believe that AGI will</p><p>come soon. One is Ray Kurzweil, who is an inventor, futurist, bestselling</p><p>author, and director of Engineering at Google. When it comes to AI, he has</p><p>left his imprint on the industry, such as with innovations in areas like text-to-</p><p>speech systems.</p><p>Kurzweil believes that AGI will happen—in which the Turing Test will be</p><p>cracked—in 2019, and then by 2045, there will be the Singularity. This is</p><p>where we’ll have a world of hybrid people: part human, part machine.</p><p>Kind of crazy? Perhaps so. But Kurzweil does have many high-profile followers.</p><p>34 www.businessinsider.com/ai-experts-call-on-amazon-not-to-sell-rekognition-</p><p>software-to-police-2019-4</p><p>Artificial Intelligence Basics</p><p>http://www.businessinsider.com/ai-experts-call-on-amazon-not-to-sell-rekognition-software-to-police-2019-4</p><p>http://www.businessinsider.com/ai-experts-call-on-amazon-not-to-sell-rekognition-software-to-police-2019-4</p><p>174</p><p>But there is much heavy lifting to be done to get to AGI.Even with the great</p><p>strides with deep learning, it still generally requires large amounts of data and</p><p>significant computing power.</p><p>AGI will instead need new approaches, such as the</p><p>traditional computer logic or preprogramming—that is, the use</p><p>of approaches like If-Then-Else statements.</p><p>Artificial Intelligence Basics</p><p>http://psych.fullerton.edu/mbirnbaum/psych101/Eliza.htm</p><p>http://psych.fullerton.edu/mbirnbaum/psych101/Eliza.htm</p><p>10</p><p>Next, there was Frank Rosenblatt, who believed that AI needed to use</p><p>systems similar to the brain like neural networks (this field was also known as</p><p>connectionism). But instead of calling the inner workings neurons, he referred</p><p>to them as perceptrons. A system would be able to learn as it ingested data</p><p>over time.</p><p>In 1957, Rosenblatt created the first computer program for this, called the</p><p>Mark 1 Perceptron. It included cameras to help to differentiate between two</p><p>images (they had 20 × 20 pixels). The Mark 1 Perceptron would use data that</p><p>had random weightings and then go through the following process:</p><p>1. Take in an input and come up with the perceptron output.</p><p>2. If there is not a match, then</p><p>a. If the output should have been 0 but was 1, then the weight for</p><p>1 will be decreased.</p><p>b. If the output should have been 1 but was 0, then the</p><p>weight for 1 will be increased.</p><p>3. Repeat steps #1 and #2 until the results are accurate.</p><p>This was definitely pathbreaking for AI. The New York Times even had a</p><p>write-up for Rosenblatt, extolling “The Navy revealed the embryo of an</p><p>electronic computer today that it expects will be able to walk, talk, see, write,</p><p>reproduce itself and be conscious of its existence.”7</p><p>But there were still nagging issues with the perceptron. One was that the</p><p>neural network had only one layer (primarily because of the lack of computation</p><p>power at the time). Next, brain research was still in the nascent stages and</p><p>did not offer much in terms of understanding cognitive ability.</p><p>Minsky would co-write a book, along with Seymour Papert, called Perceptrons</p><p>(1969). The authors were relentless in attacking Rosenblatt’s approach, and it</p><p>quickly faded away. Note that in the early 1950s Minsky developed a crude</p><p>neural net machine, such as by using hundreds of vacuum tubes and spare</p><p>parts from a B-24 bomber. But he saw that the technology was nowhere at a</p><p>point to be workable.</p><p>Rosenblatt tried to fight back, but it was too late. The AI community quickly</p><p>turned sour on neural networks. Rosenblatt would then die a couple years</p><p>later in a boating accident. He was 43 years old.</p><p>Yet by the 1980s, his ideas would be revived—which would lead to a revolution</p><p>in AI, primarily with the development of deep learning.</p><p>7 www.nytimes.com/1958/07/08/archives/new-navy-device-learns-by-doing-</p><p>psychologist-shows-embryo-of.html</p><p>Chapter 1 | AI Foundations</p><p>http://www.nytimes.com/1958/07/08/archives/new-navy-device-learns-by-doing-psychologist-shows-embryo-of.html</p><p>http://www.nytimes.com/1958/07/08/archives/new-navy-device-learns-by-doing-psychologist-shows-embryo-of.html</p><p>11</p><p>For the most part, the Golden Age of AI was freewheeling and exciting.</p><p>Some of the brightest academics in the world were trying to create machines</p><p>that could truly think. But the optimism often went to the extremes. In</p><p>1965, Simon said that within 20 years, a machine could do anything a human</p><p>could. Then in 1970, in an interview with Life magazine, he said this would</p><p>happen in only 3–8 years (by the way, he was an advisor on the 2001: A</p><p>Space Odyssey movie).</p><p>Unfortunately, the next phase of AI would be much darker. There were more</p><p>academics who were becoming skeptical. Perhaps the most vocal was Hubert</p><p>Dreyfus, a philosopher. In books such as What Computers Still Can’t Do: A</p><p>Critique of Artif icial Reason,8 he set forth his ideas that computers were not</p><p>similar to the human brain and that AI would woefully fall short of the lofty</p><p>expectations.</p><p>AI Winter</p><p>During the early 1970s, the enthusiasm for AI started to wane. It would</p><p>become known as the “AI winter,” which would last through 1980 or so (the</p><p>term came from “nuclear winter,” an extinction event where the sun is blocked</p><p>and temperatures plunge across the world).</p><p>Even though there were many strides made with AI, they still were mostly</p><p>academic and involved in controlled environments. At the time, the computer</p><p>systems were still limited. For example, a DEC PDP-11/45—which was</p><p>common for AI research—had the ability to expand its memory to only 128K.</p><p>The Lisp language also was not ideal for computer systems. Rather, in the</p><p>corporate world, the focus was primarily on FORTRAN.</p><p>Next, there were still many complex aspects when understanding intelligence</p><p>and reasoning. Just one is disambiguation. This is the situation when a word</p><p>has more than one meaning. This adds to the difficulty for an AI program since</p><p>it will also need to understand the context.</p><p>Finally, the economic environment in the 1970s was far from robust. There</p><p>were persistent inflation, slow growth, and supply disruptions, such as with</p><p>the oil crisis.</p><p>Given all this, it should be no surprise that the US government was getting</p><p>more stringent with funding. After all, for a Pentagon planner, how useful is a</p><p>program that can play chess, solve a theorem, or recognize some basic images?</p><p>Not much, unfortunately.</p><p>8 MIT Press, 1972.</p><p>Artificial Intelligence Basics</p><p>12</p><p>A notable case is the Speech Understanding Research program at Carnegie</p><p>Mellon University. For the Defense Advanced Research Projects Agency</p><p>(DARPA), it thought this speech recognition system could be used for fighter</p><p>pilots to make voice commands. But it proved to be unworkable. One of the</p><p>programs, which was called Harpy, could understand 1,011 words—which is</p><p>what a typical 3-year-old knows.</p><p>The officials at DARPA actually thought that it had been hoodwinked and</p><p>eliminated the $3 million annual budget for the program.</p><p>But the biggest hit to AI came via a report—which came out in 1973—from</p><p>Professor Sir James Lighthill. Funded by the UK Parliament, it was a full-on</p><p>repudiation of the “grandiose objectives” of strong AI. A major issue he noted</p><p>was “combinatorial explosion,” which was the problem where the models got</p><p>too complicated and were difficult to adjust.</p><p>The report concluded: “In no part of the field have the discoveries made so</p><p>far produced the major impact that was then promised.”9 He was so pessimistic</p><p>that he did not believe computers would be able to recognize images or beat</p><p>a chess grand master.</p><p>The report also led to a public debate that was televised on the BCC (you can</p><p>find the videos on YouTube). It was Lighthill against Donald Michie, Richard</p><p>Gregory, and John McCarthy.</p><p>Even though Lighthill had valid points—and evaluated large amounts of</p><p>research—he did not see the power of weak AI. But it did not seem to matter</p><p>as the winter took hold.</p><p>Things got so bad that many researchers changed their career paths. And as</p><p>for those who still studied AI, they often referred to their work with other</p><p>terms—like machine learning, pattern recognition, and informatics!</p><p>The Rise andFall ofExpert Systems</p><p>Even during the AI winter, there continued to be major innovations. One was</p><p>backpropagation, which is essential for assigning weights for neural networks.</p><p>Then there was the development of the recurrent neural network (RNN).</p><p>This allows for connections to move through the input and output layers.</p><p>But in the 1980s and 1990s, there also was the emergence of expert systems.</p><p>A key driver was the explosive growth of PCs and minicomputers.</p><p>9 The 1973 “Artificial Intelligence: A General Survey” by Professor Sir James Lighthill of</p><p>Cambridge University, www.bbc.com/timelines/zq376fr.</p><p>Chapter 1 | AI Foundations</p><p>http://www.bbc.com/timelines/zq376fr</p><p>13</p><p>Expert systems were based on the concepts of Minsky’s symbolic logic,</p><p>involving complex pathways. These were often developed by domain experts</p><p>in particular fields like medicine, finance, and auto manufacturing.</p><p>Figure 1-2 shows the key parts of an expert system.</p><p>While there are expert systems that go back to the mid-1960s, they did not</p><p>gain commercial use until the 1980s.</p><p>An example was XCON (eXpert</p><p>CONfigurer), which John McDermott developed at Carnegie Mellon</p><p>University. The system allowed for optimizing the selection of computer</p><p>components and initially had about 2,500 rules. Think of it as the first</p><p>recommendation engine. From the launch in 1980, it turned out to be a big</p><p>cost saver for DEC for its line of VAX computers (about $40 million by 1986).</p><p>When companies saw the success of XCON, there was a boom in expert</p><p>systems—turning into a billion-dollar industry. The Japanese government also</p><p>saw the opportunity and invested hundreds of millions to bolster its home</p><p>market. However, the results were mostly a disappointment. Much of the</p><p>innovation was in the United States.</p><p>Consider that IBM used an expert system for its Deep Blue computer. In</p><p>1996, it would beat grand chess master Garry Kasparov, in one of six matches.</p><p>Deep Blue, which IBM had been developing since 1985, processed 200 million</p><p>positions per second.</p><p>But there were issues with expert systems. They were often narrow and</p><p>difficult to apply across other categories. Furthermore, as the expert systems</p><p>got larger, it became more challenging to manage them and feed data. The</p><p>Figure 1-2. Key parts of an expert system</p><p>Artificial Intelligence Basics</p><p>14</p><p>result was that there were more errors in the outcomes. Next, testing the</p><p>systems often proved to be a complex process. Let’s face it, there were times</p><p>when the experts would disagree on fundamental matters. Finally, expert</p><p>systems did not learn over time. Instead, there had to be constant updates to</p><p>the underlying logic models, which added greatly to the costs and complexities.</p><p>By the late 1980s, expert systems started to lose favor in the business world,</p><p>and many startups merged or went bust. Actually, this helped cause another</p><p>AI winter, which would last until about 1993. PCs were rapidly eating into</p><p>higher-end hardware markets, which meant a steep reduction in Lisp-based</p><p>machines.</p><p>Government funding for AI, such as from DARPA, also dried up. Then again,</p><p>the Cold War was rapidly coming to a quiet end with the fall of the Soviet</p><p>Union.</p><p>Neural Networks andDeep Learning</p><p>As a teen in the 1950s, Geoffrey Hinton wanted to be a professor and to</p><p>study AI. He came from a family of noted academics (his great-great-</p><p>grandfather was George Boole). His mom would often say, “Be an academic</p><p>or be a failure.”10</p><p>Even during the first AI winter, Hinton was passionate about AI and was</p><p>convinced that Rosenblatt’s neural network approach was the right path. So</p><p>in 1972, he received his PhD on the topic from the University of Edinburgh.</p><p>But during this period, many people thought that Hinton was wasting his time</p><p>and talents. AI was essentially considered a fringe area. It wasn’t even thought</p><p>of as a science.</p><p>But this only encouraged Hinton more. He relished his position as an outsider</p><p>and knew that his ideas would win out in the end.</p><p>Hinton realized that the biggest hindrance to AI was computer power. But he</p><p>also saw that time was on his side. Moore’s Law predicted that the number of</p><p>components on a chip would double about every 18 months.</p><p>In the meantime, Hinton worked tirelessly on developing the core theories of</p><p>neural networks—something that eventually became known as deep learning.</p><p>In 1986, he wrote—along with David Rumelhart and Ronald J.Williams—a</p><p>pathbreaking paper, called “Learning Representations by Back-propagating</p><p>Errors.” It set forth key processes for using backpropagation in neural</p><p>networks. The result was that there would be significant improvement in</p><p>accuracy, such as with predictions and visual recognition.</p><p>10 https://torontolife.com/tech/ai-superstars-google-facebook-</p><p>apple-studied-guy/</p><p>Chapter 1 | AI Foundations</p><p>https://torontolife.com/tech/ai-superstars-google-facebook-apple-studied-guy/</p><p>https://torontolife.com/tech/ai-superstars-google-facebook-apple-studied-guy/</p><p>15</p><p>Of course, this did not happen in isolation. Hinton’s pioneering work was</p><p>based on the achievements of other researchers who also were believers in</p><p>neural networks. And his own research spurred a flurry of other major</p><p>achievements:</p><p>• 1980: Kunihiko Fukushima created Neocognitron, which</p><p>was a system to recognize patterns that became the basis</p><p>of convolutional neural networks. These were based on</p><p>the visual cortex of animals.</p><p>• 1982: John Hopfield developed “Hopfield Networks.”</p><p>This was essentially a recurrent neural network.</p><p>• 1989: Yann LeCun merged convolutional networks with</p><p>backpropagation. This approach would find use cases</p><p>with analyzing handwritten checks.</p><p>• 1989: Christopher Watkins’ PhD thesis, “Learning from</p><p>Delayed Rewards,” described Q-Learning. This was a</p><p>major advance in helping with reinforcement learning.</p><p>• 1998: Yann LeCun published “Gradient-Based Learning</p><p>Applied to Document Recognition,” which used descent</p><p>algorithms to improve neural networks.</p><p>Technological Drivers ofModern AI</p><p>Besides advances in new conceptual approaches, theories, and models, AI had</p><p>some other important drivers. Here’s a look at the main ones:</p><p>• Explosive Growth in Datasets: The internet has been a</p><p>major factor for AI because it has allowed for the creation</p><p>of massive datasets. In the next chapter, we’ll take a look</p><p>at how data has transformed this technology.</p><p>• Infrastructure: Perhaps the most consequential company</p><p>for AI during the past 15 years or so has been Google. To</p><p>keep up with the indexing of the Web—which was</p><p>growing at a staggering rate—the company had to come</p><p>up with creative approaches to build scalable systems.</p><p>The result has been innovation in commodity server</p><p>clusters, virtualization, and open source software. Google</p><p>was also one of the early adopters of deep learning, with</p><p>the launch of the “Google Brain” project in 2011. Oh,</p><p>and a few years later the company hired Hinton.</p><p>Artificial Intelligence Basics</p><p>16</p><p>• GPUs (Graphics Processing Units): This chip technology,</p><p>which was pioneered by NVIDIA, was originally for high-</p><p>speed graphics in games. But the architecture of GPUs</p><p>would eventually be ideal for AI as well. Note that most</p><p>deep learning research is done with these chips. The</p><p>reason is that—with parallel processing—the speed is</p><p>multiples higher than traditional CPUs. This means that</p><p>computing a model may take a day or two vs. weeks or</p><p>even months.</p><p>All these factors reinforced themselves—adding fuel to the growth</p><p>of AI.What’s more, these factors are likely to remain vibrant for many</p><p>years to come.</p><p>Structure ofAI</p><p>In this chapter, we’ve covered many concepts. Now it can be tough to understand</p><p>the organization of AI.For instance, it is common to see terms like machine</p><p>learning and deep learning get confused. But it is essential to understand the</p><p>distinctions, which we will cover in detail in the rest of this book.</p><p>But on a high-level view of things, Figure1-3 shows how the main elements of</p><p>AI relate to each other. At the top is AI, which covers a wide variety of</p><p>theories and technologies. You can then break this down into two main</p><p>categories: machine learning and deep learning.</p><p>Figure 1-3. This is a high-level look at the main components of the AI world</p><p>Chapter 1 | AI Foundations</p><p>17</p><p>Conclusion</p><p>There’s nothing new that AI is a buzzword today. The term has seen various</p><p>stomach-churning boom-bust cycles.</p><p>Maybe it will once again go out of favor? Perhaps. But this time around, there</p><p>are true innovations with AI that are transforming businesses. Mega tech</p><p>companies like Google, Microsoft, and Facebook consider the category to be</p><p>a major priority. All in all, it seems like a good bet that AI will continue to</p><p>grow and change our world.</p><p>Key Takeaways</p><p>• Technology often takes longer to evolve than originally</p><p>understood.</p><p>• AI is not just about computer science and mathematics.</p><p>There have been key contributions from fields like</p><p>economics, neuroscience, psychology, linguistics,</p><p>electrical engineering, mathematics, and philosophy.</p><p>• There are two main types</p><p>of AI: weak and strong. Strong</p><p>is where machines become self-aware, whereas weak is</p><p>for systems that focus on specific tasks. Currently, AI is</p><p>at the weak stage.</p><p>• The Turing Test is a common way to test if a machine can</p><p>think. It is based on whether someone really thinks a</p><p>system is intelligent.</p><p>• Some of the key drivers for AI include new theories from</p><p>researchers like Hinton, the explosive growth in data,</p><p>new technology infrastructure, and GPUs.</p><p>Artificial Intelligence Basics</p><p>© Tom Taulli 2019</p><p>T. Taulli, Artif icial Intelligence Basics,</p><p>https://doi.org/10.1007/978-1-4842-5028-0_2</p><p>C H A P T E R</p><p>2</p><p>Data</p><p>The Fuel for AI</p><p>Pinterest is one of the hottest startups in Silicon Valley, allowing users to pin</p><p>their favorite items to create engaging boards. The site has 250 million MAUs</p><p>(monthly active users) and posted $756 million in revenue in 2018.1</p><p>A popular activity for Pinterest is to plan for weddings. The bride-to-be will</p><p>have pins for gowns, venues, honeymoon spots, cakes, invitations, and so on.</p><p>This also means that Pinterest has the advantage of collecting huge amounts</p><p>of valuable data. Part of this helps provide for targeted ads. Yet there are also</p><p>opportunities for email campaigns. In one case, Pinterest sent one that said:</p><p>You’re getting married! And because we love wedding planning—espe-</p><p>cially all the lovely stationery—we invite you to browse our best boards</p><p>curated by graphic designers, photographers and fellow brides-to-be, all</p><p>Pinners with a keen eye and marriage on the mind.2</p><p>The problem: Plenty of the recipients of the email were already married or</p><p>not expecting to marry anytime soon.</p><p>1 www.cnbc.com/2019/03/22/pinterest-releases-s-1-for-ipo.html</p><p>2 www.businessinsider.com/pinterest-accidental-marriage-emails-2014-9</p><p>http://www.cnbc.com/2019/03/22/pinterest-releases-s-1-for-ipo.html</p><p>http://www.businessinsider.com/pinterest-accidental-marriage-emails-2014-9</p><p>20</p><p>Pinterest did act quickly and put out this apology:</p><p>Every week, we email collections of category-specific pins and boards to</p><p>pinners we hope will be interested in them. Unfortunately, one of these</p><p>recent emails suggested that pinners were actually getting married,</p><p>rather than just potentially interested in wedding-related content. We’re</p><p>sorry we came off like an overbearing mother who is always asking when</p><p>you’ll find a nice boy or girl.</p><p>It’s an important lesson. Even some of the most tech-savvy companies</p><p>blow it.</p><p>For example, there are some cases where the data may be spot-on but the</p><p>outcome could still be an epic failure. Consider the case with Target. The</p><p>company leveraged its massive data to send personalized offers to expectant</p><p>mothers. This was based on those customers who made certain types of</p><p>purchases, such as for unscented lotions. Target’s system would create a</p><p>pregnancy score that even provided estimates of due dates.</p><p>Well, the father of one of the customers saw the email and was furious, saying</p><p>his daughter was not pregnant.3</p><p>But she was—and yes, she had been hiding this fact from her father.</p><p>There’s no doubt that data is extremely powerful and critical for AI.But you</p><p>need to be thoughtful and understand the risks. In this chapter, we’ll take a</p><p>look at some of the things you need to know.</p><p>Data Basics</p><p>It’s good to have an understanding of the jargon of data.</p><p>First of all, a bit (which is short for “binary digit”) is the smallest form of data</p><p>in a computer. Think of it as the atom. A bit can either be 0 or 1, which is</p><p>binary. It is also generally used to measure the amount of data that is being</p><p>transferred (say within a network or the Internet).</p><p>A byte, on the other hand, is mostly for storage. Of course, the numbers of</p><p>bytes can get large very fast. Let’s see how in Table2-1.</p><p>3 www.businessinsider.com/the-incredible-story-of-how-target-</p><p>exposed-a-teen-girls-pregnancy-2012-2</p><p>Chapter 2 | Data</p><p>http://www.businessinsider.com/the-incredible-story-of-how-target-exposed-a-teen-girls-pregnancy-2012-2</p><p>http://www.businessinsider.com/the-incredible-story-of-how-target-exposed-a-teen-girls-pregnancy-2012-2</p><p>21</p><p>Data can also come from many different sources. Here is just a sampling:</p><p>• Web/social (Facebook, Twitter, Instagram, YouTube)</p><p>• Biometric data (fitness trackers, genetics tests)</p><p>• Point of sale systems (from brick-and-mortar stores and</p><p>e-commerce sites)</p><p>• Internet of Things or IoT (ID tags and smart devices)</p><p>• Cloud systems (business applications like Salesforce.com)</p><p>• Corporate databases and spreadsheets</p><p>Types ofData</p><p>There are four ways to organize data. First, there is structured data, which is</p><p>usually stored in a relational database or spreadsheet. Some examples include</p><p>the following:</p><p>• Financial information</p><p>• Social Security numbers</p><p>• Addresses</p><p>• Product information</p><p>• Point of sale data</p><p>• Phone numbers</p><p>For the most part, structured data is easier to work with. This data often</p><p>comes from CRM (Customer Relationship Management) and ERP (Enterprise</p><p>Resource Planning) systems—and usually has lower volumes. It also tends to</p><p>Table 2-1. Types of data levels</p><p>Unit Value Use Case</p><p>Megabyte 1,000 kilobytes A small book</p><p>Gigabyte 1,000 megabytes About 230 songs</p><p>Terabyte 1,000 gigabytes 500 hours of movies</p><p>Petabyte 1,000 terabytes Five years of the Earth Observing System (EOS)</p><p>Exabyte 1,000 petabytes The entire Library of Congress 3,000 times over</p><p>Zettabyte 1,000 exabytes 36,000 years of HD-TV video</p><p>Yottabytes 1,000 zettabytes This would require a data center the size of Delaware</p><p>and Rhode Island combined</p><p>Artificial Intelligence Basics</p><p>22</p><p>be more straightforward, say in terms of analysis. There are various BI</p><p>(Business Intelligence) programs that can help derive insights from structured</p><p>data. However, this type of data accounts for about 20% of an AI project.</p><p>The majority will instead come from unstructured data, which is information</p><p>that has no predefined formatting. You’ll have to do this yourself, which can</p><p>be tedious and time consuming. But there are tools like next-generation</p><p>databases—such as those based on NoSQL—that can help with the process.</p><p>AI systems are also effective in terms of managing and structuring the data, as</p><p>the algorithms can recognize patterns.</p><p>Here are examples of unstructured data:</p><p>• Images</p><p>• Videos</p><p>• Audio files</p><p>• Text files</p><p>• Social network information like tweets and posts</p><p>• Satellite images</p><p>Now there is some data that is a hybrid of structured and unstructured</p><p>sources—called semi-structured data. The information has some internal tags</p><p>that help with categorization.</p><p>Examples of semi-structured data include XML (Extensible Markup Language),</p><p>which is based on various rules to identify elements of a document, and JSON</p><p>( JavaScript Object Notation), which is a way to transfer information on the</p><p>Web through APIs (Application Programming Interfaces).</p><p>But semi-structured data represents only about 5% to 10% of all data.</p><p>Finally, there is time-series data, which can be both for structured, unstructured,</p><p>and semi-structured data. This type of information is for interactions, say for</p><p>tracking the “customer journey.” This would be collecting information when a</p><p>user goes to the web site, uses an app, or even walks into a store.</p><p>Yet this kind of data is often messy and difficult to understand. Part of this is</p><p>due to understanding the intent of the users, which can vary widely. There is</p><p>also huge volumes of interactional data, which can involve trillions of data</p><p>points. Oh, and the metrics for success may not be clear. Why is a user doing</p><p>something on the site?</p><p>But AI is likely to be critical for such issues. Although, for the most part, the</p><p>analysis of time-series data is still in the early stages.</p><p>Chapter 2 | Data</p><p>23</p><p>Big Data</p><p>With the ubiquity of Internet access, mobile devices, and wearables, there has</p><p>been the unleashing of a torrent of data. Every second, Google processes</p><p>over 40,000 searches or 3.5 billion a day. On a minute-by-minute basis,</p><p>Snapchat users share 527,760</p><p>photos, and YouTube users watch more than</p><p>4.1 million videos. Then there are the old-fashioned systems, like emails, that</p><p>continue to see significant growth. Every minute, there are 156 million</p><p>messages sent.4</p><p>But there is something else to consider: Companies and machines also</p><p>generate huge sums of data. According to research from Statista, the number</p><p>of sensors will reach 12.86 billion by 2020.5</p><p>In light of all this, it seems like a good bet that the volumes of data will</p><p>continue to increase at a rapid clip. In a report from International Data</p><p>Corporation (IDC) called “Data Age 2025,” the amount of data created is</p><p>expected to hit a staggering 163 zettabytes by 2025.6 This is about ten times</p><p>the amount in 2017.</p><p>To deal with all this, there has emerged a category of technology called Big</p><p>Data. This is how Oracle explains the importance of this trend:</p><p>Today, big data has become capital. Think of some of the world’s biggest</p><p>tech companies. A large part of the value they offer comes from their</p><p>data, which they’re constantly analyzing to produce more efficiency and</p><p>develop new products.7</p><p>So yes, Big Data will remain a critical part of many AI projects.</p><p>Then what exactly is Big Data? What’s a good definition? Actually, there isn’t</p><p>one, even though there are many companies that focus on this market! But</p><p>Big Data does have the following characteristics, which are called the three Vs</p><p>(Gartner analyst Doug Laney came up with this structure back in 20018):</p><p>volume, variety, and velocity.</p><p>4 www.forbes.com/sites/bernardmarr/2018/05/21/how-much-data-do-we-</p><p>create-every-day-the-mind-blowing-stats-everyone-should-</p><p>read/#788c13c660ba</p><p>5 w w w . f o r b e s . c o m / s i t e s / l o u i s c o l u m b u s / 2 0 1 8 / 0 6 / 0 6 / 1 0 -</p><p>charts-that-will-challenge-your-perspective-of-iots-growth/#4e9fac23ecce</p><p>6 h t t p s : / / b l o g . s e a g a t e . c o m / b u s i n e s s / e n o r m o u s - g r o w t h - i n -</p><p>data-is-coming-how-to-prepare-for-it-and-prosper-from-it/</p><p>7 www.oracle.com/big-data/guide/what-is-big-data.html</p><p>8 https://blogs.gartner.com/doug-laney/files/2012/01/ad949-3D-Data-</p><p>Management-Controlling-Data-Volume-Velocity-and-Variety.pdf</p><p>Artificial Intelligence Basics</p><p>http://www.forbes.com/sites/bernardmarr/2018/05/21/how-much-data-do-we-create-every-day-the-mind-blowing-stats-everyone-should-read/#788c13c660ba</p><p>http://www.forbes.com/sites/bernardmarr/2018/05/21/how-much-data-do-we-create-every-day-the-mind-blowing-stats-everyone-should-read/#788c13c660ba</p><p>http://www.forbes.com/sites/bernardmarr/2018/05/21/how-much-data-do-we-create-every-day-the-mind-blowing-stats-everyone-should-read/#788c13c660ba</p><p>http://www.forbes.com/sites/louiscolumbus/2018/06/06/10-charts-that-will-challenge-your-perspective-of-iots-growth/#4e9fac23ecce</p><p>http://www.forbes.com/sites/louiscolumbus/2018/06/06/10-charts-that-will-challenge-your-perspective-of-iots-growth/#4e9fac23ecce</p><p>https://blog.seagate.com/business/enormous-growth-in-data-is-coming-how-to-prepare-for-it-and-prosper-from-it/</p><p>https://blog.seagate.com/business/enormous-growth-in-data-is-coming-how-to-prepare-for-it-and-prosper-from-it/</p><p>http://www.oracle.com/big-data/guide/what-is-big-data.html</p><p>https://blogs.gartner.com/doug-laney/files/2012/01/ad949-3D-Data-Management-Controlling-Data-Volume-Velocity-and-Variety.pdf</p><p>https://blogs.gartner.com/doug-laney/files/2012/01/ad949-3D-Data-Management-Controlling-Data-Volume-Velocity-and-Variety.pdf</p><p>24</p><p>Volume</p><p>This is the scale of the data, which is often unstructured. There is no hard-</p><p>and-fast rule on a threshold, but it is usually tens of terabytes.</p><p>Volume is often a major challenge when it comes to Big Data. But cloud</p><p>computing and next-generation databases have been a big help—in terms of</p><p>capacity and lower costs.</p><p>Variety</p><p>This describes the diversity of the data, say a combination of structured, semi-</p><p>structured, and unstructured data (explained above). It also shows the</p><p>different sources of the data and uses. No doubt, the high growth in</p><p>unstructured data has been a key to the variety of Big Data.</p><p>Managing this can quickly become a major challenge. Yet machine learning is</p><p>often something that can help streamline the process.</p><p>Velocity</p><p>This shows the speed at which data is being created. As seen earlier in this</p><p>chapter, services like YouTube and Snapchat have extreme levels of velocity</p><p>(this is often referred to as a “firehouse” of data). This requires heavy</p><p>investments in next-generation technologies and data centers. The data is</p><p>also often processed in memory not with disk-based systems.</p><p>Because of these issues, velocity is often considered the most difficult when</p><p>it comes to the three Vs. Let’s face it, in today’s digital world, people want</p><p>their data as fast as possible. If it is too slow, people will get frustrated and go</p><p>somewhere else.</p><p>Over the years, though, as Big Data has evolved, there have been more Vs</p><p>added. Currently, there are over ten.</p><p>But here are some of the common ones:</p><p>• Veracity: This is about data that is deemed accurate. In</p><p>this chapter, we’ll look at some of the techniques to</p><p>evaluate veracity.</p><p>• Value: This shows the usefulness of the data. Often this is</p><p>about having a trusted source.</p><p>Chapter 2 | Data</p><p>25</p><p>• Variability: This means that data will usually change over</p><p>time. For example, this is the case with social media</p><p>content that can morph based on overall sentiment</p><p>regarding new developments and breaking news.</p><p>• Visualization: This is using visuals—like graphs—to better</p><p>understand the data.</p><p>As you can see, managing Big Data has many moving parts, which leads to</p><p>complexity. This helps to explain why many companies still use only a tiny</p><p>fraction of their data.</p><p>Databases andOther Tools</p><p>There are a myriad of tools that help with data. At the core of this is the</p><p>database. As should be no surprise, there has been an evolution of this critical</p><p>technology over the decades. But even older technologies like relational</p><p>databases are still very much in use today. When it comes to mission-critical</p><p>data, companies are reluctant to make changes—even if there are clear</p><p>benefits.</p><p>To understand this market, let’s rewind back to 1970, when IBM computer</p><p>scientist Edgar Codd published “A Relational Model of Data for Large Shared</p><p>Data Banks.” It was pathbreaking as it introduced the structure of relational</p><p>databases. Up until this point, databases were fairly complex and rigid—</p><p>structured as hierarchies. This made it time consuming to search and find</p><p>relationships in the data.</p><p>As for Codd’s relational database approach, it was built for more modern</p><p>machines. The SQL script language was easy to use allowing for CRUD</p><p>(Create, Read, Update, and Delete) operations. Tables also had connections</p><p>with primary and foreign keys, which made important connections like the</p><p>following:</p><p>• One-to-One: One row in a table is linked to only one row</p><p>in another table. Example: A driver’s license number,</p><p>which is unique, is associated with one employee.</p><p>• One-to-Many: This is where one row in a table is linked to</p><p>other tables. Example: A customer has multiple purchase</p><p>orders.</p><p>• Many-to-Many: Rows from one table are associated with</p><p>rows of another. Example: Various reports have various</p><p>authors.</p><p>With these types of structures, a relational database could streamline the</p><p>process of creating sophisticated reports. It truly was revolutionary.</p><p>Artificial Intelligence Basics</p><p>26</p><p>But despite the advantages, IBM was not interested in the technology and</p><p>continued to focus on its proprietary systems. The company thought that the</p><p>relational databases were too slow and brittle for enterprise customers.</p><p>But there was someone who had a different opinion on the matter: Larry</p><p>Ellison. He read Codd’s paper and knew it would be a game changer. To prove</p><p>this, he would go on to co-found Oracle in 1977 with a focus on building</p><p>relational databases—which would quickly become a massive market. Codd’s</p><p>paper was essentially a product roadmap for his entrepreneurial</p>