I wrote this book to explain the concepts of Artificial Intelligence (AI) in the form of deep learning to people in the business community who are not data scientists. You won’t find equations here. This book is a no-math book.
As a preliminary matter, my subject matter expertise is law and litigation. I will often use the legal context because it’s familiar to me, and I encourage you to take terms like “Commercial Tort Litigation,” “litigation hold notices” and any other legal terms or phrases in stride. Please don’t concern yourself with a lack of definition or further details. Just think of any of the “inside baseball” terms that I may mention as a placeholder for some term or phrase that may be familiar to you because of your expertise. Legal is my sand box.
My aim and hope is that this book will be important to you if you’re a life-long learner and you want to use your subject-matter expertise to identify innovative applications of AI for your industry and in your areas of expertise.
My other goal is to help you appreciate what the AI revolution is all about, and why the phrase “deep learning” is what people are really talking about when they use the term AI.
I’m also writing this book because I think of myself as a lifelong learner. I learned enough of the conceptual framework for deep learning to come up with more than a handful of business applications and to obtain eight patents for them. If I can climb on top of deep learning, so can you.
But aside from remotely auditing a computer science course at Stanford (which was then called CS 224d and is CS 224n now), I’m not a trained data scientist.
However, long ago, I earned a B.S. degree in Engineering Systems at UCLA, and then an M.S. degree in Environmental Engineering Science at the California Institute of Technology. But after graduating from Caltech, I went to the Gould School of Law at the University of Southern California. After earning a J.D. degree, I was a litigation specialist in California for both plaintiffs and defendants in both state and federal courts from 1975 to 2014.
Thus, while I’m not unfamiliar with technology, I didn’t have professional experience with engineering or technology. But in 2015, I began my second childhood. That’s the step that allowed me to become an inventor of software systems that make use of deep learning.
Briefly, here’s what happened.
During much of 2015, I was writing a book with computer scientist W. H. (Bill) Inmon, the father of the data warehouse. The book, Preventing Litigation: An Early Warning System to Get Big Value Out of Big Data, was published in the Fall of 2015 by Business Expert Press, the same publisher of this book.
I stand by Preventing Litigation. It’s earned a perfect five-star rating from readers and it was endorsed by one of the world’s foremost thought leaders for the legal profession, Sir Richard Susskind. He allowed me to use this line on the back cover: “As a lawyer or client, if you prefer a fence at the top of a cliff to an ambulance at the bottom, this insightful book is essential reading.”
However, while I was editing Preventing Litigation, and before it was published, I saw an article that spoke to me. The news was that venture capitalists were starting to take AI seriously and had started to back an ever-growing list of startups. I mentioned AI and deep learning only in passing, on pages 169–170, as a possible alternative to the technology that Bill Inmon, my co-author, had described in Chapters 10–14. After Preventing Litigation was published, I diverted from the technology Bill had laid out. I still had the same business application in mind, that is, a software system to help corporate counsel (i.e., in-house attorneys) see litigation risks in time to nip them in the bud. As attorney Dan Sanders put it recently in the context of risk management, the mission I had set up in Preventing Litigation was to create the software that would “protect tomorrow, today.”1
When Preventing Litigation was published, I had been living in Sequim, Washington (on the Olympic Peninsula) for less than two years. In order to meet new people, I volunteered to be a mentor to the students at Sequim High School who were working after school hours on a robot that would compete in a national competition. I was fortunate in that one of the other mentors was Michael G. Becker. Mike had been a software engineer during a 40-year career that took him from IBM to Microsoft. I mentioned deep learning to Mike one day and he told me that he had previously come across “neural networks,” and so was interested. I knew where to find the data for training a deep learning model of specific types of litigation and the data we could use as a testbed.
As luck would have it, one of the startups in that article I mentioned on page 169 of Preventing Litigation was MetaMind. When I looked up the MetaMind website, I saw an open API (“application programming interface”) by which MetaMind was making its system available to third parties like us. If we had the appropriate training data, we could train a neural network. If we had the appropriate test data, we could evaluate the model. It took us about four months to create a rudimentary proof of concept.
During much of that time, my understanding was that, due to the U.S. Supreme Court decision in Alice Corp. v. CLS Bank International,2 the United States Patent and Trademark Office (USPTO) was going to turn down every patent application for a software system on the grounds that the system was simply a task that a computer could do, and so was patent ineligible.
But on June 27, 2016, the Federal Circuit Court of Appeals published its decision in Bascom Global v. AT&T Mobility.3 In that decision, the Federal Circuit expressed the notion that practitioners had misread the Alice decision. There was a “step two” in the Alice decision, and it was that if an application for a patent set out a software system with an “inventive concept,” a patent application for such a system might be granted. In his Bascom Global opinion, Judge Raymond Chen explained that the patent at issue had passed this “step two” test and explained why: “As is the case here,” he wrote, “an inventive concept can be found in the non-conventional and non-generic arrangement of known, conventional pieces.”4
The Bascom Global opinion was inspirational. I wrote a provisional application and, with the assistance of patent attorney Dan Cotman and his colleague, Obi Iloputaife, my provisional application was filed four days later, on July 1, 2016.
This first step was crucial. I knew that the “provisional” I had filed would become my Priority Date under the new “first to file” amendment to the patent laws, and I have been building on it ever since then. By the end of 2018, I have turned my inventive concepts into seven additional U.S. software system applications that have been accepted by the USPTO and issued as patents.
Later in this book, after I’ve set the stage for the concepts of deep learning as applied to text in the context of words, I’ll be using portions of the text from those patents as examples.
That said, I’ll reiterate my hope that this book will enable you to apply deep learning to your own areas of subject matter expertise. Please read this book with that thought in mind, so that you may apply deep learning to situations that are unfamiliar to me, but which are familiar to you.
And please keep in mind that the way I expect to enable you to learn—from examples—turns out to be the way deep learning enables computers to learn.
Note: For readers of the electronic version, the links in the Notes at the end of the book are live. Readers of the print version will have to copy them into a browser.
Notes
1 Dan Sanders, Risk and Business, in Corporate Counsel-Digital, at 42 (April 2019).
2 Alice Corp. Pty. Ltd. v. CLS Bank International, 573 U.S. 208, 134 S. Ct. 2347 (2014).
3 Bascom Global Internet Servs., Inc. v. AT & T Mobility, LLC, 827 F.3d 1341 (Fed. Cir. 2016).
4 Ibid. at p. 1350.