Everyone takes a different path to Artificial Intelligence. Sure, there are some common approaches from academia, research, engineering, and the like. But AI is a weird field that has had massive transformations in its history, and I don’t think I’m too unusual in that I’ve only started to really “do AI” professionally in the last year or so. Here’s a bit more about my path to AI.
My first exposure to the study of AI was in college, at the University of Virginia. Heading to school, I planned to do pre-med and become a doctor, but organic chemistry made it clear that wasn’t going to work out. I enjoyed science, so I decided on a biology major. Compared to the crunch of pre-med, this left time for other pursuits, like elective courses.
One elective I enrolled in, and very much enjoyed, was called “Philosophy of Artificial Intelligence”. It was offered by the philosophy department — so there was no math or programming or any of that. It was mostly reading, lectures, and papers. Wikipedia offers a pretty good overview of the AI philosophy we studied: things like physical symbol systems and neurobiological explanations of consciousness. I remember this as “chess vs brains” in a sort of Cliff’s Notes way. There wasn’t significant discussion of neural networks or deep learning; perhaps this was because I was taking this class in the early 1990s during the heart of the AI Winter!
After that class, I didn’t think about AI for many years. I learned about software engineering and became a professional coder (by way of AppleScript, but that’s another story). I built process automation scripts, web sites, and databases. I wound up getting very into databases and focused almost exclusively on SQL programming for several years. This became helpful later when I realized that data engineering is a powerful (and necessary) component of data science and analytics. It’s very difficult to create statistically models and visualizations on messy data!
When I first learned about data science in 2010, the term Artificial Intelligence was generally not mentioned. D.J. Patil’s famous sexiest job of the 21st century article does not mention AI at all. Neither does the What is Data Science? chapter of the O’Reilly Doing Data Science book. This chapter does mention “statistics” and “machine learning” — and those were the terms I learned about initially, in addition to databases, “big data”, and software engineering. So that’s what I did next.
Stats & Machine Learning
I learned statistics and machine learning in a variety of ways, all self-taught. In addition to a bunch more O’Reilly books, I mostly used Dataquest to practice interactively — I found this better than just watching videos. I learned core concepts with Kahn Academy and Wikipedia. I tried Andrew Ng’s Deep Learning Coursera course, but struggled with Octave and math notation, so I gave that up. I have never had to program anything in Octave or translate math notation to code in my career (yet) so I feel OK about this choice!
Most importantly, I applied data science concepts at work. “Learn by doing” is a real thing! I was fortunate to work with folks that had more stats & machine learning knowledge than me, in a variety of roles. This included peers, managers, and people that I helped hire and manage. I was usually able to handle most of the engineering “foundations” in these teams, so it was often a two-way learning process. I would lean into data modeling, databases, analytics, cloud infrastructure, and so forth, and they would lean into stats and machine learning. We would all advance.
Don’t Believe the Hype
Along the way, my views on AI evolved. For a significant period of time, I thought there was way too much hype around AI, especially deep learning. My reasoning was: if simpler methods are sufficient, and if the data is relatively small, let’s stick with the simpler methods. Consequently, most of my modeling work during this time included linear regression, clustering, logistic regression, decision trees, and similar traditional machine learning methods. Scikit-learn was my friend (and still is!). Moreover, modeling itself was only a small percentage of work in most projects: data collection and preparation often took up most of the time; it’s the base of the pyramid after all.
More recently, I worked in a team that was building a new data science practice in an established company. Tellingly, we decided early on that “leading with AI” was probably a bad idea. In addition to the perceived complexity of deep learning, we worried that collaborators and leaders would become impatient as we ramped up, and we would lose momentum. It was a startup-like “minimum viable product” strategy within a larger org. The thinking was: better to get “quick wins” with novel datasets and simple analytics than sink time into deep learning. This strategy generally worked; data science has become an established practice in this org. But it was like hitting the pause button for me personally on AI.
When I joined Google about 10 months ago it was like hitting play, then fast-forward. AI is, of course, a big focus for Google, and I have immersed myself in it. This has included doing the machine learning crash course, tracking the latest from Google Research and DeepMind, and learning how to use AutoML to predict retail stock-outs. I’ve also studied and internalized Google’s AI Principles to help guide my thinking.
I have started looking ahead more, and I’ve broadened my outlook. This reflects my role in OCTO, where I work with customers in various industries as well as Google’s product and engineering teams. We like to think about emerging themes (McKinsey’s “horizon two”).
Everywhere I look, there are horizon two opportunities for AI, from conversational approaches (think chatbots, but better), to vision applications (especially with documents, which are everywhere), to optimization strategies (supply chain and logistics come to mind), and beyond. More human-like AI technology will improve our lives in an amazing number of ways in the future.
I’m here for it! Let’s do AI.