I recently attended a presentation by John Boyer, a Chief Architect on IBM’s Watson project. The presentation was about the Watson technology that IBM has, and continues to, develop. This was the second Watson presentation I have attended, the first was at CASCON 2014 where an IBM VP gave a keynote lecture. I must admit, after both presentations I walked away feel rather disappointed.
I must preface the remainder of this post by stating that this represents my opinion based on the information I have gathered through these presentations and reading the news. I could be entirely wrong.
The feeling of disappointment stems from what I thought Watson was. Before attending these presentations I had watched Watson play jeopardy and read multiple news stories about this fantastic new technology, it felt like Watson was a huge step forward, and it was going to meet all of our analytics needs for all of eternity and make me coffee every morning. I expected these presentations to dazzle and amaze me with new and exciting machine learning techniques or to provide some deep philosophical insight regarding AI. After all, IBM must some “secret sauce” that makes this tech work, right?
In retrospect, it is not surprising that I was disappointed, I set myself up for it. Anytime someone attributes a “god-like” status to anything, technology or other, they will inevitability be disappointed in some respect.
As far as I can tell, Watson is an aggregation of the numerous machine learning and data mining techniques developed over the last 50 or so years. Of course, IBM has improved and tuned these approaches. However, in the end they are relying the same basic approaches that the rest of the machine learning and data mining community rely upon. These have been combined in to support exactly one task: answering questions.
The entire purpose of Watson is to answer the questions based on the (potentially vast) amount of information available. This approach has allowed IBM to focus their efforts, there are numerous tasks that Watson can’t, and never will be able to, do, simply because it wasn’t designed for them. Watson isn’t some fantastically brilliant AI using some revolutionary algorithm or technique, it is just a (really big) program, and an excellent example of what contemporary machine learning is capable of at scale.
Watson is, in my opinion, a feat of engineering. Given that it uses the current state-of-the-art machine learning techniques, and not some IBM secret sauce, then it is an exceptionally well executed version of a machine learning tool. A tremendous amount of engineering effort has gone into bringing together different machine learning approaches and unifying them into package that can be deployed in a breadth of application domains (health care, finance, business management, security etc…). The complexity of a single domain, if not managed properly, is often enough to cripple most software projects; Watson takes this complexity in stride and provides value to its customers.
I think the most interesting thing about Watson, isn’t Watson itself but its application to different domains. IBM has taken approaches that have existed primarily in academia and made them accessible, on a huge scale, to the world. The value of machine learning and data mining techniques increases as we increase the amount of data processed. Watson will allow us process tremendous amounts of information, and to really understand the impact of machine learning in different application domains.
Finally, I think Watson provides us with the opportunity to understand the limitations of current machine learning techniques. What problems with these approaches will manifest at scale? At what point is machine learning no longer viable, how will we know when we reach that point, and what will we do after that?
In conclusion, I am disappointed that Watson isn’t some out of this world piece of technology, I really expected to be blown away. Upon reflection, this appears to have been a naive opinion, life lesson: “there is no silver bullet”. However, Watson is an exceptional tool that will change how we approach problems in numerous domains, but it is just that: a tool.