Thursday, June 23, 2011

The Soul of a Not-Quite-Thinking Machine

I just finished a wonderful book, Final Jeopardy: Man vs. Machine and the Quest to Know Everything by Stephen Baker. Baker had inside access to the IBM team that developed Watson, the amazing question-answering computer system that bested two human champions on Jeopardy early this year. The book quickly jumped the queue in my Amazon Kindle app after I read the sample chapter. I've always been fascinated by artificial intelligence, and I had already read a bit about Watson in a New York Times series last fall and through a number of web sources, but I somehow missed this book. It came out in February, shortly after the Watson shows had aired.

While Baker discusses some of the technology behind Watson, it's really the story of a group of talented humans who took on a great challenge. IBM had already been working on natural language-based question answering systems, but when the "Blue J" project started in 2006, they were nowhere near the level of any human player, let alone champions like Ken Jennings and Brad Rutter. The book reminded me most of Tracy Kidder's The Soul of a New Machine which I read sometime in the 1980's.

As interested as I was in Watson, I missed the original Jeopardy broadcast due to travel, and I had seen only a few brief excerpts. While reading the book, I decided I should really watch the whole three-show tournament, which I was able to do (without commercials) in about an hour thanks to YouTube. I was also curious to know more about the technology of Watson than I had learned from Baker and the other general things I had read online. I found a great paper from AI Magazine's fall 2010 issue, "An Overview of the DeepQA Project" (PDF), written by the IBM Watson project leader David Ferrucci and 11 members of his team. It is really incredible to learn about what goes on behind the scenes to allow Watson to answer virtually any Jeopardy question in 3-5 seconds with such incredible precision and confidence. "Confidence" is a big part of why this "QA" technology is different from something like Google's search technology which usually gets you into the neighborhood of an answer, but depends on having a human in the loop to determine the "right" answer. Watson has no human in the loop, so it needs to "know what it knows" to decide whether to "buzz in" on a Jeopardy question. It puts its "hypotheses" through extensive evaluation to determine its confidence in each one as a possible answer.

Of course IBM does not claim that Watson is "intelligent" or that it "thinks" anything like we do, but you could say that human brains are also massively parallel computing systems with thousands of inter-communicating subsystems. Watson's subsystems are in server racks. Our subsystems are squishy. Diversity! So let's give a warm welcome to our new computer overlord cousin, shall we?

No comments: