Friday, June 25, 2010

Watson: the Amazing IBM Question Answering Computer

IBM built a computer that can play Jeopardy! against human players.

It can figure out the question, even unusual questions with puns in them, and often give the right answer, and do it fast enough to beat good human Jeopardy! players.

This is interesting because if it can do this, it can do most any data retrieval chore. For example, it can argue with you about your phone bill, or perhaps give you emergency medical advice, or tell you what product you should buy.

It is also interesting because of the huge programing accomplishment that it took to do this. The Watson coders correlated language with millions of words of scanned text, and built its own dictionary from that. The relationships are the program. Watson has read the Bible, encyclopedias, dictionaries and literature.

Watson has been around a while it was only answering 15% correct in 2006. It was beating human players in 2008, and can beat champion players today. It used the six seconds during while the question is read for a head start.

Watson is an implementation of DeepQA which runs on a BlueGene IBM computer. BlueGen computers are massively parallel supercomputers, with computers-on-a-chip at each node.  (See my post on BlueBrain which also runs on BlueGene computers.)

Watson is not aware in the sense that a human being feels aware. Its software cannot sense awareness or anything like that. It does not know why it plays Jeopardy! or why it wants to. In this sense it is not sentient.

On the other hand Douglas Hofsteadtler, author of Goedel, Escher, Bach, and Strange Loop argues that sentience is an illusion that our brain feels.

Similarly, many believe that the Mind is a giant interior dialog we carry on with ourselves. If this computer answers questions, could it answer its own questions? Probably not -- it is not made that way. The designers would need to start over. This question answering machine could not ask a question any better than the airline reservation computer could.