On July 27th, Dr. David Ferrucci came to Fermi to give a talk about IBM Watson. He talked about the architecture of Watson, some technical details, how the team members cooperated with others.
He says that Watson got its knowledge from only 100GB of data. After pre-processing for fast calculation in the game, it became 800GB. It was all in the memory in the game, because disks are slow.
Knowledge was not strictly categorized. Like, cat is a kind of animal, pencil is a kind of tool. Watson did not do this. I think one of the reasons is that it is impossible to build categories for all the things in the world. It just had to be fuzzier to be general.
Regarding why Watson believes Toronto was a US city, the explanations were:
- Even though the question belongs to a category about US cities, the answer does not have to be a US city, according to some previous questions of Jeopardy!.
- The confidences on Chicago and Toronto were so close. (Chicago had a chance to win.)
- Cities are not categorized as said before. So Toronto MIGHT be a US city. They even have a team in NBA… This is the way how Waston discovers the answer: get clues, connect them.
My phone has to make loud sound when taking a picture. So I waited until everybody was laughing so it would cover the sound of my phone. Succeeded!