In a lecture at the University of California at Berkeley, IBM research scientist Eric Brown outlined the history of the project, and provided some details about how Watson was able to sort through a variety of structured and unstructured data in the fastest time possible. His team of 30 engineers spent four years designing the current system, and believe it has great potential for non-gimmicky purposes.
Watson runs on 90 IBM 750 servers, with 2,880 Power7 cores running on 3.55GHz processors. It has 15TB or memory and can pump out 80 teraflops. This is a commercially available configuration, but Watson's secret sauce is IBM’s DeepQA data-handling software. Brown said that to answer a question on this rig eventually took under three seconds, compared to the two days it would have taken a single processor.
The top human Jeopardy players are very, very good, with the all-time champion answering nearly two-thirds of the questions in a match with 85 to 95 per cent accuracy. In 2007, the best the IBM team could manage was around 30 per cent accuracy, so they decided to shift their approach from sifting through large amounts of structured databases to looking at more unstructured data via Hadoop.
The second big shift in strategy was the abandonment of software rules wherever possible. Brown explained, for example, that while it might seem logical to set up a rule that a data set for “month” should only include the standard twelve, January to December, this left Watson flummoxed over questions of holy months such as Ramadan. Rather than set strict rules, the team relied on a statistical analysis of evidence to weigh probabilities of a specific answer being correct.
This final score vector is then analyzed by IBM’s DeepQA code, and run through a series of testing algorithms designed to find either supporting evidence or contradictory data. Risk factors from misunderstanding the question are also thrown into the mix.
The system is not always perfect, however. In the first round of February’s contest, the contestants were provided the answer: "Its largest airport was named for a World War II hero; its second largest, for a World War II battle," but the question that Watson provided was wrong: "What is Toronto?"
Brown explained that this case showed an occasional problem with the Jeopardy format and machine learning. The category was "U.S. Cities", but that term wasn't mentioned in the clue. There are also five towns named Toronto in the US. Watson scored Toronto at around 16 per cent likelihood, with Chicago – the correct city – a few percentage points lower, so got it wrong.
One surprising revelation about Watson is the relatively small size of the data set it works with. Taking into account structured and unstructured data, the machine only has to search through around 100Gb of text data for each answer.
IBM is touring US campuses with the machine in order to seek ideas and recruits to take Watson further. The company has already touted it for medical uses, and is also touting IT technical support and government software as possible areas for expansion.
Watson will be put through its paces again on Thursday night in a match between it and teams from Stanford University and UC Berkeley – teams composed of humans, that is. ®