A slimmed-down Watson supercomputer is on its way from quiz show winner to a career in medicine, according to David Ferrucci, who led the IBM team that built the machine.
Speaking at the Alan Turing Centenary Conference in Manchester (Saturday, 23 June), Ferrucci said that after the machine’s win on the leading US quiz show Jeopardy!: “The future direction is to show how [Watson] can be applied. One of the areas we are excited about is healthcare.”
Ferrucci explained that updated versions of the machine and its software could aid diagnoses by filtering and scoring the massive amount of medical information that is now available online. Instead of answering simple questions, the machine would enter into a dialogue with doctors to work through symptoms and look through the evidence to find the most likely causes for a medical problem. He claimed Watson is highly suited to the task because of the way in which the software evolved during the Jeopardy! project.
“Watson is good at finding sources of evidence and scoring them. We are taking those techniques and adapting them to medical knowledge,” Ferrucci said, adding that optimisations are bringing down the cost of implementation.
Ferrucci explained that the original design uses a very large supercomputer. “We were under a tremendous amount of time pressure. Senior management were telling me that I was messing with the IBM brand.” As a result, the emphasis was on making Watson work rather than how it cost.
“We weren’t using the most efficient hardware,” said Ferrucci, adding that later implementations of the software that are more streamlined can run on a machine that uses a fifth of the hardware capacity of the Jeopardy! winner.
He added that future, more complete artificial-intelligence systems will most likely use a form of the Watson technology.
In contrast to standard AI systems, Watson uses a number of statistical techniques to work out which pieces of evidence are relevant to a problem. “The way Watson is architected now, it can’t do deep inferential analysis.” Ferrucci explained. “I don’t know what the answer is for AI but I know it’s the combination of these two things: statistical and logic-based techniques.”