These days we hear a lot about Watson, the IBM AI supercomputer that defeated two greatest Jeopardy players on February 16th 2011.
Watson is in size of 8 refrigerators, a cluster of ninety IBM Power 750 servers with a total of about 3000 processor cores and about 16-terabyte of memory. (For comparison the US Library of Congress contains about 10 terabyte of data)
The large memory is used hold ontological object graph of concepts, entities, properties and their relationships.
Putting it in Knowledge Modeling context Watson works using a chain of Knowledge Models as followed:
Diagnostic Model -> Explorative Model -> Analytic Model -> Selective Model
- The Diagnostic Model utilizes Natural Language Processing (NLP) to decipher the Jeopardy question and identify the "problem" (what is being asked for)
- An Explorative Model utilizes the internal memory to present all possible answers, each represented as a Hypothesis.
- The Analytic Model is in charge of evaluating and calculating the confidence level of each Hypothesis on multiple dimensions.
- And finally, the Selective Model ranks the options and chooses the best possible answer based on predefined thresholds and dimensional weightings.
According to the IBM research team Watson is using a dynamic weight setting that enables it to learn during the game. (For example adjusting its confidence level according to Jeopardy category)
To me, one of the most impressive part of Watson is the massive parallelism that makes a timely response possible.
Apparently the overwhelming victory in Jeopardy resulted in a partnership between IBM and Nuance to apply Watson technology in Healthcare industry, with the first commercial offering planned within 24 months!!