IBM Watson Heads for Jeopardy Showdown

Next week the Sun microsystems supercomputer known as "Watson" will require on two of one of the most accomplished Jeopardy players ever, Ken Jennings and Brad Rutter, in a three-game match beginning on February 14. In the event that Watson manages to finest the humans, it may represent the most important advance in machine cleverness since IBM's "Deep Blue" beat chess grandmaster Garry Kasparov in the late nineties. But this time about, the company also promises to make a business case for the technology. Unimportant pursuit this is not.

And impressive technology it's. On the hardware side, Watson is comprised of 90 Power 750 machines, 16 TB of memory space and 4 TB associated with disk storage, all housed in a relatively compact ten racks. The 750 is IBM's elite Power7-based server focused on high-end enterprise stats. (The Power 755 is actually geared toward high overall performance technical computing and differs only marginally in Computer speed, memory capacity, and also storage options.) Although the actual enterprise version can become ordered with 1 in order to 4 sockets of half a dozen-core or 8-core Power7 chips, Watson is maxed out with the 4-socket, 8-core settings using the top bin 3.55 GHz processors.

The 360 Power7 chips that make up Watson's brain represents IBM's best and brightest processor technology. Each Power7 is capable of over 500 GB/second regarding aggregate bandwidth, making this particularly adept at influencing data at high speeds. FLOPS-wise, a several.55 GHz Power7 provides 218 Linpack gigaflops. For comparison, the POWER2 SC model, which was the computer chip that powered cyber-chessmaster Deep Blue, managed a paltry 0.48 gigaflops, with the entire machine delivering a mere 11.4 Linpack gigaflops.

But FLOPS are not real story here. Watson's question-answering software presumably makes little use floating-point number crushing. To deal with the game scenario, the system had to be endowed with a somewhat advanced version of NLP. But according to David Ferrucci principal investigator for that project, it goes much beyond language smarts. The software system, called DeepQA, additionally incorporates machine learning, knowledge representation, and deep analytics.

Even so, the complete application rests on first understanding the Jeopardy indications, which, because they utilize colloquialisms and often unknown references, can be difficult even for humans. This is exactly why this is such an excellent test case for NLP. Ferrucci says the power to understand language is determined to become very important aspect of computers. "It has to become that way," he states. "We just cant imagine a future without this."

But it's the analysis component that we associate with real "intelligence." The strategy here reflects the open up domain nature of the problem. According to Ferrucci, this wouldn't have made feeling to simply construct any database corresponding to all possible Jeopardy clues. This type of model would have reinforced only a small small fraction of the possible subject areas available to Jeopardy. Instead their approach was to use "as is" information resources -- encyclopedias, dictionaries, thesauri, plays, books, etc. -- and increase the risk for correlations dynamically.

The strategy of course is to complete all the processing in real-time. Contestants, no less than the successful ones, must provide an answer in a few seconds. When the program was run on a lone 2.6 GHz CPU, it got around 2 hours to be able to process a typical Risk clue -- not a extremely practical implementation. But once they parallelized the algorithms throughout the 2880-core Watson, they were able to cut the processing time from a couple of hours to between 2 and also 6 seconds.

Even at that, Watson doesn't just cough up the answers. It forms hypothesis based on evidence it finds scores all of them at various confidence ranges. Watson is programmed not to buzz in until this reaches a confidence for at least 50 percent, although this parameter can be self-adjusted depending around the game situation.

To attain all this, DeepQA utilizes an ensemble of calculations -- about a million lines of code --- to collect and score the proof. These include temporal thought algorithms to correlate occasions with events, statistical paraphrasing algorithms to assess semantic context, and geospatial reasoning to correlative locations.

It can furthermore dynamically form associations, at training and at video game time, to connect disparate ideas. For instance it can learn in which inventors can patent info or that officials can submit resignations. Watson furthermore shifts the weight that assigns to different algorithms based on which kinds are delivering the much more accurate correlations. This facet of machine learning allows Watson to get "smarter" the more it performs the game.

Read More: Here