Monday, February 21, 2011

a really elaborate hardware sales pitch...,

The Economist | Four years in the making, Watson is the brainchild of David Ferrucci, head of the DeepQA project at IBM’s research centre in Yorktown Heights, New York. Dr Ferrucci and his team have been using search, semantics and natural-language processing technologies to improve the way computers handle questions and answers in plain English. That is easier said than done. In parsing a question, a computer has to decide what is the verb, the subject, the object, the preposition as well as the object of the preposition. It must disambiguate words with multiple meanings, by taking into account any context it can recognise. When people talk among themselves, they bring so much contextual awareness to the conversation that answers become obvious. “The computer struggles with that,” says Dr Ferrucci.

Another problem for the computer is copying the facility the human brain has to use experience-based short-cuts (heuristics) to perform tasks. Computers have to do this using lengthy step-by-step procedures (algorithms). According to Dr Ferrucci, it would take two hours for one of the fastest processors to answer a simple natural-language question. To stand any chance of winning, contestants on “Jeopardy!” have to hit the buzzer with a correct answer within three seconds. For that reason, Watson was endowed with no fewer than 2,880 Power750 chips spread over 90 servers. Flat out, the machine can perform 80 trillion calculations a second. For comparison’s sake, a modern PC can manage around 100 billion calculations a second.

For the contest, Watson had to rely entirely on its own resources. That meant no searching the internet for answers or asking humans for help. Instead, it used more than 100 different algorithms to parse the natural-language questions and interrogate the 15 trillion bytes of trivia stored in its memory banks—equivalent to 200m pages of text. In most cases, Watson could dredge up answers quicker than either of its two human rivals. When it was not sure of the answer, the computer simply shut up rather than risk losing the bet. That way, it avoided impulsive behaviour that cost its opponents points.

Your correspondent finds it rather encouraging that a machine has beaten the best in the business. After all, getting a computer to converse with humans in their own language has been an elusive goal of artificial intelligence for decades. Making it happen says more about human achievement than anything spooky about machine dominance. And should a machine manage the feat without the human participants in the conversation realising they are not talking to another person, then the machine would pass the famous test for artificial intelligence devised in 1950 by Alan Turing, a British mathematician famous for cracking the Enigma and Lorenz ciphers during the second world war.

It is only a matter of time before a computer passes the Turing Test. It will not be Watson, but one of its successors doubtless will. Ray Kurzweil, a serial innovator, engineer and prognosticator, believes it will happen by 2029. He notes that it was only five years after the massive and hugely expensive Deep Blue beat Mr Kasparov in 1997 that Deep Fritz was able to achieve the same level of performance by combining the power of just eight personal computers. In part, that was because of the inexorable effects of Moore’s Law halving the price/performance of computing every 18 months. It was also due to the vast improvements in pattern-recognition software used to make the crucial tree-pruning decisions that determine successful moves and countermoves in chess.