IN this period of widespread economic uncertainty, in which technology vendors have to come up with all kinds of creative ways to cajole clients into spending more on their products and services, IBM's Manoj Saxena claims to have an interesting problem with demand for the product under his charge.

"I have to put my shoulder on the door to prevent everyone from coming in and saying, 'We want Watson,'" says Saxena, general manager for Watson Solutions, IBM Software Group. Saxena is referring to the latest in the IT giant's series of supercomputers, named for the company's founder Thomas Watson.

Watson demonstrated its prowess in a big way last year when it beat two top contestants in US TV game show, Jeopardy, which not only tests participants' general knowledge but also their language skills.

Watson was first developed in 2006, when IBM Research, the company's unit in charge of deep-dive R&D, took up the challenge of creating a computer that could understand human speech and answer questions with greater speed and accuracy than humans, while continuously learning and improving. Dr David Ferrucci, a senior scientist with IBM, was given the go-ahead to assemble the team and assets needed in a process described by Saxena as akin to "going to the IBM candy store".

In a sense, developing Watson is becoming a company tradition. An earlier IBM supercomputer, dubbed Deep Blue, drew on its ability to calculate 200 million positions in a second to defeat world chess champion Garry Kasparov in a series of matches in 1996 and 1997. What makes Watson's achievements more remarkable than Deep Blue's is the vast improvement in the way these computers are made to think, guided by what is called "evidence-based probabilistic architecture".

The third generation

The first generation of computers, seen from today's perspective, were but glorified tabulators. The second generation, epitomised by Deep Blue, were programmable. Watson, designed a decade later, has gone beyond knowing how to calculate digits and other so-called "structured data". It has learnt how to learn "unstructured data", like the human language on top of machine language. "The focus is not on search but discovery, where applications are not deterministic but probabilistic. We are working with systems of engagement, not just systems of record. This is the beginning of where computers are going to reason and learn," says Saxena, at the recent IBM InterConnect conference held in Singapore.

Jeopardy is a good test because, unlike chess playing, where moves are structured and there is a definite number of possibilities, the data that Watson has to analyse is a series of questions not merely factual but laden with metaphors, synonyms and all the complexities found in the English language.

Saxena's favourite example was "Chicks dig me", famously uttered by Bill Murray in the 1981 movie Stripes. When Watson encountered that line for the first time, it refrained from responding and, instead, observed how the other (human) contestants responded. It figured out whether the phrase was a metaphor or simile, or whether there were other ways of looking at it. "It is the ability to constantly adapt and learn that makes Watson unique," he says.

While it is one thing to create a Watson to win Jeopardy, it is another to design other Watsons to be really useful in the world outside TV studios. In fact, Saxena has drawn up a rather long list of requirements Watson needs to meet. For example, whereas Jeopardy machine was a "single use, single session" machine, in the real world, if it is to be used in, say, healthcare, there could easily be tens of thousands of users concurrently.

Read the original here:

IBM takes Watson super computer to next level