IBM's Watson Computer and the Future of Artificial Intelligence

Last Tuesday, Dr. David Ferrucci spoke at The Computer History Museum in Mountain View with The Financial Times' Richard Waters about IBM's Watson computer and the future of artificial intelligence (AI).

Watson is the most successful AI computer since Deep Blue, the chess-playing computer that played against and defeated chess grandmaster Garry Kasparov in 1997.

Watson's mission was to compete against the top champions of the popular trivia show Jeopardy. Dr. Ferrucci was in charge of the Watson project at IBM, and explained to a packed audience how the project came to be and where IBM is taking this technology in the future.

Artificial intelligence began to experience a renaissance in 1995 when research began on teaching computers how to digest and comprehend knowledge in the manner that humans do -- natural language. Natural language comprehension, the goal of Watson, really strikes at the heart of AI. After the success of Deep Blue, IBM executives wanted to find the "next big thing". An executive came up with the idea of a computer that could compete playing Jeopardy. It took a couple years to gain mindshare within IBM to take on what seemed like a harebrained idea, but Dr. Ferrucci stepped up to the plate.

Watson was significantly more challenging to build than its predecessor. As Dr. Ferrucci explained, when Deep Blue was created to play chess, its goal was to follow the rules and strategies of a well-defined game. There is no background knowledge or natural language processing needed to play chess; as such it is well-suited for a machine to play.

The project ran the risk of killing Ferrucci's career if it failed, but he described it as "a challenge too irresistible not to try". Over four years, 27 IBM researchers built Watson with Ferrucci.

Watson uses ~80 kilowatts of electrical power (approximately equal to 700 incandescent light bulbs), 20 tons of air conditioned cooling capacity, 15 terabytes of RAM (the approximate size of 4 million MP3 songs), and 2,880 CPU cores. It's built inside of a self-contained box, so as to constrain it in a way that would make clear that it was not communicating with the outside world or with humans while playing. It holds information equivalent to ~1,000,000 books.

At its core, Watson relies on computing many different possible answers to a question and generating and ranking probabilities about the correctness of each possible response. Watson will only answer a question if the highest probability is over a certain threshold (for example, 50%).

In order for Watson to buzz in, it:

1. Hears the spoken question, and turns that speech into text.
2. Generates queries for more information and gets back data.
3. Compiles a series of competing hypotheses from that data.
4. Ranks each competing hypothesis based on supporting data that it finds.
5. Picks and buzzes in with the top-ranked result (assuming it surpasses the "correctness" threshold).

In order to fairly interact with the Jeopardy game as a human does, Watson was outfitted with a robotic hand to depress the buzzer.

While Watson may no longer be competing against Jeopardy contestants its underlying technology, Deep QA, has a variety of applications in the real world. In the field of health care, Deep QA can reason over language and can deliver explanations of diagnoses it finds in natural language.

While Deep QA will never replace doctors in performing a diagnosis, it is a vast improvement over existing computer systems that perform diagnoses by giving doctors a 5,000 world deductive proof rather than an readily understandable explanation. Deep QA's abilities are applicable across law, education and defense.

To learn more about Watson, check out the TED talk, Final Jeopardy! and the Future of Watson.