IBM's Watson computer soundly defeated the two highest-rated Jeopardy former champions earlier this year, signifying the dawn of a new age of computer-assisted decision-making. It was also a massive marketing coup for IBM and Jeopardy. Stephen Baker's "Final Jeopardy" reports on how Watson came to this latest achievement, along with hints of what comes next.
Originally computers were limited to rapidly processing data according to predefined rules, first for individual questions (eg. military applications), then larger and larger batches of commercial data, and eventually immediate 'real-time' response. Next came IBM's 'Deep Blue' defeat of the reigning 1997 chess champion, Garry Kasperov, demonstrating its ability to analyze millions of scenarios/second to select the moves most likely to win - an impressive new mathematical achievement. Winning Jeopardy, however, required the computer to analyze English natural-language words and phrases, despite confusing puns, idioms, allusions, double-meanings, trick questions, and the importance of syntax, and then select the most likely correct answer from a giant electronic database containing Project Gutenberg (33,000 highly rated e-books), dictionaries, taxonomies, thesauri, encyclopedias, news articles, book abstracts, famous works of music and art, etc., without an Internet connection - all within a maximum of five seconds. Watson also had to determine optimal betting strategies, taking into account current point rankings, as well as the potential earnings and betting opportunities remaining.
David Ferrucci, an artificial intelligence (AI) expert within IBM was selected to lead its Jeopardy effort. He first declined the request, believing it impossible to build a system within the requested three years that could beat the best-known Jeopardy champions. Later, in response to a second request, he decided that it could be done and led a team of 25 scientists that accomplished the task within four years.
Some thought the task would simply require a Google-like program; Google, however, does not actually answer questions - it simply provides information for humans to do so. On the other hand, simple questions (eg. What is the highest mountain?) represent about 30% of Jeopardy queries, per Baker. It was the remaining 70% or so that mostly concerned Ferrucci and his colleagues, as well as Ken Jennings' fast responses and historical correct answer rate of about 92%. The second bit of 'good news' for the team was their realization that Watson didn't need to know the content of books, artwork, plays, music, sitcoms with the level of detail required to eg. compare and contrast their contents - just key points and protagonists.
Initial Watson responses, says Baker, took 90 minutes to two hours - obviously inadequate for Jeopardy. Ferrucci and his IBM team solved that problem by reconfiguring Watson to consist of 90 IBM server computers, each with four 8-core processors, and using 15 TBs (terabytes) overall of rapid access (solid state) memory (RAM), and then dividing its tasks into thousands of stand-alone jobs distributed among 80 teraflops of processing power for parallel processing. (1 TB of memory can hold 220 million text pages; a teraflop = 1 trillion operations/second.)
Unfortunately, Baker's book does not provide details on how these Linux-managed algorithms and processors are able to work together and in sequence. Neither do we learn much about how AI takes eg. any subject and determines what types of words are most and least associated with it, or Watson's use of some 100+ algorithms to analyze a question in different ways, the results of which are then ranked by another set of algorithms (eg. "Was the contender alive that year?"). (New York Times, 6/20/2010).
"Final Jeopardy" also provides few details about the hardware, requiring two 20-ton A/C units to cool, and operating at 3.3 gigahertz (gigahertz = billion cycles/second).
Much of the Watson project's time was taken up with devising ways to prevent errors uncovered in its constant preparation competitions with lesser-known Jeopardy champions. The group also was worried about Watson making IBM look bad by losing badly or spitting out inappropriate four-letter words. Then there were battles with Sony (Jeopardy's owner) over requiring Watson to use a mechanical finger to 'buzz in' (IBM agreed; a number of contestants proved faster by anticipating host Trebek's speaking), allowing Watson to 'see' the correct response after every question (allow Watson to improve its understanding about what the category was looking for - Sony agreed), and insuring that the category/question selection was done in a new, incontestably fair manner (Sony agreed - IBM did not want Watson's achievements to be challenged on the basis of some bias. Visual clues, however, were excluded, because of Watson's inability to process such.)
Baker also tells us that the best Jeopardy champions had their own preparation and competition strategies. Some contestants used 'Anki,' a site that provides electronic flashcards for hundreds of study fields, and J! Archive, a site that helped both IBM and various contests determine the likelihood of various categories being used. The best, including Watson, deliberately searched for Daily Double questions (more likely found in the higher-dollar questions), especially in Double Jeopardy, as a means of both keeping those potential game-changers out of opponents' hands and possibly boosting one's own standing, if needed. Google Scholar was also used by some contestants (and probably IBM as well) to refine Final Jeopardy betting and overall answer generation.
Watson's ultimate triumph against Ken Jennings and Brad Jennings in "Final Jeopardy" is a bit anticlimactic, though not without a few scary blunders by Watson. The biggest - identifying Toronto as an American city. Strangely, despite Watson's triumph, IBM was simultaneously struggling to avoid cancellation of a contract with the California Dept. of Education to answer much simpler questions such as "Do 8th-grade students enrolled in Algebra I perform better, on average, if their teacher has a credential in math?" (Huffington Post, 2/22/2011).
The first computer I worked on (1967) had 16 K (K = thousand units) of memory and a 10 K processor speed (thousand operations/second). Watson's memory is over 1 billion times greater, and its processor speed 8 billion times faster. China's 'Big Red,' the world's fastest computer, is over 30X faster than Watson; Big Blue, however, is expected to counter with 'Sequoia' in 2011 - about ten times faster yet. Thus, this year will bring processing speeds about 2.5 trillion times as fast as the mid-1960s, and 300 times as fast as Watson. More efficient algorithms will undoubtedly accompany the increased power.
Baker sees Watson as already able to easily digest and synthesize the tens of thousands of medical research papers published every year to help physicians make diagnoses and treatment plans. (A general lack of medical research meta-analyses is a major problem in medicine that future Watson's could likely provide.) IBM is now moving in this direction, starting by incorporating improved speech recognition software from Nuance Communications, and working with medical experts. Call centers, legal work, offer other potential, though less glamorous, applications.
Bottom-Line: Stephen Baker's "Final Jeopardy" gives us a glimpse into the future, where bytes better brains, just as previously machines and robots beat brawn. Watson's victory brings us to the point where the future becomes unpredictable, per Ray Kurzweil and others, due to our inability to imagine the capabilities of artificial intelligence increasing at exponential speeds, simultaneously pursuing and combining revolutionary discoveries in biology, nanotechnology, medicine, etc. ("The Singularity is Near") Readers will be left wondering what humans will do in the future, what incredible advancements in science and medicine are already underway, whether society becomes more controlled using this power (Bill Joy, co-founder of Sun Microsystems), etc.?