Garry Kasparov on IBM's Watson

<i>Worse, by definition they do not understand what they do not understand and so cannot avoid them</i><p>Kasparov didn't seem to see what I did. Watson seemed very consistent in knowing what it did not know. There was maybe two questions I recall where it actually got the question wrong with 50%+ certainty. I believe it answered, "leg" when it should have been "mising a leg". The other it answered the 20s when the answer was the 10s. And I think for neither of those the percentage was much beyond 50%.<p>Also Kasparov seems to miss that Watson in medicine would be used with humans. I doubt a doctor will say, "Watson says to cut off his left leg -- I would just given him aspirin for the headache, oh well. Hopefully cutting off this leg makes his head feel better."<p>What Watson hopefully will do is help diagnosis. Especially tricky ones.<p>There's a great story in a book I read, I wish I could recall the name, but it begins with a lady who has some stomach issue that she has for like 20 years. Everyone thinks its in her head. She finally happens upon a doctor who happens to have seen something like this before, she gets diagnosed and healed. But she had to live with it for like 20 years after seeing doctor after doctor. Watson would be able to greatly help situations like this, I hope.<p>UPDATE: The book is "How Doctor's Think". Here's an excerpt that talks about this case, <a href="http://harvardmedicine.hms.harvard.edu/bulletin/winter2007/7.php" rel="nofollow">http://harvardmedicine.hms.harvard.edu/bulletin/winter2007/7...</a> -- just in case anyone cares. :-)

Here are the reasons I was disappointed in Watson's showing (despite handily beating the human competitors). The most obvious was that Watson' auto-clicker was a big advantage over human thumbs, so that Watson got 100% of the points for clues to which all competitors knew the answer (if you asked Watson and the two humans "what's five plus five", Watson would win, but that's not necessarily proof of any sort of computer superiority).<p>The second reason is that IBM was representing Watson as something of a big push in knowledge representation (I just watched a video where they talk about Watson's "informed judgments" about complicated questions for instance). It looks instead like Watson just has an improved ability to disambiguate words relative to previous systems and to do quick lookups that match those words with nearby key terms.<p>For example, on the clue "Rembrandt's biblical scene 'Storm on the Sea of' this was stolen from a Boston museum in 1990", Watson correctly answered "Galilee". But its next two answers were "Gardner Museum" and "Art theft"; no one who "understood" the question in any conventional sense would even consider these as answers because they don't make any sense. Clearly, Watson looked for instances of "Rembrandt", "Storm on the sea of", "stolen", or other phrases from the clue in its text corpus, and found that "Galilee", "Gardner Museum", and "art theft" all frequently occurred when together (because the painting was stolen from the Gardner museum in an instance of art theft), and relatively rarely when not together. "Galilee" probably won out of these three because Watson is tuned to Jeopardy clue styles (whenever there is a quoted phrase in a clue followed by the word 'this', it's always asking for the answer that completes the phrase).<p>Similarly, Watson was far less confident on the clue "You just need a nap!" You don't have this sleep disorder that can make sufferers nod off while standing up." It still got the right answer of "Narcolepsy", but with a relatively low confidence of 64%. "Insomnia" had a confidence of 32% despite clearly being the opposite sort of sleep disorder, and "deprivation" appeared at 13%, despite not being a sleep disorder. Here Watson gets confused because the only term of the clue that appears more frequently with "narcolepsy" than "insomnia" is "standing up"; my guess is that if "standing up" had been replaced by some oddly phrased, uncommonly occurring synonym, Watson wouldn't have been able to come up with an answer, despite the clue conveying exactly the same information.<p>This kind of cleverness is certainly impressive, but it seems like it's an advance in tuning existing techniques to the format of Jeopardy, not an advance that will spark other successful projects down the line. IBM's goal of giving us "the computer from Star Trek" doesn't seem any closer; I don't see any evidence that Watson could have answered a question that required more thought or understanding than a simple text search. If there was the question "how many kings ruled England in between Henry the Fourth and Henry the Eigth" (8), then Ken and Brad would have been able to answer relatively easily, while my guess is that Watson would be stumped.

Not quite clear on why people keep pointing to the 'Toronto' question as proof that Watson is fundamentally flawed in some irreconcilable way.

Kasparov is spot on in that Watson's DeepQA has yet to prove itself in a meaningful way. If it proves itself as an effective medical advisor, that will be far more impressive than the Jeopardy win (as impressive as that was in itself).<p>I think everyone was disappointed in the applicability of the Deep Blue accomplishment in other fields. Were any of the special purpose ASICs used to defeat Kasparov used in any other application? As far as I know a significant part of the Deep Blue development team left IBM relatively soon after the accomplishment.

As noted in Paul Hoffman's recent book "King's Gambit", after his matches with both "Deep Blue" and "Deep Junior" Kasparov was exhausted:<p>"As with Deep Blue, he had once again let an encounter with a machine play games with his head. He had been obsessed with the idea that Deep Junior would never tire. 'The machine is never distracted by an argument with its mother," he told me, 'or a lack of sleep.'<p>And in the linked piece Kasparov alludes to the reported next approach IBM wants to take with Watson - support in medicine.<p>Kasparov's human reaction to his encounters with Watson's distant cousins brings up one obvious benefit in the use of technology like Watson for supporting medical decision-making - simply that such software will be less likely to miss something. Software is less likely to miss considering a diagnosis, ordering a crucial test, or following up on a finding - unlike the fallible 'I' who may have skipped a class in med school, or was up all night on call and just can't think straight, or am just occasionally more stupid than usual.<p>Diagnosis is the first thing people think of with technology like this, but in my opinion that's not the big problem Watson should tackle. Medical diagnosis in and of itself (dramatizations like the TV show 'House' notwithstanding), is not really that difficult 99% of the time. When you hear hoofbeats, you're very likely going to find horses and not zebras. A future Dr. Watson might occasionally be very helpful in pointing out very obscure (but uncommon) diagnoses. However, in my opinion the most helpful thing a Dr. Watson could provide is collecting, evaluating, and comparing evidence and outcomes as they are developed globally and locally (ie across broad swaths of medicine, but also within a single physician's own patient population), continuously educating the physician, and monitoring cases.<p>There is plenty of untapped medical data/evidence out there, but it's almost all hidden away in plain sight...text/natural language. I have to agree with Kasparov here, in that the primary advancement Watson represents was in moving farther down the path from syntax to semantics.