Instead of Jeopardy, I always wanted Watson to go on "are you smarter than a fifth grader" and throw questions like "think of the second color of the rainbow. What large bird starts with the same first letter?"<p>Or: "If I call A 5, B 4, and C 3, then what do I call D?"<p>As a scientist (working on cancer, no less), more than just churning through terabytes of data, it strikes me that THESE sorts of questions are the ones you'll need to be able to answer. There's also a lot of "reading between the lines" you'll need to do, and a lot of "small pattern extrapolation". For example, the drug candidate I'm working on is one of a family of about let's say, 16 known compounds. All of them kill cancer cells, fourteen are cardiotoxic. The two that are not, lack a single oxygen atom. Maybe that oxygen atom lets the molecule block the mitochondrion (sort of important for heart cells)? Why? Because the molecule family looks sort of like Coenzyme Q10 - if you use your imagination and remember some of your chemical reactivity rules from Grad-level Ochem - and supplementation of Q10 alleviated the cardiotoxicity in one study on rats for one of the molecules.<p>Ultimately, my previous boss was bold enough to latch onto that and give it a shot - can we take this molecule that is the most potent of the group (but unfortunately cardiotoxic), delete the homologous oxygen atom and make something that still works and is less cardiotoxic?
As far as I'm aware there have been zero peer reviewed publications showing Watson actually doing something in medicine. When a doctor thinks he or she has a better approach than what is being done, they use the scientific method to find out of their hypothesis holds up. So far I've seen lots of IBM conference presentations and press conferences. But where is the data? Where are the increased survival rates? Where are the reduced medical errors compared to controls? <i>That</i> is science. Everything else is snake oil unless you have that.
This is truly incredible stuff. Hats off to IBM for pushing the field of machine learning forward.<p>This statement really struck me:
"“In the Jeopardy days [in 2011], Watson was running off 90 servers and could store 15 terabytes of memory,” said Gold. That’s not your typical household integration. “Now it’s far more affordable and runs 240 times faster, so we can do far more with less.” Moreover, the size of IBM’s server is now smaller than a pizza box and can fit in any data center."<p>How is this possible?
Even if all some AI system did was take a list of symptoms, anonymize them, suggest a diagnosis, and then take the MD's actual diagnosis, track outcomes, and compare results over time, it would be a huge advancement.<p>Right now, unless I misunderstand something structural about our medical system, most of the data gathered by MDs are wasted, because there is no digital record of much of it and patient privacy legalities prevent wholesale entering of symptoms and case outcomes in to any sort of freely available (inter)national database.
Where can one learn more about how Watson actually works?<p>...starting from the IBM website and Wikipedia one mostly gets to whitepapers and high level descriptions. I hope at least bits of it are open source or at least have the math and logic underlying them well documented.