Features

GPS, Question 1) What first attracted you do data mining / machine learning and what education did you receive ?

Simon Funk

(answer continued from KDnuggets News 07:n08, item 4)

A few other inspirational tidbits stand out: I recall one talk I sat in on about using backprop to map from English letters to the phonemes of spoken English. The really interesting thing about this was that after the network was trained, they were able to create a sort of clustered hierarchy of categories for the letters of the alphabet, which is something linguists had been working on by hand for a long time. The results that popped out of this network in a day matched the linguists' graphs to a tee, but kept going where the linguists had left off!

I'm bushing over the details which I probably have wrong anyway, but I just remember getting this strong sense that wow: here was a class of tools that not only could learn things on their own, but can teach us. The way my Netflix algorithm discovers and reveals its own movie categories is a more recent example of the same thing, and I find it just as fun and inspiring now as then!

Another pivotal memory is, some time after college, reading Lynch and Granger's olfactory model. They go into some detail about how the olfactory cortex learns a hierarchical breakdown of smells through a cyclic process of successive refinement using reverse inhibition. Aside from just the interest in seeing how this sense of ours is actually implemented as an algorithm, it turns out that due to the timing characteristics of the various groups of neurons involved, the process works best when the stages of refinement cycle through at about four hertz--which is why we go "sniff sniff sniff..."!

So, here's this overt behavior we do, and the reasons for it trace back to the peculiarities of the particular algorithm our brain has evolved for learning and categorizing smells. If we can get inside of that, what about getting inside how we make our most basic choices, deciphering the mechanism behind the will itself? In fact this is largely what Granger is now pursuing at Dartmouth, so watch for some willful robots scurrying about their halls soon. (I may be headed there this summer myself.)

What's this?