Features

GPS: 5) What motivated you to share you code and ideas (we note that only one other top leader described their method) ?

Simon Funk:

There's no limit to what you can accomplish if you don't care who gets the credit. - anonymous

The bottom line is, life is short and I want to see things happen. I know for many this competition is between teams, but to me it's just part of a grander competition with time. The way I see it, there are a lot of pieces that need to fall into place to make some really fun and interesting things happen within our lifetimes, and this contest touches on one of those pieces. People may be viewing SVD as just a neat mathematical trick for minimizing RMSE, but when you take a step back there are a lot of interesting ties with other learning algorithms for pattern recognition, Bayesian inference and the like.

For instance, with this criss-cross Hebbian learning law where the activation from one set of nodes controls the plasticity of the other, just by changing the way the activations are normalized within each group you can shift from minimizing rmse to maximizing likelihood. In other words, nearly the same algorithm will do Bayesian learning. And of course one has to wonder if something similar isn't going on in the brain, in the cortex specifically, where bottom-up and top-down connections converge in a sort of 69 symmetry reminiscent of this relationship. In short, there are some general principles here which I hope will seed some ideas and further progress in the wider domain.

What's this?