Hacker News Reader

Comment by T_S_ | Hacker News Reader

Not sure neural net expert Yann LeCun (yann.lecun.com) would agree with your neural networks comment. Neural nets have taken on a life of their own since the time they were originated and actually have little direct connection with neuroscience these days.

View on HN

This is true, but the very good deep learning results are very recent (from less than 5 years ago), and they generally require very different models and/or training strategies from the standard multilayer perceptron you find in most AI textbooks. So it's an omission, but I'm not aware of any textbooks that cover these new neural networks in depth, as basic research in this area is still going in a frantic pace.

Yeah, I'm finding that there's little book-format info on deep neural networks as well. Similarly for belief networks. Gotta hunt down some journals, and wield some Google-fu.

The most promising approaches for neural nets these days involve learning generative models using Restricted Boltzmann Machines (RBMs). Geoffrey Hinton at UToronto has some good seminal papers on these, like http://www.cs.toronto.edu/~hinton/absps/ncfast.pdf. See http://www.cs.toronto.edu/~hinton/ for excellent videos of digit generation / classification.

As a note, this approach outperforms all others (sans tweaks), whereas neural nets used to be beaten by SVMs and the like.

Actually, the best recent results in this area involve discriminatively trained stacked auto-encoders using second-order algorithms (for which the code, as of Sept 2010, is still private to the group at the university of toronto), as described in James Martens's paper at this year's ICML: http://www.cs.toronto.edu/~jmartens/docs/Deep_HessianFree.pd...

Reply on news.ycombinator.com