Brains, Sex, and Machine Learning

Here is a GREAT talk by Geoffrey Hinton (the Prof running said lab) [this video] where he explains the method. Unfortunately, even though it was posted three times to HN [link] it never made the fron page. Here is the my summary and comment: " Great talk. I don't know much about artificial neural networks (ANN) and even less about natural ones, but I have the feeling that I learnt a lot from this video. If I understand correct, Hinton uses so many artificial neurons compared to the amount of learning data, that you would usually see an overfitting effect. However, his ANN's randomly shut of a substantial part (~50%) of the neurons during each learning iteration. He calls this "dropout". Therefore, a single ANN represents many different models. Most models never get trained, but they exist in the ANN, because they share their weights with the trained models. This learning method avoids over specializing and therefore improves robustness with respect to new data but it also allows for arbitrary combination of different models which tremendously enlarges the pool of testable models. When using or testing these ANNs you also "dropout" neurons during every prediction. Practically, every rerun predicts a different result by using a different model. Afterwards, these results are averaged. The more results, the higher the chance, that the classification is correct. Hinton argues, that our brains work in a similar way. This explains among other things a) Why are neurons firing in a random

In addition to what bravura said, Geoff Hinton's group at UT has recently introduced a new approach to training neural networks that they call "dropout". You can read about it in their paper ( [link] ) or watch Hinton describe it in a recent talk he gave at Google ( [this video] ). Roughly speaking, dropout training provides a strong regularizing effect through a sort of model averaging that is conceptually related to the well-known bagging approach from which random forests derive their power and flexibility. Dropout training has already produced state-of-the art results on several time-worn standard benchmarks and helped Hinton's group win a recent Kaggle competition (for an overview of their approach, see: [link] ). I've played around with this a bit over the last few weeks, and have a Matlab implementation publicly available from my Github at: [link] .

Speaking of bagging: Geoff Hinton talks a bit about model averaging and has a very stimulating pair of hypotheses related to this, which are: a) genetic recombination in the form of sex might be "bagging across genes" that prevents tight co-adaptation (= overfitting on the evolutionary timescale) b) the brain uses noisy discrete firing instead of continuous communication because it allows for "bagging across network topologies". The bulk of his talk is about how dropout in neural nets can be used to accomplish some of the same performance benefits that model averaging gives on other algorithms (at a much smaller relative computational cost). Here's the talk: [this video]

>As for contrastive divergence with Restricted Boltzmann Machines overcoming the vanishing gradient problem. That's true, but has anyone even tried doing this with 7-bit floating point and demonstrated it even works? I wasn't referring to RBMs here, although it does look like I misremembered the details. See slides 74-76 of this presentation for a brief sketch [0]. In these feed-forward MLPs, the information content from each neuron is capped at one bit (which is quite a bit less than seven). I seem to recall him saying something similar (and with more detail) in a Google Tech Talk a few months later [1], although I'm not sure. [0]: [link] [1]: [this video]