[jbelanich]

You're correct in that neural networks as a model have been around for a long time. However, those networks were restricted to be shallow because backpropogation didn't work well on networks with many hidden layers. Only recently have researchers developed learning procedures that can learn these deep architectures efficiently, using some clever unsupervised learning techniques. And surprisingly, they are finding that these deep networks perform remarkably well, beating the state of the art in a number of benchmarks.

You are also right that you do need a lot of processing power to get neural networks to work well. But that is changing rapidly. Hinton's convolutional neural network has the state of the art in the ImageNet benchmark, yet was trained using significantly less power than google brain. Regardless, you don't need google scale computation to get deep networks to work well. The point of google brain is to see how far one could push neural networks.

[jcrites]

> Only recently have researchers developed learning procedures that can learn these deep architectures efficiently, using some clever unsupervised learning techniques.

Would you mind naming some of these techniques, if you're familiar with them? I'd like to take a deeper look.

[visarga]

Restricted Boltzmann Machines. https://www.youtube.com/watch?v=AyzOUbkUf3M

This video drives the point home, and is made by the author of this technique.