On: Neural Networks Demystified

[limsup]

people are mostly interested in what is successful and backprop/gradient descent (with certain specific architectures) is what has been dropping error rates on all sorts of tasks in recent years.