Another interesting thing is that perceptron can approximate any non-linear function with just a single hidden layer.

http://cstheory.stackexchange.com/questions/17545/universal-...