I second that thought. There is a pretty well cited paper from the late eighties called "Multilayer Feedforward Networks are Universal Approximators". It shows that a feedforward network with a single hidden layer containing a finite number of neurons can approximate any continuous function. For non continous function additional layers are needed.