Neural Network Design

Question

I am playing around with Neural Networks and wondered how the experts decided how many neurons is enough or say know what range would be proficient. Does one decide based on heuristics or are there logical amounts based on the application?

Say for example I wanted a simple network to be able to tell the letter I with say a grid of 9x10.

0 0 1 1 1 1 1 0 0
0 0 0 0 1 0 0 0 0
0 0 0 0 1 0 0 0 0
0 0 0 0 1 0 0 0 0
0 0 0 0 1 0 0 0 0
0 0 0 0 1 0 0 0 0
0 0 0 0 1 0 0 0 0
0 0 0 0 1 0 0 0 0
0 0 0 0 1 0 0 0 0
0 0 1 1 1 1 1 0 0

So this would have obviously 90 input neurons. Let's say for simplicity, one output neuron (0 or 1 if it is the letter I or not). My question is, based on this network and application design, how many hidden layer neurons is enough to produce less error? How does one come to such a conclusion?

Answer 1

Here's a FAQ that looks helpful. I don't think there's any know formula or anything specific, it probably depends on the input domain.

Actually for something that simple you wouldn't need very much.