Practical Neural Network Recipies in C++

180

Chapter 10

required. The tendency is invariably to overestimate the requirement.

It is not at all unusual to have a problem with hundreds of input

neurons and several output neurons which only requires five or so

hidden neurons. When in doubt, start low and work up as needed. It

is unfortunate that using fewer hidden neurons often increases the

likelihood of the learning algorithm becoming trapped in a local

minimum. Additional weights can create new channels through which

gradient descent is able to pursue a global minimum. In practice,

though, the tradeoff is rarely worthwhile. Stick with the minimum

number necessary to solve the problem ...