Practical Neural Network Recipies in C++

100

Chapter 6

This section has provided the foundation algorithms for

training a multiple-layer feedforward network. The mean square

error, given an epoch of input presentations and desired output

targets, can be found. The weight gradient vector of this error can also

be computed. All that remains is to apply any of several standard

numerical techniques to optimize the network. The next sections will

focus on the most important such algorithms.

Training by Backpropagation of Errors

Backpropagation was the first practical method for training a multiple-

layer feedforward network. Its presentation in [Rumelhart and

McClelland,

1986]

was almos ...