Practical Neural Network Recipies in C++

10

Chapter 1

of the network's weights, is called an epoch. The number of samples

in the subset is called the epoch size. Some researchers use an epoch

size of one, meaning that the weights are updated after each training

case is presented. The author usually favors using the entire training

set for each epoch, as this favors stability in convergence to the

optimal weights. Compromises between these extremes are popular.

When the epoch size is less than the entire training set, it is important

that the subset be selected randomly each time, or ugly oscillations

may occur. Epochs of training are repeated until the network's perfor-

mance is satisfactory ...