The role of the number of training samples on weight initialisation of artificial neural net classifier

The number of training samples is one of the important factors that determines the number and the depth of the local minima of the pattern error surface of an artificial neural network (ANN). In the case of a small training sample, the local minima are deeper and their number is large. Simulation studies have shown that the addition of noise to each training vector in each training sweep helps to flatten the pattern error function and thus to reduce the negative local minima effects. The nonrandom initialization of the weights of the ANN can also be useful. When solving important pattern recognition problems by means of ANN classifiers, the authors recommend using several starting-point-determining techniques and performing training several times.<>