Estimating hidden unit number for two-layer perceptrons

The authors concentrate on estimating how many hidden units are needed for large nets. The design method is adapted from a more rigorous one developed for smaller nets where all possible input/output relationships are presented in the training vectors. It provides a 'ballpark' estimate of the minimum number of hidden units required by a two-layer perceptron for a provided training subset of the possible input/output pairs. Typically, the number of hidden units is slightly overestimated in this approach. To obtain an estimate of the number of hidden units for a fully connected net with n output units, it is necessary to obtain an estimate of the number of conflicts contained in the individual binary responses that must be learned by each ouput unit. The estimate produced is data dependent, as the number of conflicts for an output unit depends on the specific responses of the output unit to the input vectors contained in the training set.