Estimating hidden units for two-layer perceptrons

A method of estimating the number of hidden units required by a two-layer perceptron learning binary mappings using back propagation of error signals is presented. In order 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 output unit. A conflict is a set of input/output relationships that require incompatible weight solutions when the responses of an output unit are learned on a single layer perceptron. The estimate produced is data-dependent, since 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