Abstract
A neural network is often expected to have the noise-resisting ability, which is the property that the outputs keep the desired signals against the inputs perturbed with some noise to some extent. In this paper, the learning problem which realizes the noise-resisting ability is formulated as an optimization problem minimizing the maximum valued output error function, and “Restriction Learning Method” is proposed to solve it.
This learning starts by solving a restricted learning problem using standard input patterns without noises as the training signals. In progress of learning, other input patterns with noises which give the maximum errors are generated iteratively under the synaptic weights obtained by a restricted problem, and the corresponding error functions are supplemented into the set of error functions in the restricted problem. Lastly, when the necessary and sufficient number of inputs with noises as the training signals are generated, the learning terminates in acquisition of the resistibility against arbitrary perturbation within a certain range. The learning results for the simple recognition problem show that the noise-resisting ability is obtained by the restriction learning method.
