LEARNING CAPABILITY ASSESSMENT AND FEATURE SPACE OPTIMIZATION FOR HIGHER-ORDER NEURAL NETWORKS

A technique for evaluating the learning capability and optimizing the feature space of a class of higher-order neural networks is presented. It is shown that supervised learning can be posed as an optimization problem in which inequality constraints are used to code the informati on contained in the training patterns and to specify the degree of acc uracy expected from the neural, network. The approach establishes: (a) whether the structure of the network can effectively learn the traini ng patterns and, if it can, a connectivity which corresponds to satisf actorily learning; (b) those features which can be suppressed from the definition of the feature space without deteriorating performance; an d (c) if the structure is not appropriate for learning the training pa tterns, the minimum set of patterns which cannot be learned. The techn ique is tested with two examples and results are discussed.