Learning of category boundaries based on inverse recall by multilayer neural network

The author describes a method for inverse recall by a multilayer neural network with constraints. He also describes a query learning algorithm using the inverse recall for learning category boundaries in a pattern space. The method considered finds patterns near category boundaries and automatically produces a graded set of patterns between the adjacent categories. Seed patterns are selected based on the instability of the output signals of the network when the pattern is slightly perturbed. If another category´s response increases, that category is selected as a target. Patterns are generated by inverting the network´s mapping near the boundary. Since the inverse map is undetermined, constraints based on the domain are used to choose a particular inverse. The author first demonstrates the inverse recall method using handwritten numerals and Chinese characters. Then he demonstrates the automatic seed selection method on a two-dimensional distribution, and shows that the method cuts the number of patterns needed for boundary learning in half compared to a naive method