Connectionist Learning with High-Order Functional Networks and its Application

Abstract

We consider the problem of representing and learning of continuous complex patterns in the 2-dimensional space by muli-layer networks. We propose a new network architecture, high-order functional networks with Chebychev polynomials as input units. It is shown that the structure of characterizing continuous patterns can be generalized by the high-order functional networks that are trained with backpropagation. The property of extracting the specific features of continuous patterns without losing the essential meanings of their original patterns is indispensable to many pattern information processing problems. We propose a structural pattern classification method with high-order functional networks. We apply this methodology to the task domain of a speech recognition and show its advantages over the traditional non-structural approach.