A new scheme which incrementally generates neural networks for distorted handprinted Kanji pattern recognition

Abstract

We present a recognition system that incrementally generates neural networks to solve the problem of error caused by sample distribution overlap among categories. The first stage neural network performs the easiest task which is to separate mostly non-overlapping distributions, and leaves the difficult tasks such as separating overlapped distributions to the neural network(s) generated in the following stage(s). The new system improves its performance by assigning tasks to neural networks according to the degree of task difficulty and forms a specialized neural network. The new system achieves higher performance for the recognition of distorted Kanji patterns than the traditional neural networks which consist of only one neural network. The ability of the system to eliminate overlapped distributions is confirmed by analyzing the output distribution of the hidden units.