Neocognitron learning by backpropagation
Michihiro Ohno, Masato Okada, Kunihiko Fukushima · Systems and Computers in Japan · 1995
Abstract In the neural network for pattern recognition, when the selectivity of the feature‐extracting cell is lowered to enhance the generalizing power, a tendency is produced that patterns with similar shapes but belonging to different categories are confused. Because of this property, it has been difficult to en‐hance the discriminating power to separate patterns with similar shapes without deteriorating the generalizing power. If patterns are observed, however, focusing on the features which differ between similar patterns, it will be possible to discriminate similar patterns. In the earlier neocognitron, the feature‐extracting cell has uniform sensitivity within the receptive field. Then, it is impossible to discriminate patterns by emphasizing the local feature, which is located at a particular position in the receptive field. This paper proposes a method of learning, where the sensitivity of the feature‐extracting cell is made variable depending on the position in the receptive field, and the sensitivity is adjusted by the error backpropagation. After the learning, the feature‐extracting cell exhibits a high sensitivity at the position where similar patterns have different features. Thus it is made possible to classify similar patterns without deteriorating the generalizing power.