Abstract
The neocognitron is a multi-layered convolutional network that can be trained to recognize visual patterns robustly. This paper discusses training methods for the neocognitron. The add-if-silent rule is used for training intermediate layers of the hierarchical network. By the add-if-silent rule, a new cell is generated when all postsynaptic cells are silent. The generated cell learns the activity of the presynaptic cells in one-shot, and its input connections will never be modified afterward. At the highest stage of the hierarchical network, the method of interpolating-vector is used, not only for classifying patterns during recognition, but also for the training. In the highest stage, each feature-extracting cell is given a label indicating the class name of a training vector, and its preferred feature is represented by a reference vector in the multi-dimensional vector space. We assume plane segments spanned by every trios of reference vectors of the same label. During the recognition phase, the label of the plane segment nearest to the test vector shows the result of classification.
