The ability of the neocognitron to recognize patterns is influenced by the selectivity of feature extracting cells in the networks. This selectivity can be controlled by the threshold of the cells. In the previous work, use of different thresholds for feature extracting cells in the learning and recognition was proposed: The thresholds in the recognition phase are set low enough to maintain the generalization ability. The thresholds in the learning phase, however, are set high to generate a suffiicent number feature-extracting cells. We verify this method using large hand-written character database. Using this method, we obtained a recognition rate of 97.4
展开▼
