USE OF DIFFERENT THRESHOLDS IN LEARNING AND RECOGNITION

The neural networks' ability to robustly recognize patterns is influen ced by the selectivity of feature-extracting cells in the networks. Th is selectivity can be controlled by the threshold values of the cells. This paper proposes to use different threshold values for feature-ext racting cells in the learning and recognition phases, when an unsuperv ised learning with a winner-take-all process is used to train the netw ork. During the recognition phase, better performance is achieved when the thresholds are set lout enough to maintain the generalization abi lity. The thresholds in the learning phase, however, should be kept hi gher than in the recognition phase. If the thresholds in the learning phase are made as low as in the recognition phase, a sufficient number of feature-extracting, cells cannot be generated in the network becau se of the competition among the cells. The effectiveness of adopting t wo different threshold values is demonstrated by computer simulation, taking the 'neocognitron' as an example.