Use of different thresholds in learning and recognition

Abstract

The neural networks' ability to robustly recognize patterns is influenced by the selectivity of feature-extracting cells in the networks. This selectivity can be controlled by the threshold values of the cells. This paper proposes to use different threshold values for feature-extracting cells in the learning and recognition phases, when an unsupervised learning with a winner-take-all process is used to train the network. During the recognition phase, better performance is achieved when the thresholds are set low enough to maintain the generalization ability. The thresholds in the learning phase, however, should be kept higher 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 because of the competition among the cells. The effectiveness of adopting two different threshold values is demonstrated by computer simulation, taking the ‘neocognitron’ as an example.