A LEARNING-MODEL OF THE FEATURE-DETECTING CELLS FOR UNSUPERVISED PATTERN-CLASSIFICATION

A modified neural network unsupervised learning scheme by the feature- detecting cell is proposed. We improve the performance in learning cat egories by adding a modulation system and a competing system to the co nventional feature-detecting cell model. With the aid of the modultion system and the competing system, the cluster prototype which is close t to the input pattern will win the competitions and a winner dominate d learning will be controlled by the properly assigned bias values. Th e proposed model has the following features: it guarantees the corresp onding feature-detecting cell of each input pattern to be formed regar dless of the initial weights and the duplication (of the feature-detec ting cells formation for each sampled input pattern) can be reduced.