NEOCOGNITRON WITH DUAL C-CELL LAYERS

The neocognitron is a hierarchical neural network model capable of def ormation-resistant pattern recognition. In the hierarchical network of the neocognitron, feature extraction by the S-cells and a blurring op eration by the C-cells are repeated. The ability of each S-cell to rob ustly extract deformed features is created by the blurring operation o f the C-cells placed in front of the S-cell. In the conventional neoco gnitron, the amount of blurring produced by the C-cells is uniform in the receptive field of each S-cell. An S-cell would accept a much larg er deformation if a nonuniform blurring could be produced in such a wa y that a larger blurring is generated in the periphery than at the cen ter of the receptive field. This is desirable as discrepancies between a training pattern and a deformed stimulus pattern usually become lar ger in the periphery than at the center of the receptive field. In ord er to produce such a nonuniform blurring economically, we propose a ne ocognitron with a dual C-cell layer A layer of C-cells in an intermedi ate stage of the network is divided into two sub-layers: one with a sm aller blurring and the other with a larger blurring. Each S-cell in th e succeeding layer receives input connections from the low-blur C-cell layer al the center of its connecting area while also receiving conne ctions from the high-blur C-cell layer at its periphery. Computer simu lation has shown that the new neocognitron recognizes characters more robustly than the conventional neocognitron.