The competition algorithm of the hypercolumn neural network
Tarek El. Tobely, Naoyuki Tsuruta, Makoto Amamiya · 2004
The Hypercolumn neural network model (HCM) is an unsupervised competitive network consisting of hierarchical layers of hierarchical self-organizing map (HSOM) neural network arranged as the cell plans of the Neocognitron (NC) neural network. HCM combines the advantageous of both HSOM and NC while rejecting their disadvantage and are seen to alleviate many difficulties associated with image recognition applications, where it can recognize images with varying object size, position, orientation, and spatial resolution. However, due to the hierarchical structure of the HCM model, the network spends a long time in the recognition. The HCM model is introduced with a new competition algorithm to reduce the network recognition time into the real-time range. The proposed competition algorithm is based on selecting a subset from the most discriminate codebook of the network weights. This can drastically reduce the network recognition time into the range of real-time rate.