Unsupervised feature extraction by hypothesis of feature centres

The unsupervised training technique of the neocognitron is not robust to all types of pattern. Difficulty is experienced for patterns whose second order features occur at widely different distances apart. This creates problems in choosing the size of the circular area in which to search for a feature. The authors propose an improvement to the neocognitron to overcome this. The modification is to direct the network at the second level of hierarchy to concentrate at specific positions of the input space which correspond to the centres of features. For the positions to be specified unsupervised, it was necessary to incorporate into the network a hypothesis for the centre of a feature. The experimental results show that inclusion of the hypothesis leads to a more robust unsupervised training technique up to the second level of hierarchy than that used in the neocognitron. The new architecture of the network is designed to retain the tolerance to shift, the deformation of input patterns and also the simplicity and modularity of the neocognitron