Prime feature extraction in pyramid network

A simple features extraction method for a connection weights fixed multi-layer feedforward pyramid network is proposed in this paper. Pyramid network development was biologically motivated, and an innovative pattern recognition strategy is follows which divides the mapping work into neural network and coordinate feature extraction and code matching algorithms within a conventional Von Neumann computer. Benefiting from its unique structure and simple feedforward mapping method, pyramid network can perform topology preserving mapping from a high dimensional space to a lower dimensional lattice. At the lower dimensional output layer, a code of the input pattern can be constructed by a set of gradient magnitudes of output layer neurons output and the average value of those outputs in a sequence of centre to surrounding neurons. Consequently learning and classification modules can be simplified by using algorithms to store feature vectors and matching these codes in a conventional memory. Experiments show that a small size set of 3×3 vectors can sufficiently represent many input pattern geometry features