Nonorthogonal visual image coding by a laterally inhibitory neural network

A two-layered, laterally connected neural network is proposed for modeling a nonorthogonal visual coding system. If the code primitives are given in advance (as biologically), it can be shown that the connection weights between input and output layers are just these primitives, while the lateral connection weights are formed by their inner products. In order to gain insight into the detailed nature of the network, Hebbian and anti-Hebbian rules are chosen for governing the modifications of feedforward and lateral connection weights, respectively. When the network is fed with random noises, it can self-organize according to these learning rules to develop masks resembling nonorthogonal receptive fields of simple cortical cells, as opposed to those models based on principal component analysis which seek to yield orthogonal feature detectors. At the same time it can perform optimal nonorthogonal image coding with respect to the code primitives being formed.<>