An infomax-based learning rule that generates cells similar to visual cortical neurons

A learning rule for a visual neural network is derived according to an information-maximization approach. Each basic module of the assumed neural network consists of two simple units and one complex unit. Each simple unit calculates a linear summation of its input by using its synaptic weights, and the complex unit calculates a squared sum of the outputs of the simple units. The learning algorithm updates the synaptic weights of simple units so that the information obtained from the output of the complex unit is increased. Simulation of the algorithm showed that it generates Gabor-wavelet-like weights similar to those observed in visual cortical neurons (simple cells). It also showed that, after the training, the responses of the complex unit are similar to those reported for a complex cell.