Biologically Inspired Algorithm Based On Error Minimization

摘要

This work presents a model for unsupervised neural network adaptation which generates a local error provided by excitatory and inhibitory interaction between the outputs of neighboring neural structures. When applied to images, this algorithm yields an efficient code for natural images and the emergence of simple-cell like receptive fields. In its development we make use of second order statistics only and by spatial interactions we extract information beyond correlation. Therefore, we conclude that the excitatory and inhibitory local interactions in distributed systems are sufficient to sparsify the L2 norm solution, which normally requires higher order statistics.