Abstract

L. T. Maloney has described a method of calibrating a visual system consisting of input units and output units that are initially imperfectly connected. The information available is the known translation and the output images. The desired condition forcing adjustment of the connecting weights is that the output images be related by the translation. We have simulated three algorithms. Two of them resemble the Widrow-Hoff algorithm except that one of the output images is translated and used as the reinforcer for the other. The third is a combination of these two and includes the gradient descent solution for a single motion presentation. The translation condition is not affected by a space-invariant linear filter, so we force a unique solution by fixing the receptive field of one unit. For appropriate image subspaces and positioning and motion paradigms, all three methods can correctly train the weights for connecting an irregular array to a regular array, independent of the initial weights.

© 1989 Optical Society of America