Abstract

A three-layer nonrecurrent network of simple nonlinear elements can learn to compute the slant of a set of uniformly textured planar surfaces. During the learning phase, input training images are processed with Gabor filters as present in early mammalian vision and these filter responses are presented simultaneous with the known output slant values. Connection weights are adjusted to minimize error between the computed slant and the actual slant using an iterative gradient descent learning algorithm. After convergence, the operational phase begins, in which the network is used with the weights fixed at values determined during learning to compute the slant of input images. The network is shown to achieve an excellent correlation with the actual slant of the surface. The weights between the first and second layer units of the network can be interpreted biologically as receptive fields. The qualitative characteristics of some of the fields observed are oriented diagonal in nature, similar to those seen in Visual Area 17, but in a novel multidimensional space consisting of spatial frequency, orientation, and image location.

© 1989 Optical Society of America