No abstract

Image movements relative to the retina are essential for the visual perception of stationary objects during fixation. Here we have measured fixational eye and head movements of the turtle, and determined their effects on the activity of retinal ganglion cells by simulating the movements on the isolated retina. We show that ganglion cells respond mainly to components of periodic eye movement that have amplitudes of roughly the diameter of a photoreceptor. Drift or small head movements have little effect. Driven cells that are located along contrast borders are synchronized, which reliably signals a preceding movement. In an artificial neural network, the estimation of spatial frequencies for various square wave gratings improves when timelocked to this synchronization. This could potentially improve stimulus feature estimation by the brain.

We consider a Perceptron with N i input units, one output and a yet unspecified number of hidden units. This Perceptron must be able to learn a given but arbitrary set of input-output examples. By sequential learning we mean that groups of patterns, pertaining to the same class, are sequentially separated from the rest by successively adding hidden units until the remaining patterns are all in the same class. We prove that the internal representations obtained by such procedures are linearly separable. Preliminary numerical tests of an algorithm implementing these ideas are presented and compare favourably with results of other growth algorithms.

We show that the generalization ability of simple Perceptron-like devices is strongly enhanced by allowing the network itself to select the training examples. Analytic and numerical results are obtained for the Hebb and for the optimal Perceptron learning rule, respectively.

No abstract