GENERALIZATION IN THE PROGRAMMED TEACHING OF A PERCEPTRON

According to a widely used model of learning and generalization in neu ral networks, a single neuron (perceptron) can learn from examples to imitate another neuron, called the teacher perceptron. We introduce a variant of this model in which examples within a layer of thickness 2Y around the decision surface are excluded from teaching. That restrict ion transmits global information about the teacher's rule. Therefore f or a given number p = alpha N of presented examples (i.e., those outsi de of the layer) the generalization performance obtained by Boltzmanni an learning is improved by setting Y to an optimum value Y-0(alpha), w hich diverges for alpha --> 0 and remains nonzero while alpha < alpha( c) approximate to 5.7. That suggests programed learning: easy examples should be taught first.