Selecting examples for perceptrons

Abstract

A technique is introduced for analysing the learning process of perceptrons which continually select their own examples (the most efficient training algorithm yet devised). The authors predict a 42% reduction in the number of examples 'wasted' in training an Ising perceptron, compared to the case in which examples are random. Optimally stable spherical perceptrons may also be taught significantly more efficiently, and their results compare well with an existing numerical simulation.