Learning strategy for the binary perceptron

Abstract

Pursuing the work of Penney and Sherrington, we determine the optimal continuous-weight perceptron which, on clipping, correctly predicts the largest number of weights for the binary perceptron with maximum stability. We calculate the fraction of correctly predicted binary weights when only the continuous weights stronger than a certain threshold are clipped. We finally carry out simulations for a perceptron with 50 weights to test the practicability of different learning strategies.