ON THE GENERALIZATION ABILITY OF DILUTED PERCEPTRONS

A linearly separable Boolean function is learned by a diluted perceptr on with optimal stability. A different level of dilution is allowed fo r teacher and student perceptron. The learning algorithms used were th e optimal annealed dilution and Hebbian dilution. The generalization a bility, i.e. the probability to recognize a pattern which has not been learned before, is calculated in replica symmetry.