Noise robustness in the perceptron

摘要

Within the framework of online supervised learning, algorithms which lead to optimal generalization ability in the boolean per-ceptron are obtained. Restrictions on the available information during learning lead to different optimal algorithms. Knowledge of the noise level is required for optimal performance to be achieved, and misestimation of that quantity may lead to partial or complete loss of the generalization ability. Results are shown in terms of robustness phase diagrams.