作者: Manfred Opper , David Haussler
DOI: 10.1103/PHYSREVLETT.66.2677
关键词:
摘要: The generalization error of the Bayes optimal classification algorithm when learning a perceptron from noise-free random training examples is calculated exactly using methods statistical mechanics. It shown that if an assumption replica symmetry made, then, in thermodynamic limit, less than canonical stochastic algorithm, by factor approaching \ensuremath{\surd}2 as ratio number to weights grows. In addition, it approximations can be achieved algorithms use two-layer neutral net learn perceptron.