Learning Capability of Feedforward Networks
前馈神经网络的学习能力

A new grouping subnetwork use to improve Huang's significant result on the learning capability of two-hidden-layer feedforward networks (TLFNs) is presented. The result that feedforward netwroks with 2(2N)~((1/2)) 2 hidden neurons can learn any \$N\$ distinct samples with any arbitrarily small error be obtained. New method ensures the result is valid not only for the sigmoid function, but for a wide class of activation functions.