Exact Solution for On-Line Learning in Multilayer Neural Networks

We present an analytic solution to the problem of on-line gradient-descent learning for two-layer neural networks with an arbitrary number of hidden units in both teacher and student networks.