Consistent inference of probabilities in layered networks: Predictions and generalization

The problem of learning a general input-output relation using a layered neural network is discussed in a statistical framework. By imposing the consistency condition that the error minimization be equivalent to a likelihood maximization for training the network, the authors arrive at a Gibbs distribution on a canonical ensemble of networks with the same architecture. This statistical description enables them to evaluate the probability of a correct prediction of an independent example, after training the network on a given training set. The prediction probability is highly correlated with the generalization ability of the network, as measured outside the training set. This suggests a general and practical criterion for training layered networks by minimizing prediction errors. The authors demonstrate the utility of this criterion for selecting the optimal architecture in the continuity problem. As a theoretical application of the statistical formalism, they discuss the question of learning curves and estimate the sufficient training size needed for correct generalization, in a simple example.