Error estimation and error bounds for neural networks
Hualou Liang, Guiliang Dai · 2002
A method is proposed to estimate the standard error of predicted values in multilayer perceptron (MLP). It is based on likelihood theory. It holds for all feedforward networks, irrespective of the topology or the specific task at hand. In addition, the bounds on a neural network with perturbed weights and inputs is analytically derived. The bounds obtained are applicable to both digital and analog network implementations. By computer simulation, the validity of the proposed methods has been illustrated.