Backfitting neural networks (Q1775949)

The paper presents a rather simple algorithm based on backfitting and cross validation for fitting perceptron models in regression and classification problems. This new algorithm is quite different from the usual backpropagation or the Levenberg-Marquardt (LM) algorithm and ensures a better convergence than the backpropagation. It also applies the resampling techniques to select an ideal network architecture (ideal number of hidden nodes) automatically using the training data itself and reduces the training time (compared to LM algorithm). Such resampling technique helps to avoid the problem of overfitting that one faces for usual perceptron learning algorithms without model selection scheme. The performance of the proposed algorithm is illustrated through case studies and simulation experiments which deal with regression (4 examples) and classification (8 examples).