Improving speech recognition learning through lazy training

Multilayer backpropagation, like most learning algorithms that can create complex decision surfaces, is prone to overfitting. We present a novel approach, called lazy training, for reducing the overfit in multiple-layer networks. Lazy training consistently reduces generalization error of optimized neural networks by more than half on a large OCR dataset and on several real world problems from the UCI machine learning database repository. Here, lazy training is shown to be effective in a multilayered adaptive learning system, reducing the error of an optimized backpropagation network in a speech recognition system by 50.0% on the TIDIGITS corpus.