Avoiding overfitting caused by noise using a uniform training mode

In the continuous function approximations using multilayer feedforward neural networks, there is the problem of overfitting if the training data are corrupted by noise. This paper presents a practical approach to conduct the learning of the neural networks to be uniform for each sample with an individual error tolerated by the noise level of the training data, rather than conventionally considering total error of all the training samples. The experimental results using the joint inversion of seismic and well data in geophysical exploration as illustrating examples show that the uniform training mode can effectively avoid overfitting phenomenon. In addition, as by-pass products, it can greatly reduce the training computation and has potential for recognizing and removing the outliers from the training data.