A method to estimate prediction intervals for artificial neural networks that is sensitive to the noise distribution in the outputs

The use of confidence estimation techniques on neural networks outputs plays an important role when these mathematical models are applied in many practical applications. However, few of these techniques have the capability to deal with variable noise rate in the predictions over the domain, making the assumptions about the reliability of these outputs become not suitable with their real accuracy. In this paper an extension to the non-linear regression method to estimate prediction intervals for feed forward neural networks is presented. The main idea of this method is that residuals variance should be estimated in function of the input data and not as a constant. Thus, using clustering techniques, distinct estimates of the residuals variance are made and then used to obtain new prediction intervals. Proceeding in this manner, the experiments results show that this approach can lead to prediction intervals that better reflect the confidence level of the neural network outputs.