Computing the probability density in connectionist regression

We introduce a non-parametric method for determining the degree of uncertainty in prediction and show its use in a regression problem. We designed and tested a neural network which performs a prediction and gives a measure of the precision of the prediction. Outputs consist of a set of normalized exponential units which produce an estimate of the continuous probability density function p(ylx). The variance of this distribution can be interpreted as the uncertainty in the classification. The distribution can also be used for calculations of percentiles and higher order moments. >