Bayesian input selection for neural network classifiers

In this paper we discuss the use of the Bayesian posterior probability distribution over weight space and Receiver Operating Characteristic curves in a neural network input selection algorithm. The posterior distribution is obtained by combining the likelihood function based on training data and a prior distribution based on expert knowledge. To numerically calculate the marginalisation, Markov Chain Monte Carlo methods are used. We demonstrate the technique on the problem of ovarian cancer classification. The resulting input selection is then used to train a neural network that significantly outperforms the Risk of Malignancy Index, a traditionally used diagnostic aid.