Evaluation of the log-likelihood cost function for learning in feed-forward networks

The minimisation of the log-likelihood function of (E_L) corresponds to maximising the conditional likelihood of a training set with independent binary coded outputs. Applying back-propagation to the minimisation of E_L requires only trivial modifications to the standard formulae, and can improve the stability and learning speeds. The paper reports an investigation into whether, and to what extent, the generalisation of a fixed sized net can be improved by adopting this cost function