Within the context of a particular statistical discrimination task, we make a quantitative comparison between the performance of a feed-forward neural network and the information-theoretic optimal performance. We also address the ability of such networks to generalize and the effect of network architecture on performance.