BACKPROPAGATION USES PRIOR INFORMATION EFFICIENTLY

The ability of neural net classifiers to deal with a priori informatio n is investigated. For this purpose, back-propagation classifiers are trained with data from known distributions with variable a priori prob abilities, and their performance on separate test sets is evaluated. I t is found that back-propagation employs a priori information in a sli ghtly suboptimal fashion, but that this does not have serious conseque nces on the performance of this classifier. Furthermore, it is found t hat the inferior generalization that results when an excessive number of network parameters are used can (partially) be ascribed to this sub optimality.