A probabilistic model for evaluation of neural network classifiers