Explaining student grades predicted by a neural network

We have trained a backpropagation trained feedforward neural network to predict student performance in a large undergraduate computer science subject at the University of New South Wales. The prediction uses partial grades from during the teaching session to predict the final grade. The exam mark which is the major component (60%) of the overall grade is not used. The purpose of this network is to allow students to predict the final grade they are likely to achieve based on current performance, and obviously to improve their performance if the predicted grade is below their expectations. By itself, however, the network is not adequate as it provides no feedback as to why their performance merits a particular grade. We therefore generate an explanation of the conclusion reached by the neural network for predicting particular student grades.