Towards Bridging the Gap Between Pattern Recognition and Symbolic Representation Within Neural Networks