A net with feedback for recognising deformed patterns

Describes a neural net which has been used to classify hand printed characters. It is shown that conventional feed-forward nets and discriminant function classifiers are successful where the classes form distinct clusters in feature space, but that this is difficult to achieve for unconstrained hand-printed characters. By training a net with perfect exemplars of each class it is possible to achieve the required separation. These are then generalised by a process of feedback to variable gain binary feature detectors. The net is called NIFTI (net incorporating feedback to inputs)