Incremental character recognition with feature attribution

Abstract

The neural network learning algorithm presented in the paper splits the problem of handwritten digit recognition into easy steps by learning character classes incrementally: at each step, the neurons most relevant to the considered class are fixed so that subsequent classes will not disrupt the knowledge already acquired, but will be able to use it. A new relevance measure is also defined, for which a cheap approximation can be computed. The advantage of the attribution scheme starts to show even for small experiments, but should become more obvious as the number of classes increases. Picking only a few relevant features for each class, and sharing them between classes, constrains learning and improves generalization. Experiments were limited to pre-segmented digits, but our use of a spatio-temporal network architecture makes their extension to unsegmented strings straightforward.