Learning translation invariance by self-supervised neural networks

We discuss the critical factors of translation invariant learning within the framework of competitive learning. Even if we have well preprocessed input feature patterns which make the problem linearly separable, there remains to be handled the conflict between the sparse nature of the input patterns and correlation-based learning. Motivated by a batch-mode solution of the problem we introduce some local mechanisms to avoid this conflict. A self-supervised learning scheme is proposed which uses the local activity preservation feedback threshold as a reinforcement signal and weight adaptation in a positionally local manner. This scheme can handle different object classes with overlapping features, what makes it more general when compared to the recently proposed Modified Hebbian learning for the translation invariant classification of four lines of different orientation.