A natural object recognition system using Self-organizing Translation-Invariant Maps

Translation-invariant representations are a key aspect of any object categorization system. In most neural realizations, these representations are either non-adaptive and often non-optimal, or adaptive and redundant. Especially in the case of natural images containing complex features, the neural code becomes highly inefficient. In this paper, we introduce an adaptive network model which overcomes a part of this redundancy by searching for translational invariance. The model implements a Self-organizing Translation-Invariant Map (STIM) using an unsupervised competitive learning rule. The latter employs generalized weightsharing in combination with a symmetric, symmetry-breaking filter. A Kohonen-neighborhood update scheme ensures a proper covering of the feature space.