Temporal representation of space for adaptation to distortions among 2 dimensions

Abstract

Connectionist systems for spatio-temporal pattern recognition usually use a spatial coding of time or a mixed, spatial and temporal, representation. In this paper, we propose a temporal coding of spatial information by means of relative temporal delays of spatial events. These spatial events, translated into temporal events by scanning the 2D-space, feed a guided propagation network for handwritten records recognition. The network can adapt to (i) global pattern translation, (ii) local spatial shift of spatial events around their original position, (iii) suppression of some spatial events, and (iv) global pattern transformation.