Using sparse trace neural network to extract temporal-spatial invariance

The relationship between invariance extraction and temporal processing in the visual system is far from being fully explored. A temporal-spatial invariance extraction model is proposed from the point of view of neuro-coding. Temporal self-adaptation is introduced into a neural invariance extractor to code some underlying physical parameters of input patterns and to learn traces of given classes of patterns. A trace is defined as the weighted temporal average over the output of the invariance extractor which leads to a simple trace neural network (TNN). A sparse trace neural network (STNN) is then proposed to produce pattern representations which can be easily post-processed. The behaviour of the STNN can be explained as forming some interesting temporal-spatial order structures and is compared qualitatively with some other models. Simulation results show that STNN achieves better performance than several previously proposed models