Abstract: We present a neural system for the recognition of objects from realistic images, together with results of tests of face recognition from a large gallery. The system is inherently invariant with respect to shift, and is robust against many other variations, most notably rotation in depth and deformation. The system is based on Dynamic Link Matching. It consists of an image domain and a model domain, which we tentatively identify with primary visual cortex and infero-temporal cortex. Both domains have the form of neural sheets of hypercolumns, which are composed of simple feature detectors (modeled as Gabor-based wavelets). Each object is represented in memory by a separate model sheet, that is, a two-dimensional array of features. The match of the image to the models is performed by network self-organization, in which rapid reversible synaptic plasticity of the connections ("dynamic links") between the two domains is controlled by signal correlations, which are shaped by fixed inter-columnar connections and by the dynamic links themselves. The system requires very little genetic or learned structure, relying essentially on the rules of rapid synaptic plasticity and the a priori constraint of preservation of topography to find matches. This constraint is encoded within the neural sheets with the help of lateral connections, which are excitatory over short range and inhibitory over long range.
Keywords: neural networks, dynamic link matching, face recognition, translation invariance, window of attention