hide
Free keywords:
-
Abstract:
A model for position invariant pattern recognition is presented. Although not demonstrated here, the system is insensitive to distortions. Recognition is based on labeled graph matching. The system consists of two layers of neurons, an input layer, and a memory and recognition layer. The latter consists of subnets to represent individual patterns. In both layers, patterns are represented by labeled graphs. Nodes are “neurons,” labels are local feature types, links are implemented by excitatory connections and represent topology. Recognition is driven by spontaneous dynamic activations of local clusters in the input layer. Network dynamics is able to selectively activate with good reliability corresponding clusters in memory layer. Few cluster activations suffice to identify the subnet and pattern corresponding to the graph in the input layer. The system has been implemented and tested with the help of simulations.
