Learning control units for invariant recognition

Abstract A control unit is a group of synapses that are consistent with each other in terms of transformation parameters. We showed previously that these high-order interactions led to very fast dynamic link matching (DLM) for self-organization of mappings between two patterns, as needed in invariant recognition and many other visual tasks. We present here a model to learn from examples control units and the connection between them, by generalized Hebbian plasticity. Examples for learning are consistent mappings formed, for instance, by conventional DLM. Simulation results for 1D patterns showed that the learned results resemble those designed by hand, and thus were very encouraging.