Unlearning algorithm in associative memory

摘要

We incorporate into a synthesis procedure for a class ofndiscrete-time neural networks an unlearning capability. The proposedntechnique increases storage capacity while maximizing the domain ofnattraction of each desired pattern to be stored. Making use of learningnand forgetting capabilities, neural networks generated by the methodnadvanced herein are capable of learning new patterns as well asnforgetting learned patterns without the necessity of recomputing thenentire interconnection weights and external inputs. The unlearningnalgorithm developed is then utilized off-line to equalize the basins ofnattraction for each desired pattern to be stored, and to minimize thennumber of spurious states. Specific examples are given to illustrate thenstrengths and weaknesses of the methodology advocated herein