Associative mapping network without redundant local minima

Abstract

Abstract When neural networks are used for associative memory or associative mapping, there always exist in their corresponding energy functions unexpected local minima to which networks may converge. These unexpected local minima, corresponding to spurious outputs or crosstalks, are caused by improperly distributed stored patterns, especially those that are not mutually orthogonal or not linearly independent. An associative mapping network, named the correlation network, is thus proposed to solve this problem. This non-linear network directly accesses its stored patterns and does not require high-order dimension expansion, by introducing the thresholds which carry the correlating information among stored patterns. Furthermore, correlation networks would never have redundant local minima in their energy functions for any pattern stored in the network.