Hierarchical Concept Formation in Associative Memory Models and its Application to Memory of Motions for Humanoid Robots

Abstract

In associative neural networks with nonmonotonic activation functions which store hierarchically correlated patterns, bifurcations of attractors take place depending on the parameter of nonmonotonicity. With hierarchically correlated storage patterns, attractors are the stored patterns when the nonmonotonicity is large, and new emergent patterns at around the centers of clusters when the nonmonotonicity is small. The phenomenon itself was shown by the authors (2005) by simulations and applied to memory systems for humanoid robots, which store feature vectors of motion patterns and maintain specific and emergent conceptual representations of motions of humanoid robots. In this paper, we theoretically describe the hierarchical bifurcations of attractors in nonmonotonic associative memory models and discuss the correspondences between the theory and simulation results