A Voronoi-based energy function for stable associative memory in modern hopfield networks

Abstract

The advent of the Modern Hopfield Network (MHN) has enabled associative memory models to handle continuous-valued data. Although MHNs exhibit high memory capacity, their recall performance tends to be unstable depending on the distribution of stored patterns. To address this issue, we propose a novel energy function based on Voronoi partitioning that enables stable memory retrieval independent of the configuration of stored patterns. Experimental results demonstrate that the proposed method achieves higher recall accuracy across a wide range of pattern sets compared with conventional MHNs.