Dense Associative Memory with Epanechnikov Energy

  • 2025-06-12 16:16:31
  • Benjamin Hoover, Zhaoyang Shi, Krishnakumar Balasubramanian, Dmitry Krotov, Parikshit Ram
  • 0

Abstract

We propose a novel energy function for Dense Associative Memory (DenseAM)networks, the log-sum-ReLU (LSR), inspired by optimal kernel densityestimation. Unlike the common log-sum-exponential (LSE) function, LSR is basedon the Epanechnikov kernel and enables exact memory retrieval with exponentialcapacity without requiring exponential separation functions. Moreover, itintroduces abundant additional \emph{emergent} local minima while preservingperfect pattern recovery -- a characteristic previously unseen in DenseAMliterature. Empirical results show that LSR energy has significantly more localminima (memories) that have comparable log-likelihood to LSE-based models.Analysis of LSR's emergent memories on image datasets reveals a degree ofcreativity and novelty, hinting at this method's potential for both large-scalememory storage and generative tasks.

 

Quick Read (beta)

loading the full paper ...