pipette
ESEspañol

Emergence of diffusion models from associative memory through memorization and generalization

Bao Pham, Gabriel Raya, Matteo Negri, Mohammed Javeed Zaki, Luca Ambrogioni, Dmitry Krotov

Peer-reviewed journal

In the authors' words

Abstract Dense Associative Memories (DenseAMs) are generalizations of Hopfield networks that store training data at local minima of the energy landscape. When the amount of training data exceeds the critical memory storage capacity, new local minima (different from the training data) emerge. These emergent local minima, known as spurious states, can hinder memory retrieval. We examine diffusion models (DMs) through the DenseAM lens, viewing their generative process as an attempt of a memory retrieval. In the small data regime, DMs create distinct attractors for each training sample, akin to DenseAMs below the critical capacity. As the training set size increases, DMs transition to generalization. We identify a critical intermediate phase, predicted by DenseAM theory, characterized by the emergence of spurious states. In generative modeling, these previously overlooked states are no longer negative artifacts but rather are the first signs of generative capabilities. Their existence is demonstrated across a wide range of architectures and datasets.

Main resultThe abstract does not state a limitation.

In plain words AI summary

Dense Associative Memories store training data as energy landscape valleys. When training data exceeds capacity, new valleys appear that differ from training data. These extra valleys, called spurious states, can block memory retrieval. The authors study diffusion models as memory retrieval attempts. With small data, diffusion models make separate attractors for each training sample. As data grows, diffusion models shift to generalization. A middle phase emerges where spurious states appear, as theory predicts. In generative modeling, these states are not flaws but first signs of generation ability. The authors show these states exist across many architectures and datasets.

Written by Kimi K2.5 using only the abstract above. Jev checked that every sentence is supported by the abstract sentences it cites, and removed any that were not. Point at a sentence to highlight its source.

Appeared: Tuesday, October 27. Nature Communications. Peer-reviewed journal.

DOI: 10.1038/s41467-026-78727-y