Summary: A collaborative research effort has introduced Centered Daydreaming, a refined learning rule for Hopfield networks that combines simultaneous daytime learning with continual pruning. By shifting the network’s focus from absolute pixel values to deviations from a running average, the team pushed storage capacity to its theoretical limit—one full memory per neuron (100%)—even for highly biased, realistic datasets.
Key Facts
- The 13% storage bottleneck overcome: A classical Hopfield network typically stores only about 13% as many memories as neurons. Pushing beyond that threshold produces spurious attractors—false memories that interfere with retrieval.
- 100% theoretical capacity attained: The original Daydreaming framework (2025) merged learning and cleaning phases so the network could simultaneously acquire valid memories while eliminating spurious ones, enabling storage up to the theoretical maximum of one memory per neuron.
- Robustness to realistic biased data: While Daydreaming performed well on balanced laboratory datasets, it struggled with real-world, strongly biased inputs (for example, overexposed or very dark images where one pixel state dominates).
- Biologically plausible local learning: Previous fixes for data bias relied on global computations across the entire network, which are implausible biologically. Centered Daydreaming preserves local update rules that align with how biological neurons operate, communicating mainly with neighbors and using limited local information.
- Focus on averages and differences: Instead of using raw pixel values, Centered Daydreaming centers inputs by subtracting a running average, making the network sensitive to meaningful deviations rather than overwhelming background commonalities.
- Durable retrieval under strain: Extensive tests with strongly biased, realistic datasets show that Centered Daydreaming maintains accurate pattern retrieval, suggesting a path toward interpretable, energy-efficient AI systems inspired by statistical physics.
Source: SISSA
During waking hours the brain encodes new experiences; during sleep it consolidates important memories and prunes irrelevant activity. That biological pattern inspired improvements in Hopfield networks, a well-known class of associative-memory models.
In 2025, Federico Ricci-Tersenghi and collaborators proposed the Daydreaming algorithm, which simultaneously learns new patterns and suppresses spurious attractors, vastly improving capacity. The algorithm’s success, however, was initially limited to balanced datasets where pattern statistics are symmetric.
A new study, published in the Journal of Statistical Mechanics: Theory and Experiment (JSTAT), extends that approach. Ricci-Tersenghi and colleagues in Japan present Centered Daydreaming, a local and bias-tolerant variant that reliably handles strongly skewed, real-world patterns.
A classical associative memory model
John Hopfield’s networks, introduced in 1982, model associative memory through interconnected artificial neurons. They are able to retrieve stored patterns from partial or noisy cues: present a degraded image of a tree, and the network converges to the stored “tree” pattern.
However, the simplest Hopfield model has a severe capacity limit—roughly 0.13 times the number of neurons. Beyond that, the dynamics produce spurious attractors: hybrid configurations that do not correspond to any true memory and can trap the network in erroneous recall.
Daydreaming: concurrent learning and pruning
To reduce spurious attractors, researchers developed dreaming methods inspired by sleep: after training, the network is allowed to explore its states and remove false attractors. But prolonged offline cleaning risks catastrophic forgetting—valid memories can be erased.
Daydreaming avoids that trade-off by interleaving learning and cleanup continuously: the network reinforces correct patterns while dissolving spurious states. This concurrent strategy raised capacity toward the theoretical maximum, effectively allowing one stored memory per neuron.
Yet Daydreaming in its original form assumes unbiased patterns and breaks down when many inputs share the same value, as in heavily bright or dark images where most pixels are identical across examples.
Centered Daydreaming: emphasize differences, preserve locality
Global corrections used by earlier approaches are biologically unrealistic because neurons operate on local signals. Centered Daydreaming addresses this by re-centering inputs: instead of processing absolute pixel states, each neuron updates based on the deviation of its input from a local moving average. This mirrors how biological systems often detect salient features by comparing to an internal baseline (for example, recognizing faces by deviations from an average face).
By focusing learning on differences rather than shared background content, Centered Daydreaming restores discriminative power even when datasets are heavily biased. Importantly, it retains local learning rules, making the method both biologically plausible and computationally efficient.
Empirical results in the study show that centering dramatically enlarges basins of attraction compared with centered pseudo-inverse alternatives. Analysis of coupling-matrix eigenvalue spectra suggests that the broader spectrum produced by Daydreaming-style updates contributes to larger, more robust attraction basins.
Understanding how simple, brain-inspired models selectively emphasize relevant deviations can guide the design of AI systems that are more transparent, robust to real-world data bias, and energy efficient.
Key Questions Answered:
A: Spurious memories are unintended attractors that mix elements of true patterns into distorted hybrids. They clutter the network’s energy landscape so that a noisy input can be pulled toward a hybrid state, producing confident but incorrect recalls—analogous to hallucinations in the model.
A: Traditional post-training cleaning can overcorrect and eliminate valid memories. Daydreaming runs learning and cleaning concurrently, reinforcing true patterns while dissolving spurious attractors as they arise, which prevents catastrophic forgetting and permits much higher storage efficiency.
A: Standard Hopfield rules assume balanced statistics. When most pixels share the same value, shared background dominates learning and obscures the subtle differences that distinguish memories. Centering subtracts that common baseline so the network prioritizes informative deviations.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full.
- Additional editorial context was added by staff.
About this AI memory research news
Author: Federica Sgorbissa
Source: SISSA
Contact: Federica Sgorbissa – SISSA
Image: The image is credited to Neuroscience News
Original Research: Open access.
Paper: “Daydreaming algorithm for Biased Patterns” by Mikiya Doi, Masayuki Ohzeki and Federico Ricci-Tersenghi.
Journal: Journal of Statistical Mechanics: Theory and Experiment.
DOI: 10.1088/1742-5468/ae8249
Abstract
Daydreaming algorithm for Biased Patterns
The Daydreaming algorithm, introduced in Serricchio et al. (2025, Neural Networks 186 107216), simultaneously reinforces stored patterns and suppresses spurious attractors to boost Hopfield-model storage capacity. While effective for both uncorrelated and correlated data, the original formulation assumed largely unbiased patterns. Biased patterns—where one state dominates—pose greater challenges for associative-memory models.
This work reformulates Daydreaming for biased patterns by adopting a centered representation motivated by the pseudo-inverse rule. The authors define retrieval dynamics and an energy function in the centered space, derive the corresponding update rule for Centered Daydreaming, and compare it to a centered pseudo-inverse rule by examining retrieval maps and eigenvalue spectra of coupling matrices.
Results show that Centered Daydreaming produces larger basins of attraction than the centered pseudo-inverse approach. The improvement appears linked to the broader eigenvalue spectrum generated by the Daydreaming updates; intermediate coupling matrices interpolating between pseudo-inverse-like and Daydreaming spectra support this interpretation. Overall, a broader spectrum correlates with enlarged attraction basins and more robust retrieval under biased conditions.