Daydreaming Clears an AI’s False Memories, Even With Lopsided Data

A Hopfield network is one of the simplest machines that can remember: a few hundred artificial neurons wired to one another, each holding one of two values. Train one on patterns in which four entries out of five carry the same value, hand it pure noise, and let it settle. The state it lands on agrees with the pattern you wanted about 60 percent of the way. That looks like partial recovery, but every neuron has simply flipped to the majority value, and 60 percent is precisely the score you get when the wiring holds no usable information at all.

False Memories Take Up Most of the Room

That empty 60 percent is the shape of a problem that has trailed these networks since John Hopfield described them in 1982, work that shared the 2024 Nobel Prize in Physics. Storing a memory means setting the connection strengths so that the pattern becomes a resting state. Hand the network a corrupted copy, let every neuron look at what its neighbors are doing and take the sign of their weighted sum, and the whole configuration slides downhill until it stops moving.

The set of starting points that slide into a given memory is called its basin of attraction, and the width of that basin, not merely whether the memory sits still once you hand it over intact, is what decides whether the network is any use.

The classic wiring rule fills up fast. Past roughly 14 percent of the neuron count, retrieval errors run away, and the leftover space fills with spurious attractors: stable states that blend stored memories into configurations the network was never taught. Federico Ricci-Tersenghi, a theoretical physicist at Sapienza University of Rome and one of the authors, calls them “false memories, attractors of the dynamics that do not correspond to any real memory.”

The standard countermeasure borrows from sleep: let a trained network run from random starting points, weaken whatever it falls into, and the invented states should dissolve. Push that unlearning too long and the real memories go with them, a failure known as catastrophic forgetting. The biological version is better behaved, since sleep reactivates and stabilizes what was encoded during the day rather than only erasing. The Daydreaming algorithm, published in 2025 by Ricci-Tersenghi and colleagues, folded the two operations into one, strengthening a stored pattern and weakening a spurious state at every step, and drove capacity toward one memory per neuron. It also assumed balanced patterns.

Bias is not an exotic condition. Overexposed photographs, mostly dark images, and neural codes in which only a small fraction of cells respond to any given stimulus all produce patterns where one value dominates. When most entries agree, the stored patterns come to resemble each other, and the interference between them stops averaging out to nothing. It acquires a mean, and that mean shoves every neuron the same direction regardless of which memory is being retrieved.

Working Only on What Changes

The repair is to stop storing the patterns and start storing their departures from the average. For each neuron, the researchers compute the value it typically takes across the whole set of stored patterns, then subtract that value before the connection strengths are built. What gets written into the wiring is no longer “this neuron was on” but “this neuron was on when it usually is not.” During recall, the same subtraction runs in reverse: each neuron takes the weighted sum of how far its neighbors currently sit from their own averages, adds its own average back in, and takes the sign.

Ricci-Tersenghi uses faces to explain the point. If every photograph is a close-up against a similar background, most pixels are near-identical everywhere, and that shared information swamps the learning. “If, instead, we work only on what changes relative to the average face, the differences emerge clearly,” he says. Where the arithmetic happens matters just as much. The pseudo-inverse rule, the standard analytical alternative, requires inverting a matrix built from every stored pattern at once, an operation no synapse could perform. The centered update needs only the two neurons at either end of a connection and their running averages. “It is much more realistic for each decision to be made locally,” Ricci-Tersenghi says.

The Plateau Held at Every Bias Tested

The comparison ran on networks of 500 neurons holding 200 patterns, twenty independent networks per setting, with the bias dialed from balanced up to 80 percent. The pseudo-inverse rule degraded steadily. At balance it recovered patterns from starting points already about three-quarters correct; by 80 percent bias it needed roughly 90 percent agreement before it could finish the job, which leaves no room for noise. Centered Daydreaming reached essentially perfect recall from starting points around 70 percent correct, and hit that same threshold at every bias strength studied. The uncentered original, run on the same biased data, had not formed a plateau at all after 512 training epochs.

Why the broader basins appear is a separate question, and the researchers went after it directly. The connection matrix built by Daydreaming has a distinctly different fingerprint from the pseudo-inverse one: its positive eigenvalues spread across a wide range, with the largest more than twice the pseudo-inverse maximum, instead of clustering at a single value. So they took the trained matrix apart, held its eigenvectors fixed, and slid the leading eigenvalues toward their common average by a controlled amount. Halfway there, the recall plateau was unchanged. All the way, with the leading modes weighted equally as in the pseudo-inverse case, the plateau clearly shrank. The spread itself, not just which directions the network favors, is doing the work.

Structured Data Did Not Fully Cooperate

None of this was tested on photographs. The patterns are synthetic binary strings, each entry independently taking the majority value with a set probability, and the images in the explanation are analogy rather than data. The closest the paper comes to realistic structure is an appendix, where patterns are built from a small set of shared underlying features plus a deliberate global bias. Centered Daydreaming still formed a recall plateau there. But retrieval came out slightly worse after 2,048 and 4,096 training epochs than it had been earlier, so performance does not simply climb and settle. And the network’s overlap with the underlying features never approached one. The authors write that the mechanism in correlated biased data is not yet understood.

What the result does establish is narrower and firmer. A learning rule that updates using only what is available at a single connection can hold its ground against a data property that wrecks the analytical alternative. That’s why people keep returning to these models: backpropagation, the workhorse of modern deep learning, needs information to travel long distances through a network in a way brains appear not to manage, and local rules are the leading candidates for what they do instead. Ricci-Tersenghi suggests that understanding how simple brain-inspired models separate what matters from what is shared could eventually feed into AI that is easier to interpret and cheaper to run.

The loose thread sits in that appendix. When the researchers gave the biased structured model a special feature vector set to plus one everywhere, the quantity tracking how well the network aligned with it stayed flat near 0.5 no matter where the dynamics started. The network could find its memories. It could not find the thing they had in common.

Frequently Asked Questions