When “Interference” Helps the Brain Remember Better
Memory is rarely a clean filing system. One experience resembles another; old associations intrude on new ones; the neural representation of a face, place or object inevitably shares features with other memories. In models of the brain, this interference has traditionally been treated as a problem: the more stored memories overlap, the more they interfere with each other, eventually making reliable recall impossible. But what if some of that interference is useful?
A new study by Marco Benedetti and Nicolas Brunel, both of Bocconi University’s Department of Computing Sciences, together with Enzo Marinari of Sapienza University of Rome’s Department of Physics and Ulises Pereira-Obilinovic of the Allen Institute’s Neural Dynamics program, turns a long-standing assumption about neural memory networks on its head.
Published in Physical Review Letters, the study shows that the tiny, unwanted traces of other memories that appear while a network is recalling one particular memory can collectively improve its performance. Rather than simply adding disruptive noise, these “spurious overlaps” can reduce the average input received by neurons, suppress unnecessary activity and ultimately allow the network to store more memories.
A crowd of tiny errors becomes something useful
The starting point is the class of models pioneered by physicist John Hopfield. In a Hopfield-type “attractor network”, a memory is not stored in one particular neuron. Instead, it is encoded in the pattern of connections across a population of neurons. When the network receives a partial cue, its collective dynamics can steer it toward the stored pattern that best matches that cue, rather as recognizing a few notes may bring an entire melody to mind. In the language of the model, the memory is an “attractor”: once the network gets sufficiently close to it, its activity tends to converge toward that state.
But a network contains many memories, and while it is retrieving one of them, its current state also has small overlaps with all the others. Those are the spurious overlaps. Traditional models have treated them essentially as interference: individually negligible, collectively noisy. The new work identifies a very different collective effect, that ensues when memories are stored in the brain not according to the Hopfield learning rule, but according to a more realistic one, inferred from in vivo neuroscience experiments.
Each spurious overlap is tiny: in a sufficiently large network, they individually disappear. But there are many of them. Brunel, Benedetti and colleagues show how added together, they noticeably lower the average synaptic input received across the network. Crucially, this reduction affects disproportionately neurons that should remain quiet during recall (background neurons), and “foreground” neurons — the ones that should be strongly active when a particular memory is retrieved. Lowering the overall input has only a modest effect on the foreground neurons because they are already receiving strong enough signals. But it can keep background neurons safely below their activation threshold. The result is a cleaner, sparser neural representation: fewer irrelevant neurons become active, and that in turn reduces noise in the network.
“This global suppression promotes sparser retrieval states, in which only a small fraction of neurons are strongly active.”
The numbers make the effect unusually tangible. In the researchers’ mean-field analysis, background neurons account for only 3–5% of neurons firing above half of the maximum rate when the spurious-overlap effect is included. When that effect is artificially removed, their share jumps to 20–25%. In other words, what looks like interference at the level of individual memories becomes a filtering mechanism at network level.
Why earlier models missed it
Classical Hopfield-style models generally use learning rules with a particular mathematical symmetry: averaged across stored memories, the postsynaptic component of learning balances out to zero. Under those conditions, spurious overlaps may make neuronal inputs more variable, but they do not systematically push their average level up or down.
The learning rule used in this study is different. It is based on earlier work that inferred properties of synaptic learning from recordings of neuronal activity in living animals. In that rule, the average postsynaptic effect is negative. That apparently technical difference changes the collective behavior of the network. Many small overlaps now add up to a reduction in average neural input.
The authors argue that this may point to something broader about biological memory systems:
“Our findings suggest that the design principles of biological networks may have evolved to harness the collective effects of spurious overlaps, providing a new perspective on the mechanisms underlying associative memory.”
The important point is not that biological brains have been shown to use exactly this mechanism. This is a theoretical model. But because its learning rule is motivated by neurobiological data, the result raises the possibility that properties which look inefficient in simplified models may serve useful functions in more realistic ones.
A biological design principle?
The broader implication is not simply that one particular neural-network model can store more patterns than expected. It is that learning rules inferred from biological data appear to exploit a collective effect that conventional models had effectively designed away. The paper goes so far as to suggest that these biologically derived rules sit close to an optimum for memory storage. Their negative postsynaptic bias produces enough suppression to keep background neurons quiet, without suppressing the neurons needed to represent the memory itself. Too little suppression loses the benefit; too much would begin to erase the signal.
That balance leads the authors to reinterpret spurious overlaps elegantly:
“This self-regulating mechanism highlights the importance of spurious overlaps as a homeostatic feature of the network, rather than purely a source of degradation.”
Efficient memory may not depend on eliminating every source of interference. It may instead depend on exploiting interference in the right way. What looks like noise at the level of an individual memory can, at the level of the whole network, become part of the mechanism that makes remembering possible.
