Brain's Neural Drift May Use Sudden Jumps to Maintain Stable Perception

The brain faces a peculiar problem that neuroscientists have only recently begun to fully appreciate. When researchers record the activity of neurons over days or weeks, they find something unsettling: the neural patterns that represent a fixed object or experience keep changing. A neuron that fires when a mouse reaches a particular spot in a maze fires at a slightly different spot the next day. The visual cortex's response to the same image drifts measurably from one week to the next. This phenomenon, called representational drift, happens without any obvious change in behavior or learning. Yet somehow the brain maintains a stable, coherent perception of the world despite this constant neural reshuffling. How?

A team of researchers approached this puzzle by asking a deceptively simple question: what if the way neurons change matters as much as the fact that they change? They built a mathematical model of how a downstream neural circuit could track and correct for drift, using only the statistical patterns in neural activity itself—no external feedback, no knowledge of what the animal actually sees or does. The model revealed something striking. When neurons change their tuning gradually, like a dial turning slowly, small changes are easy to confuse with random noise. But when neurons make sudden jumps—switching from one preferred firing location to another in a single day—those jumps stand out clearly from the background noise. This distinction becomes critical as drift accelerates. The researchers found that sudden jumps, what they call heavy-tailed statistics, allow a downstream circuit to compensate for drift using fewer neurons or tolerating faster drift rates than would be possible with gradual changes alone.

To test whether the brain actually uses this strategy, the team analyzed two existing datasets of neural recordings from mice. The first came from the posterior parietal cortex, a region involved in navigation and reaching, where researchers had tracked individual neurons over two weeks as mice performed a virtual maze task. The second came from visual cortex, where neurons were recorded weekly over longer intervals as mice viewed natural scenes and oriented gratings. In both datasets, the researchers looked for the statistical signature of sudden versus gradual drift by examining how the correlations between pairs of neurons changed over time. In the posterior parietal cortex, where drift happens relatively quickly, they found that roughly 74 percent of tuning changes appeared to be sudden jumps rather than gradual shifts. In visual cortex, where drift is slower, the proportion dropped to about 58 percent, with more gradual changes mixed in. The difference was subtle but consistent: faster-drifting regions showed a stronger bias toward sudden changes.

The implications cut in multiple directions. For neuroscience, the finding suggests that representational drift may not be a bug in the brain's system but something closer to a feature—a consequence of learning and plasticity that the brain has learned to live with by structuring how individual neurons change. The sudden jumps could arise from the nonlinear properties of neurons themselves: as synaptic weights gradually shift, a neuron's inputs may suddenly cross a threshold, causing an abrupt change in what it responds to. Or neurons might suddenly recruit new inputs from upstream circuits. The researchers emphasize that their adaptive decoder is an idealized model, a best-case benchmark rather than a claim about how the brain actually implements this correction. Real neural circuits rely on known plasticity mechanisms like Hebbian learning and homeostatic plasticity, which are constrained by biology in ways their model is not. Still, their analysis provides a bound on what is theoretically possible and suggests that the brain's actual mechanisms operate somewhere within that bound.

The work also hints at a principle that could matter for artificial intelligence. Neural networks trained on streams of new data often suffer from catastrophic forgetting—they lose what they learned before when learning something new. The researchers suggest that if artificial networks were trained to make sudden, concentrated changes to their weights rather than distributed, gradual ones, they might avoid this problem more effectively. The brain's solution to drift, in other words, might offer a template for building more stable learning systems.

What remains unknown is substantial. The analysis drew on recordings from only two brain regions under different experimental conditions, making it impossible to say whether the pattern holds broadly. The researchers cannot yet explain the mechanistic link between gradual changes in synaptic weights—which is what learning rules typically produce—and the sudden jumps in neural tuning they observe. And the datasets they analyzed captured relatively small populations of neurons; a fuller picture would require simultaneous recordings of both drifting populations and the downstream circuits that read from them. But the work establishes a new lens for thinking about drift: not as a problem the brain solves despite, but as a phenomenon whose statistical structure the brain may have evolved to exploit.