Interacting Adaptive Processes with Different Timescales Underlie Short-Term Motor Learning

You practice a movement, you get better, and then someone takes away all feedback. Your body spontaneously drifts back toward the old error all by itself, with no new information arriving. That's not a malfunction. It's the signature of two separate learning systems running at different speeds inside you. Until Smith, Ghazizadeh, and Shadmehr published this work, nobody had a clean explanation for why it happens. Motor adaptation is the brain's routine recalibration of movement in response to consistent errors. In the lab, researchers typically study it by introducing a perturbation — for example, a robotic arm that pushes your reach sideways every single time — and watching the brain gradually correct for it. Because the task is precise and repeatable, motor adaptation has become one of the best windows we have into how learning actually works, trial by trial. But the field had a problem. Researchers had catalogued several behavioral phenomena that a simple learning model couldn't explain all at once. The first is savings: if you learn a perturbation, wash out that learning until behavior looks normal again, and then relearn the same perturbation, you relearn it faster the second time. Something survives the apparent washout. The second is anterograde interference: learning one perturbation slows down your ability to learn the opposite perturbation next. The third is spontaneous recovery: after you've extinguished an adaptation, the original learned pattern can drift back on its own, with no new errors driving it. There are also rapid unlearning and rapid downscaling — the ability to reverse or shrink a learned adaptation surprisingly quickly. The standard model — a single adaptive state with one time constant, where each trial's update is simply a retention factor times the previous state plus a learning rate times the current error — cannot produce all of this simultaneously. A single process predicts a single timescale for everything. It cannot show faster relearning after washout, cannot predict a rebound in the absence of new feedback, and gets interference effects wrong. Smith and colleagues summarized the problem bluntly: current models of trial-to-trial adaptation fail to account for savings, spontaneous recovery, anterograde interference, rapid unlearning, and rapid downscaling. That's a lot of unexplained territory. Their solution is elegant in its simplicity. They introduced two adaptive processes operating in parallel, each responding to motor error, but with opposite trade-offs. One process has high sensitivity to error but poor retention — it updates strongly after each mistake but forgets quickly. The other has low sensitivity to error but strong retention — it updates weakly but holds onto what it learns. The net motor output on any trial is the sum of these two internal states. In words, here is how the model works. The fast internal state on the next trial equals a retention factor times its current value plus a learning rate times the current error. The slow internal state does the same, but with a larger retention factor and a smaller learning rate. The net motor output is the sum of the two states, and the error is the difference between the environmental perturbation and that net output. Four relationships, two processes. Smith and colleagues fit this model to human force-field reaching data, and the parameter estimates are striking in how clearly they separate. The slow process had a retention factor of 0.99 — it keeps almost everything from trial to trial — and a learning rate of just 0.02. The fast process had a retention factor of only 0.59, meaning it loses roughly 40 percent of its state on every trial, but it compensates with a learning rate of 0.21, ten times higher than the slow process. By the end of the first learning block, the slow module already accounted for more than half of total adaptation. Even early in training, the persistent, low-sensitivity process is doing the heavy lifting. Now, here is where the model makes a prediction that seems almost paradoxical. Suppose you run a standard adaptation-then-extinction experiment. During extinction, errors drive the fast state toward zero quickly because its retention is low. But the slow state, which barely decays between trials, is still carrying a strong signal from the original adaptation. Once errors go to zero — once you clamp the error channel so there's nothing new to learn from — the fast state continues to decay while the slow state hangs on. So the net output, the sum of the two, should briefly drift back in the direction of the original adaptation. This is spontaneous recovery — a rebound — from nothing but the math of forgetting at different rates. To test this, Smith and colleagues built a novel paradigm using a robotic manipulandum — a device that participants held while making reaching movements. The robot could impose a viscous curl force field, perturbing the reach sideways in a consistent direction. It could also create an error-clamp channel that constrained lateral displacement to under 0.6 millimetres, effectively holding error at zero. Fourteen right-handed participants adapted to one force field, then experienced a brief exposure to the opposite field to wash out their compensation. After this, they performed a series of error-clamp trials with no perturbation and no new feedback. The model's prediction held. After the washout block, the first error-clamp trials showed near-zero or even oppositely directed forces. By roughly trials twelve to fifteen of the error-clamp block, the originally learned force pattern reemerged on its own. It grew, peaked, and then decayed back toward zero. The rebound was significant with a p-value below 0.0001 for all fourteen participants combined, and with a p-value below 0.01 for each seven-participant subgroup independently. There were no new errors, no new information — just two systems decaying at different rates, and the slower one briefly winning. That result is the empirical anchor of the paper. Smith and colleagues show that the same architecture — no additional mechanisms, no new assumptions — also accounts for the other four phenomena. Savings falls out naturally. After initial learning and washout, the slow state remains biased toward the original adaptation even when net behavior appears normal. When the perturbation returns, that biased slow state provides a head start, and relearning is faster. The model also predicts that inserting null trials before relearning should reduce savings because even the slow state will decay somewhat, which matches what Kojima and colleagues found in saccade adaptation. Anterograde interference follows from the same logic. When you switch to the opposite perturbation, the slow state — still carrying a strong signal from the first adaptation — resists updating in the new direction. The effective time constant for second learning is longer. The model predicts that this interference should grow as the first adaptation period gets longer, which gives experimenters a clean test. Rapid unlearning and rapid downscaling come from the fast process. Because the fast state has low retention and high error sensitivity, it can reverse or decay quickly when the environment changes. De-adaptation can appear faster than initial adaptation precisely because the fast module swings hard and decays hard. The model predicts that these speedups should diminish as initial training gets longer — the slow state becomes increasingly dominant, and it's the stubborn one. One framework. Four previously unconnected phenomena. All falling out from the interaction of two numbers: how much you learn per trial and how much you keep. Where might these two processes live in the brain? Smith and colleagues are careful here, and appropriately so. The cerebellum is a strong candidate, particularly for the fast process — cerebellar lesions produce dramatic deficits in the rate of motor adaptation. Work by Medina and colleagues from eyelid conditioning is suggestive: cerebellar cortex behaves like a rapidly updated module while the cerebellar nuclei develop a slower, more durable response, consistent with a cascade where fast cortical learning slowly trains a more persistent downstream structure. Recordings from motor cortex during force-field tasks also reveal two classes of memory cells whose outputs partially cancel — a pattern the authors suggest could reflect their fast and slow systems operating in motor cortex. Smith and colleagues acknowledge the modules could reside in the cerebellum, motor cortex, both, or distributed across multiple structures. The behavior constrains the architecture; it doesn't uniquely identify the anatomy. What this work ultimately argues is something deeper than any single phenomenon. Even within a single training session — within minutes — the brain is running at least two parallel learning algorithms at different speeds. The rich, sometimes counterintuitive behaviors of motor learning don't come from one process doing its job. They come from two systems in constant negotiation, each with its own memory, each pulling behavior in a direction that makes sense only when you know the other one is there. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.