Three decades of discoveries

Neural control of movements

Our laboratory combines mathematics, human behavior, and neurophysiology in nonhuman primates to understand how the brain predicts the consequences of actions, learns from errors, and turns neural activity into movement.

Recent work · 2022–2026

A neural language for the cerebellum

How can thousands of Purkinje cells act together to control a single movement?

Individual Purkinje cells do not simply report the direction of an eye movement through their firing rates. A cell can fire during movements in many directions, yet each spike has a measurable downstream influence—a potent vector in the space of movement.

By measuring the consequences of spikes and combining these vectors across simultaneously active cells, we found that populations of Purkinje cells can construct a motor command through superposition and cancellation. Components that would move the eyes in unwanted directions cancel, while the remaining activity helps guide the eyes toward their goal.

Related work examines how complex spikes perturb movements, how climbing fibers distinguish sensory events from movement plans, how synchronized Purkinje cell activity shapes output, and how the cerebellum controls the deceleration of tongue movements.

Figure from Fakharian et al. (2025) illustrating Purkinje cell potent vectors and their superposition
Neural computation in the Purkinje-cells is like a tug-of-war. Each neuron pulls along its own potent vector, trying to induce a movement (displacement of the flag), but many of the vectors cancel. There is a movement only when there is an imbalance of activities in the population. When there is an error in the movement, a teaching signal via the climbing fiber instructs the neuron on how to change its firing rate, thus altering how hard that neuron will pull along its potent vector.
2010–2020

Movement vigor reflects the value of action

The brain decides not only what to do, but how energetically to do it.

Movements toward valuable goals tend to be more vigorous. This observation led us to propose that movement vigor provides a window into the brain's estimate of the subjective value of an action, incorporating expected reward, effort, and time.

Studies of eye movements, reaching, and foraging connected decision-making to movement execution: the computations that influence which action is worth taking can also influence the speed and energy with which it is performed.

We move faster towards goals that we consider more valuable.
We move faster towards goals that we consider more valuable. Our research has considered this problem from a neuroeconomic standpoint, suggesting that the reward at stake discounts the energetic cost of the movement, invigorating our actions.
2014 onward

A memory of errors

Learning from an error depends on the errors that came before it.

Motor adaptation is not governed by the current error alone. We found that the history of experienced errors changes how strongly the brain responds to a new error. Repeated errors of a consistent kind can increase sensitivity to future errors, while inconsistent errors can reduce it.

This idea adds a second kind of memory to sensorimotor learning: not only a memory of the motor command, but also a memory that adjusts the learning rate itself.

Figure from Herzfeld et al. (2014) showing how error history changes learning
History of error alters error sensitivity. (A) Reaching paradigm with force-field perturbations. The yellow circles note a perturbations state, and z indicates probability of remaining in that state. The slow, medium, and rapidly switching environments are shown. We measured error sensitivity through probe trials in which participants experienced a constant perturbation, sandwiched between two error-clamp trials. (B) Learning from error in the probe trials, plotted as a percentage of the ideal force.
2000 onward

Learning unfolds across multiple timescales

The multiple timescales of memory give rise to spontaneous recovery after extinction

Adaptation reflects the interaction of processes with different learning and retention properties. A fast process responds strongly to error but forgets quickly; a slow process learns more gradually but retains what it has learned.

This framework explains patterns of adaptation, forgetting, and savings that cannot be understood as a single uniform memory. It provides a mathematical account of how short-lived and persistent changes coexist during skill acquisition. Importantly, it provides an explanation as to why extinction training does not erase memories, and instead is followed by spontaneous recovery.

Figure illustrating fast and slow adaptive processes in motor learning
When one practices a new motor task, a motor memory forms. This memory is composed of two components: a fast state that learns quickly but tends to forget with passage of time, and a slow state that learns slowly but tends to retain its information. After learning, if we reverse the task and produce extinction, the two states sum to produce a cancellation, but the result is not a return to baseline. Rather, during extinction, the fast state is competing with the slow state. After extinction, with passage of time the slow state declines gradually, whereas the fast state declines rapidly. This gives rise to the phenomenon of spontaneous recovery of the memory.
Foundational work · 1994 onward

Internal models: learning to predict the world

The brain learns how the world responds to our actions—and uses those predictions to move.

In our early experiments, people reached while a robotic device applied unfamiliar forces to their hands. Their initial movements were distorted. With practice, the trajectories straightened; when the forces were removed, the hands deviated in the opposite direction.

These aftereffects revealed that the nervous system had learned a predictive representation of the new dynamics rather than merely reacting to each disturbance. The force-field adaptation paradigm became a widely used experimental approach to studying internal models, sensory prediction errors, and the role of the cerebellum in motor learning.

Hand trajectories during initial exposure, adaptation, and aftereffects in a robotic force field
Learning to reach in a force field. The field is produced by a robotic manipulandum and the brain learns to construct an internal model of the field.

Across these projects, a common question connects our earliest experiments to our most recent work: how does the nervous system transform experience into predictions, and predictions into coordinated movement?