We trained mice to use a joystick to retrieve water. To test their cerebellar learning, we imposed a perturbation (gain = 2x; lower panel) that they needed to overcome to retrieve the reward.

Like humans wearing prism goggles, mice learn to adapt their behavior in only a few trials (dark blue). They retain a memory of this adaptation, which they need to "unlearn" when returned to control conditions (light blue).

But how can mice learn and form a memory so quickly if climbing fibers only provide a binary error/no error signal to Purkinje cells (as in classic supervised learning models)?

Imaging complex spikes in populations of Purkinje cells during learning reveals the answer: bands of cells respond similarly. Together, these cells encode both the magnitude and the direction of the behavioral error.

While the complex spike error signal from a single pull can only provide miniscule LTD/LTP of parallel fibers onto a Purkinje cell, the combined effect across the population of Purks is robust enough to influence the very next joystick pull.

If you prefer a talk to a paper, I walk through the full results and the story in this video: www.youtube.com/watch?v=rNQY...