Organization of reward and movement signals in the basal ganglia and cerebellum

Abstract

The basal ganglia and the cerebellum are major subcortical structures in the motor system. The basal ganglia have been cast as the reward center of the motor system, whereas the cerebellum has been hypothesized to be involved in adjusting sensorimotor parameters. Recent findings of reward signals in the cerebellum have challenged this dichotomous view. To directly compare the basal ganglia and the cerebellum we recorded from oculomotor regions in both structures from the same monkeys. We partitioned the trial-by-trial variability of the neurons into reward and eye-movement signals to compare the coding within and across structures. Reward expectation and movement signals were the most pronounced in the output structure of the basal ganglia, intermediate in the cerebellum, and the smallest in the input structure of the basal ganglia. These results indicate different information convergence properties within these structures and demonstrate differences in internal computations.

Population coding of distinct categories of behavior in the frontal eye field

“…We first quantified the extent of directional tuning encoded on the different tasks for single neurons. To do so, we utilized the partial ω 2 ( ω 2 p , see methods) as a measure of the effect size of the directional tuning in the different conditions (Olejnik and Algina, 2000; Larry et al, 2022). ω 2 p serves as an unbiased estimator of the variability explained by an experimental variable, divided by the sum of this same variability and a noise term (Fig.…”

“…To quantify the directional tuning of a neuron over time, we calculated the partial ω 2 effect size ( ω 2 p ). ω 2 p is a common effect size measure used in ANOVA designs that is often preferred over other effect size measures since it is unbiased (Olejnik and Algina, 2000, 2003; Okada, 2013; Larry et al, 2022). We counted the number of spikes in bins of 100 ms, resulting in a distribution of the number of spikes emitted by the neuron in each condition and each bin.…”

“…This contrasts with other effect size estimators based solely on the average activity, which can be strongly biased by the trial-by-trial variability. The ω 2 p also has the advantage that it can be used with multiple conditions (such as different directions of movement) unlike other measures such as the choice probability that are defined for two task conditions (Larry et al, 2022).…”

Predictive reward-prediction errors of climbing fiber inputs integrate modular reinforcement learning with supervised learning

“…These forward models likely involve a network loop that includes the inferior olive nucleus, cerebral cortex, basal ganglia, and cerebellum. They operate anticipatorily and collectively support modular reinforcement learning (Bostan and Strick, 2018;Chabrol et al, 2019;Wagner and Luo, 2020;Watabe-Uchida et al, 2017;Yamazaki and Lennon, 2019;Kostadinov and Häusser, 2022;Larry at al., 2024). In this framework, each functional module, represented by Purkinje cells, engages in supervised learning designed for specific actors, as depicted in Fig 5A.…”