Cerebellar climbing fibers convey behavioral information of multiplex modalities and form functional modules

Abstract

Cerebellar climbing fibers (CFs) convey sensorimotor information and their errors, which are used for motor control and learning. Furthermore, they represent reward-related information. Despite such functional diversity of CF signals, it is still unclear whether each CF conveys the information of single or multiple modalities and how the CFs conveying different information are distributed over the cerebellar cortex. We performed two-photon calcium imaging from cerebellar Purkinje cells (PCs) in mice engaged in a voluntary forelimb lever-pull task and demonstrated that CF responses in 93% of PCs could be explained by the combination of multiple behavioral variables, such as lever movement, licking, and reward delivery. Neighboring PCs exhibited similar CF response properties, formed functional clusters, and shared noise fluctuations of responses. Taken together, individual CFs convey behavioral information on multiplex variables and are spatially organized into the functional modules of the cerebellar cortex.

Dynamic organization of cerebellar climbing fiber response and synchrony in multiple functional modules reduces dimensions for reinforcement learning

“…These results demonstrated that the cerebellum reduces dimensions of activity in a large number of neurons to a much smaller number of components, each of which is driven by a synchronization scheme that conforms to a specific task. Interestingly, we also found that individual anatomical zones and even a single CF could contain signals from multiple components [53][54][55][56] .This study provided the first evidence simultaneously supporting the two major theories of cerebellar functions in a single task 18,[57][58][59][60][61][62] , and should contribute to resolution of the longstanding controversy 63 . This study also unveiled the secret of cerebellar functional architecture: learning from small samples is achieved by compartmentalization (reduced degrees of freedom) due to synchronization and dynamics; therefore, it may contribute to new-generation AI designs.…”

“…These results demonstrated that the cerebellum reduces dimensions of activity in a large number of neurons to a much smaller number of components, each of which is driven by a synchronization scheme that conforms to a specific task. Interestingly, we also found that individual anatomical zones and even a single CF could contain signals from multiple components [53][54][55][56] .…”

“…Thus, monitoring population activities as learning proceeds, combined with causal analysis of neuronal responses and behavior changes, is needed to understand diverse CF functions across TC populations. This is particularly crucial because CFs may multiplex motor, cognitive and reward-related information, as found in previous studies [53][54][55][56] . In Figure 6 we highlighted this possibility by showing a mixed tensor representation of 6,445 individual neurons in the cerebellar cortex.…”

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

“…PLS was necessary because reward and sensorimotor variables are considerably correlated with each other and we need to find meaningful correlations under this difficult situation of multicollinearity. PLS revealed relatively moderate zonal distributions of different variables, probably because each zone and even each neuron contain multiple functional components as demonstrated by previous studies (Markanday et al, 2021; Ikezoe et al, 2022; Hoang et al, 2022). Therefore, in the third analysis, we decomposed spiking activity into four tensor- components (TCs) by tensor component analysis (TCA) and then examined functional representations of each TC at a trial basis.…”

“…One of the reasons why relatively moderate anatomical distributions of different functions are found in Fig. 3 is that each zone and even each neuron contain multiple functional components as demonstrated in Hoang et al (2022) as well as in previous studies demonstrating multiplexed representations (Markanday et al 2021; Ikezoe et al 2022). In order to overcome this difficulty due to multiplexing for revealing precise functions of each component, we next examine functional representations of each tensor component utilizing Q-learning and trial-based analyses, while incorporating timing information of each spike instead of broadly computing the average spike rate over a wide temporal window.…”