Cerebro-cerebellar networks facilitate learning through feedback decoupling

Abstract

Behavioural feedback is critical for learning in the cerebral cortex. However, such feedback is often not readily available. How the cerebral cortex learns efficiently despite the sparse nature of feedback remains unclear. Inspired by recent deep learning algorithms, we introduce a systems-level computational model of cerebro-cerebellar interactions. In this model a cerebral recurrent network receives feedback predictions from a cerebellar network, thereby decoupling learning in cerebral networks from future feedback. When trained in a simple sensorimotor task the model shows faster learning and reduced dysmetria-like behaviours, in line with the widely observed functional impact of the cerebellum. Next, we demonstrate that these results generalise to more complex motor and cognitive tasks. Finally, the model makes several experimentally testable predictions regarding cerebro-cerebellar task-specific representations over learning, task-specific benefits of cerebellar predictions and the differential impact of cerebellar and inferior olive lesions. Overall, our work offers a theoretical framework of cerebro-cerebellar networks as feedback decoupling machines.

Emergent Low-Frequency Activity in Cortico-Cerebellar Networks with Motor Skill Learning

“…Our increased M1–cerebellum LFP coherence with skill learning is consistent with this observation. Neural network models of cortico-cerebellar networks show that cerebellar feedback improves the rate of learning and that the cerebellar network also carries task representation ( Boven et al, 2023 ). Our experimental data support this notion as well.…”

Cerebellar-driven cortical dynamics enable task acquisition, switching and consolidation

“…Such combination of fast and gradual learning is reminiscent of recent experimental results which suggest significantly faster timescales of plasticity in the hippocampus compared to the prefrontal cortex during a cognitive task 82 . Moreover, the consolidation period can be related to the idea that a task-optimised cerebellum can be utilised as a cortical teacher 30,31 . It is in principle possible for cerebellar-thalamo-cortical projections to support this dual role of the cerebellum as both a driver and teacher of cortical states.…”

“…5F), we follow the method in 26 . In particular, we divide the presentation period evenly into 3 time windows - [1-15, 16-30, 31-45] - and fit the model choice according to a logistic regression model where ŷ denotes the predicted model choice probability, S is the sigmoid logistic function, E i = # R i − # L i is the different in the total number of ‘right’ and ‘left’ inputs in window i , and β i is the respective weight on that window. ŷ is fitted to minimise the negative log likelihood of the observed model decisions.…”

“…Indeed, an ever-growing array of clinical 22 , functional imaging 23,24 , and optogenetic [25][26][27] studies support an important cerebellar contribution to cortical activity in both motor and non-motor domains. Recently, two hypotheses on the computational role of cortico-cerebellar loops have been put forward [28][29][30][31] . The first asserts that the cerebellum reinforces cortical-dependent goal-directed behaviour by appropriately steering or stabilising cortical states in real-time 28,29 .…”

“…The first asserts that the cerebellum reinforces cortical-dependent goal-directed behaviour by appropriately steering or stabilising cortical states in real-time 28,29 . The second also promotes the cerebellum as a facilitator of goal-directed cortical transitions, but it does so indirectly via teaching signals which lead to cortical plasticity 30,31 . Whilst these two views may co-exist, it is the former that is well placed to operate under weakly plastic cortical networks.…”