Abstract
Animal survival depends on behavioural adaptation to the environment. This is thought to be enabled by plasticity in the neural circuit. However, the laws which govern neural plasticity are unclear. From a functional aspect, it is desirable to correctly identify, or assign “credit” for, the neurons or synapses responsible for the task decision and subsequent performance. In the biological circuit, the intricate, non-linear interactions involved in neural networks makes appropriately assigning credit to neurons highly challenging. In the temporal domain, this is known as the temporal credit assignment (TCA) problem.This Thesis considers the role the cerebellum – a powerful subcortical structure with strong error-guided plasticity rules – as a solution to TCA in the brain. In particular, I use artificial neural networks as a means to model and understand the mechanisms by which the cerebellum can support learning in the neocortex via the cortico-cerebellar loop. I introduce two distinct but compatible computational models of cortico-cerebellar interaction.
The first model asserts that the cerebellum provides the neocortex predictive feedback, modeled in the form of error gradients, with respect to its current activity. This predictive feedback enables better credit assignment in the neocortex and effectively removes the lock between feedforward and feedback processing in cortical networks. This model captures observed long-term deficits associated with cerebellar dysfunction, namely cerebellar dysmetria, in both the motor and non-motor domain. Predictions are also made with respect to alignment of cortico-cerebellar activity during learning and the optimal task conditions for cerebellar contribution.
The second model also looks at the role of the cerebellum in learning, but now considers its ability to instantaneously drive the cortex towards desired task dynamics. Unlike the first model, this model does not assume any local cortical plasticity need take place at all and task-directed learning can effectively be outsourced to the cerebellum. This model captures recent optogenetic studies in mice which show the cerebellum as a necessary component for the maintenance of desired cortical dynamics and ensuing behaviour. I also show that this driving input can eventually be used as a teaching signal for the cortical circuit, thereby conceptually unifying the two models.
Overall, this Thesis explores the computational role of the cerebellum and cortico-cerebellar loops for task acquisition and maintenance in the brain.
| Date of Award | 5 Dec 2023 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Rui Ponte Costa (Supervisor) & Paul T Chadderton (Supervisor) |