Probing cerebello-cortical motor learning via recurrent neural networks – CC-RNN

This project uses recent developments in recurrent neural networks, a type of artificial neural network that can be trained to perform many different kinds of tasks. By leveraging this technique, we train networks that have the same cerebello-cortical architectures to perform motor and cognitive tasks, and then reverse engineer how these models are able to learn. We can then analyse how the resulting trained model is able to perform many different tasks using the same data science techniques that are used to analyse large-scale neural recordings. This can reveal how the neural representations may differ in the neocortex and in the cerebellum, and how this compares to representations observed in real data. We also use a type of machine learning called reinforcement learning to understand how the cerebellum could use reward information to help guide its role in controlling behaviour, including its role in cognition.

The main results of this project will be to understand how the cerebellum and the neocortex are able to work together to control behaviour. One result will be which sites of plasticity and which architectural features are key in reshaping neural representations as they are transformed between the neocortex and the cerebellum in their looped pathways. This will help to identify whether interregional deep pathways linking the two regions constitute an important source of plasticity during motor learning, which is still poorly understood. Another result of this project will be to understand how the cerebellum could be involved in cognition using reinforcement learning, a type of machine learning algorithm that can teach agents how to control their behaviour to achieve the most rewards. Recent experimental evidence has shown surprising evidence of reward signals in the cerebellum, especially since the structure of the cerebellum is already thought to be perfectly optimised for sensory error based learning. One of the main results of this project is to propose a reinforcement learning interpretation of the cerebellar cortex. Future work will address how reinforcement learning in the cerebellum could be integrated with the neocortex during cognitive tasks.

A focus of this project is to propose the first models of how cerebellar error learning can modify neocortical representations for more complex behavioural control, and how the cerebellum can be involved in cognitive function using reinforcement learning.

The scientific production of this project will be multiple scientific articles on cerebello-cortical interaction and cerebellar reinforcement learning for cognition.

Accumulating evidence suggests that naturalistic behaviours are driven by interactions between brain regions traditionally studied in isolation. For example, recent work has shown that the cerebellum and the motor cortex, two highly interconnected regions that are each crucial for motor behaviour, coordinate during motor learning and execution. However, progress towards understanding cerebello-cortical interaction has been hindered by the drastically different circuitries of these regions, coupled with the inaccessibility of the deep subcortical pathways connecting them. As a result, it remains unclear how to integrate leading theories of cerebellar and cortical function in motor control.

Recent theoretical and experimental evidence has argued that the motor cortex flexibly generates the spatiotemporal patterns necessary for complex movements by exploiting its rich dynamics. This concept is strongly influenced by the emergence of recurrent neural networks (RNNs), a class of AI-inspired models that can be taught to solve a wide range of motor and cognitive tasks via changes in recurrent connectivity. In contrast, the feedforward, evolutionarily-conserved circuitry of the cerebellum is thought to be optimized for learning sensorimotor relationships and error-based adaptation of movements. Yet this classic view of cerebellar motor control is not able to explain recent evidence of reward information in cerebellar Purkinje cells, nor its mysterious role in cognition.

I propose to use RNNs to probe how these two distinct neural circuits work together during motor learning. I will extend the traditional neocortical-like RNN architecture to incorporate a cerebellar-like module capturing its divergent-convergent structure and the sparser interregional pathways. I will train this cerebello-cortical RNN (CC-RNN) to perform motor and cognitive tasks using machine learning methods, and will use systems neuroscience and statistical learning tools to analyze the resulting learned structural and functional properties. With this approach, I will address four questions that have been challenging to answer with experiments or traditional circuit modeling: 1) how cerebellar and motor cortical representations co-emerge during learning, 2) how cerebellar adaptation reshapes motor representations in the neocortex, 3) the impact of cerebello-cortical scaling on complex behaviours, and 4) the role of cerebellar adaptation in cognition. More broadly, this project will lead to a fuller understanding of how cortical and subcortical regions interact to produce complex behaviours.