AbstractCurrent accounts of cognition propose that the brain uses sensory information to build internal models and make predictions about the world. If predictions fail, it may be appropriate to update such models to ensure desired outcomes are still achieved. However, knowing exactly when to update beliefs and action plans – and by how much – is a crucial challenge. This thesis argues that the Locus Coeruleus (LC), a small brainstem nucleus which provides the brain’s main
source of noradrenaline, plays a critical role in navigating this challenge by allowing the brain to strike the right balance between fixing and updating beliefs.
Using the theoretical framework of Active Inference (AI), I describe the LC as a response system triggered by errors in the inference of states and actions. AI simulations reproduce experimentally observed LC spiking patterns, and phasic and tonic firing (often considered to be distinct ‘modes’) are unified, appearing naturally as emergent features of belief updating. I also demonstrate that AI agents can update their internal models more effectively by using state-action
prediction error to optimise learning rate.
In a series of acute experiments in rodents I developed an approach to test theoretical predictions, using silicon probes to record multiple LC units and optogenetics to manipulate their activity. This produced the largest known dataset of optogenetically identified LC units and enabled the first assessment of the reliability of ‘traditional’ methods of cell identification. I also demonstrated that the nucleus is characterised by low levels of pairwise cell correlation and
displays complex internal interactions – underlying the LC's inhomogeneity.
Finally, I piloted the use of a two-arm bandit experiment using behaving rodents. Whilst results were inconclusive, this project demonstrated the feasibility of using a simple behavioural task to test the predictions of the Active Inference models – an approach which could be developed in future experiments.
In summary, through complementary computational and experimental avenues this work demonstrates the rich internal dynamics of the LC and outlines its crucial computation role within a predictive brain.
|Date of Award||24 Mar 2020|
|Supervisor||Anthony Edward Pickering (Supervisor), Matt W Jones (Supervisor) & Rosalyn J Moran (Supervisor)|