Neural world models supporting flexible behaviour

Thesis

A core component of intelligent behaviour is our ability to flexibly adapt to changes in the external world or internal goals. In order to achieve such flexible behaviour, it has been argued that humans and animals maintain sophisticated internal models – or cognitive maps – of the environment that monitor the state of the world and encode the relationships between objects. This thesis examines 1) how internal models may be monitored and updating in a changing world, and 2) the general principles under which internal models may be learnt, how they may be used for flexible and efficient planning, as well as how they may be generalised to new experiences.

In the first part of the thesis, we investigate the neural substrates of maintaining flexible behaviour in a changing environment. In order to maintain flexibility, the uncertainty in internal models guiding behaviour needs to be increased when the current behavioural strategy no longer accurately reflects the environment. We use functional MRI to measure an internal model representing participants’ beliefs about the state of the world in medial orbitofrontal cortex. We show that changes in pupil diameter – thought to index the neuromodulator noradrenaline – predict increases in uncertainty in this internal model. Anterior cingulate cortex tracked the certainty in participants’ beliefs and predicted changes in pupil diameter. Together these results provide evidence for neuromodulatory control of the flexibility of internal models in dynamic environments.

In the second part of the thesis, we investigate how internal models may be learnt, generalised to novel situations and used for flexible and efficient planning. Cells in the brain that fire when an animal is in certain spatial locations are an internal model of space. They also provide “ground truth” representations of a 2-dimensional environment allowing us to investigate general principles under which they may emerge. We show these cells can be thought of as representing the underlying structure or topology of the environment. We further show how these representations can be used for flexible and efficient planning of arbitrary structures, as well as how they may be generalised to novel situations allowing reuse of previously learnt knowledge. These results suggest the way we process space may be an instance of general coding algorithms responsible for organising knowledge of all different kinds.

Together, these studies demonstrate how neural world models may support flexible behaviours.

Authors: Muller, T

• DOI: 10.5287/ora-8rwveaver
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English