Cognitive maps for causal inference

It has long been appreciated that humans and other animals form internal models that encode environmental states and the relationship between those states. These representations are commonly referred to as “cognitive maps” and are thought to provide a scaffold for flexible behaviour, giving rise to higher cognitive functions like inference, knowledge transfer, or continual learning that have proven extremely challenging to reverse engineer in artificial intelligence research.

Thus far, research has conceived cognitive maps predominantly as prospective representations that link states in their chronological order to predict future events, e.g., to decide which action is most likely to produce a desired outcome. However, recent work in non-human animals suggests that the brain also forms and uses retrospective representations that link states in their reverse order to infer candidate causes of rare but significant outcomes. Prospective and retrospective maps are computationally distinct and may serve different functions, supporting decisions and learning respectively. However, it is currently unknown how these two types of maps are distinctly represented in the human brain, and how the information they provide is integrated for adaptive behaviour (e.g., if retrospectively inferred causes serve to revise predictive models and/or if prediction errors serve to trigger retrospective information search). This PhD project aims to fill these gaps using a computational cognitive neuroscience approach.

The PhD candidate will develop novel tasks that vary prospective and retrospective action-outcome associations independently (e.g., by changing the frequencies of predictable and unpredictable rewards) and use neuroimaging to identify networks of brain regions that track changes in prospective and retrospective association strength. The project will also examine the information flow between prospective and retrospective maps, and their role in habit formation, maintenance, and extinction, which could have important clinical implications (e.g., in the context of addiction).

The project brings together supervisory expertise in representation learning and brain imaging (Dr Muhle-Karbe), decision neuroscience (Prof Apps), electrophysiology and computational modelling (Dr Froemer). The PhD candidate will receive in-depth training in cutting-edge methods from psychology, neuroscience, and computational science, allowing them to develop a diverse set of valuable skills for a career within or outside of academia.

Interested candidates are encouraged to get in contact ([email protected]) for informal discussion before submitting an application.