Neural mechanisms for flexible behaviour in humans and artificial neural networks

Humans are remarkably good at generalising prior knowledge and using it to facilitate learning in novel situations, which depend on the ability to discover latent task structures within the environment and effectively abstract knowledge from the context it was originally obtained in. Current literature on cognitive flexibility indicates the existence of a rostrocaudal gradient of abstraction of task representations within the frontal cortex. However, the cued task-switching approach used by most studies fails to capture the knowledge abstraction and generalisation processes, and in that, misses the key aspects of human intelligent behaviour. Recently, a non-cued approach has been successfully used to show that humans learn tasks in a hierarchical manner whenever the environment allows it, even when it conveys no immediate behavioural advantage. An underlying mechanism has been proposed, in which the search for a latent task structure is carried out by the prefrontal cortex (PFC) and conjunctive coding of stimulus-action associations, by posterior frontal cortex, consistent with the rostro-caudal axis of abstraction. Here I propose a method to study how such hierarchical task structuring is reflected in the internal object representations in the PFC. Subjects will learn a task without explicit instructions about its latent structure, whilst undergoing an fMRI scan. I will use representational similarity analysis to track the development of abstract object representations in PFC, which should allow me to test whether the representations of stimuli sharing a higher-order context will become clustered in rostral PFC regions, in line with the proposed rostro-caudal organisation. In part 2 of the study, I will test the subjects on a positive transfer task to afford the examination of the following prediction: the similarity indices for stimuli governed by the abstract rule from part 1 will be higher than those at corresponding time points in the previous task. Extensions to this study could explore additional techniques such as computational modelling with the use of neural networks, as well as the investigation of related topics, for example how task representations using shared stimuli are formed without overrating existing knowledge.