Representation learning for continual task performance

Thesis

Humans have the remarkable ability to learn continually without forgetting, and adapt their behaviour to changing situational demands. While previous work has focussed on elucidating the mechanisms that underlie flexible context-dependent processing of information, much less is known about the format in which information is represented in the human brain, and how this promotes continual task performance. The aim of this DPhil was to develop computationally informed theories of representation learning for context-dependent processing, and test these in behavioural and neuroimaging recordings from healthy human participants.

Across a series of neural network simulations, behavioural and neuroimaging studies and re-analyses of freely available datasets with recordings from macaque FEF, I gathered evidence in support of earlier theories of cognitive control, which postulated that prefrontal cortex implements gating strategies that favour task-relevant over task-irrelevant information in the service of context-specific task goals. In chapter 3, I propose a computational framework to study representation learning for contextdependent decisions with artificial neural networks, and demonstrate how the same architecture can learn either high-dimensional and task-agnostic or low-dimensional and task-specific representations. In chapter 4, I tested predictions from these simulations in fMRI recordings from human participants who learned to perform a similar context-dependent decision task, finding that representations in fronto-parietal regions were highly task-specific, with relevant information from distinct tasks mapped onto orthogonal coding axes. In chapter 5, I introduce a model of human continual learning, in which the gating signal is learned by a simple Hebbian mechanism. Lastly, in chapter 6, I tested whether previously reported benefits of blocked over interleaved training generalised to abstract rules and whether they promote cross-domain transfer.

Taken together, this thesis introduces a computational theory of continual representation learning and provides empirical evidence that the human brain uses gating strategies to represent relevant information in context-specific subspaces.

Authors: Flesch, T

• DOI: 10.5287/bodleian:GwXNqGoyd
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: artificial neural networks; continual learning; fMRI
• Subjects: Psychology, Experimental; Computational neuroscience; Cognitive neuroscience