Representation learning for continual task performance

Humans have the remarkable ability to learn multiple categorisation tasks in sequence without forgetting. For example, we may first learn how to categorise fruit by their shape, and later by their colour (“ripeness”). While previous work has focussed on elucidating the mechanisms that underlie flexible context-dependent information processing, much less is known about the format in which information is represented in the human brain, and how this promotes continual task performance. In this talk, I will present results from a series of neural network simulations, behavioural and neuroimaging studies that investigated the computational underpinnings of continual representation learning. Using artificial neural networks as mathematical toolkit and building on earlier work on cognitive control, we present a theory of context-dependent gating that can explain how several tasks are learned with minimal interference, and provide empirical evidence that representations in dorsal portions of prefrontal cortex are consistent with this theory. The results indicate that the brain has evolved mechanisms to map task-relevant feature dimensions onto separate, orthogonal coding axes, and filter out irrelevant information from the input signal.