Peter Dayan, Gatsby Computational Neuroscience Unit, UCL
Reverse replay during human decision problem solving
The fast ‘replay’ of sequences of neural representations has been suggested as supporting learning and online planning. However, it has largely been studied in spatial tasks in rodents. I will show how we came upon reverse replay during our latest attempt to use the decoding of MEG data to capture the process of human model-based planning in a non-spatial sequential decision-making problem. During epochs in which our subjects were planning, their brains spontaneously visited representations of approximately four states in the problem in fast sequences lasting on the order of 120 milliseconds. These sequences followed backward trajectories along the permissible paths in the task. I will discuss the possible implications of this finding.
This is joint work with Zeb Kurth-Nelson, Marcos Economides and Ray Dolan.
Mark Humphries, University of Manchester
Population activity in rule-learning prefrontal cortex reveals signatures of internal models
The idea that brains represent the world with probabilities, and compute using probability distributions, is a powerful explanation for a range of behavioural phenomena. But evidence that such probabilistic internal models are implemented by neurons has largely come from theoretical studies, and largely focussed on cortical sensory processing. But what about higher cortical regions? And learning? And action? In this talk I will present evidence that moment-to-moment population activity in rodent medial prefrontal cortex is generated from a probabilistic internal model of a maze task. We further show that sampling of population activity, and by implication the internal model, is updated by trial outcomes when learning the maze’s contingency rules. Our work suggests sample-based internal models are a general computational principle of cortex.