Neural mechanisms of learning a predictive world model

A fundamental aspect of mammalian behaviour is the ability to learn how the world works from experience to predict what will happen and plan what to do. Recent interdisciplinary advances from neuroscience, psychology and machine learning suggest that mammals, from mice to humans, abstract a “world-model” or “cognitive map” from experience, enabling appropriate behaviour in new situations via prediction and planning. The neural mechanisms that enable learning of such a world model are becoming clear through experiments in which rodents learn about their spatial environment while the activity of large populations of neurons are being recorded. We hypothesise that neurons in the hippocampus (a brain region required for learning and memory) encode states within a task and the sequence in which they are experienced, replaying these sequences to train nearby neocortical areas to capture the common transition structures of similar tasks to support prediction in new situations. In spatial foraging tasks, where “states” correspond to places, the hippocampal neurons will resemble “place cells”, while neurons in entorhinal cortex that capture the structure of transitions between places will resemble “grid cells”, and analogous patterns of neural responses can be predicted for non-spatial tasks. We aim to test the key predictions of this model, taking advantage of our virtual reality (VR) system for mice. VR allows us to manipulate a task’s transition structure in ways not previously possible, while also performing large-scale neuronal recordings and immediate inhibition of specific sequences of neural activity. A series of experiments will directly assess how hippocampal neurons encode the sequences of states experienced within a task, and how the overall structure of the transitions between states within a task is learned across multiple experiences. We will test predictions for how the task transition structure is represented by entorhinal neurons, and whether and how the replay of specific sequences of hippocampal neural activity causes these representations to be learned. The use of VR allows us to observe mice learning any transition structures that we want, including those that are impossible in real spatial environments. The outcome of this work will be a fundamental advance in cognitive neuroscience towards a mechanistic understanding of the rules of life governing complex behaviour in mammals.