Title: Linking learning to behaviour via neural representation geometry.
Abstract: Recent developments in experimental neuroscience enable us to simultaneously record the activity of a large number of neurons and to target specific components of neural circuits with genetic techniques. These methods pave the way to a deeper understanding of how the brain’s collective dynamics instantiate learning and behaviour. Obtaining such a conceptual understanding from large, high-dimensional neural datasets, however, requires concurrent advances in theoretically driven circuit modelling. By examining network mechanisms that underlie the emergence of complex behaviour, I will show that neural representation geometry may be the key approach to tying animal behaviour and learning to circuit mechanisms. I will demonstrate how neural activity from cortical areas unfolds through temporal sequences of neural states organized in behavioural plans or episodes. Then, I will illustrate how these neural representations in premotor areas drive the generation of actions, allowing us to predict the intention to act. Finally, I will investigate how key geometrical aspects of these representations emerge, de novo, as a result of learning to predict upcoming sensory experience. Altogether, I will show how both learning mechanisms and behavioural demands shape the geometry of neural representations. I will build these results by using theoretical and computational techniques that combine dynamical systems, machine learning, and statistical physics approaches.