Condensed Matter Seminar: Neural representation geometry: a mesoscale theoretical approach linking learning to complex behavior

Recent developments in experimental neuroscience allow us to record the activity of many neurons at the same time, opening the door to a grounded theoretical understanding of how the brain's collective dynamics instantiate learning and behavior. I will demonstrate how neural representation geometry - an emergent theoretical approach - may hold the key to such an endeavor. We will proceed in three steps. 1) We will start by establishing a connection between the sequential dynamics of complex behavior and geometrical properties of neural representations. To this end we will leverage the insight of attractor neural networks. 2) We will then link these geometrical properties to underlying circuit components. Specifically, we will uncover connectivity mechanisms that allow the circuit to control the geometry of its representations using novel statistical physics results. 3) Finally, we will investigate how key geometrical structures emerge, de novo, through learning. To answer this, we will analyze the learning of representations in feedforward and recurrent neural networks trained to perform multiple tasks using machine learning techniques. As a result, we will show how both learning mechanisms and behavioral demands shape the geometry of neural representations. Along the way, I develop theoretical and computational techniques that combine dynamical systems, statistical physics and machine learning approaches.