Title: Unifying neural population dynamics, manifold geometry, and single-cell selectivity
Abstract: Single neurons show complex, heterogeneous responses during cognitive tasks, often forming low-dimensional manifolds in the population state space. Consequently, it is widely accepted that neural computations arise from low-dimensional population dynamics while attributing functional properties to individual neurons is impossible. I will present our recent work bridging single-neuron heterogeneity to manifold geometry and population dynamics. First, we developed a flexible modeling approach for simultaneously inferring single-trial population dynamics and tuning functions of individual neurons to the latent population state. Applied to spike data recorded during decision-making, our model revealed that all neurons encode the same dynamic decision variable, and heterogeneous firing rates result from diverse nonlinear tuning of single neurons to this decision variable. Second, using a firing-rate recurrent network model, we mathematically prove that responses of single neurons cluster into functional types when population dynamics are confined to a low-dimensional linear subspace, with the number of distinct response types equal to the linear dimension of the neural manifold. We confirm these predictions in recurrent neural networks trained on cognitive tasks and brain-wide neural recordings from mice during a decision-making behavior. Our findings show that low-dimensional population dynamics can be understood in terms of functional cell types, and random mixed selectivity emerges only in the limit of high-dimensional dynamics.