Non-parametric discovery of population dynamics from large-scale neural activity recordings

Recent advances in neurotechnology enabled activity recordings from many neurons simultaneously, allowing us to study how neural populations are coordinated to drive behavior. Current population-analysis methods are based on fitting parametric models to data. However, these ad hoc models often distort dynamical features and result in ambiguous model comparisons. To overcome these limitations, we develop a non-parametric framework for discovering population dynamics directly from the data without a priori model assumptions. Our framework is based on latent Langevin dynamics, in which driving forces are directly optimized to effectively search the entire space of possible dynamics. The framework incorporates diverse, non-linear relationships between population-dynamics and firing-rates of single neurons. We derive a gradient descent algorithm for optimization over the space of continuous functions and use cross-validation and early stopping for regularization. Our framework accurately recovers qualitatively different population-dynamics simultaneously with diverse firing-rate profiles of single neurons.