Roxana Zeraati: Adaptive behavior across timescales

Animals behave adaptively on a broad range of timescales. I will discuss how we combined neural data analysis, computational models, and machine learning approaches to uncover mechanisms underlying such adaptivity across different cognitive tasks. Applying a Bayesian method we show that ongoing spiking activity unfolds across distinct timescales that adapt to behavioral states and correlate with behavioral timescales. Then using mechanistic and task-optimized recurrent neural networks we find the link between neural timescales, state of dynamics, and task performance. Our results suggest that nonlinear recurrent interactions are a key mechanism for developing adaptive timescales in biological and artificial neural networks.