Constructing a Hierarchical Kalman Filter for Extraction of Adaptive Neural Population Dynamics
Danilo Perez Jr, New York University
Cognition requires integrating and using information at multiple timescales. For example, during a value-based decision making task, neurons in the Orbitofrontal Cortex (OFC) are known to reflect the value of offered rewards on individual trials, while also adapting to changes in reward distributions over trials. However, how populations of neurons represent cognitive variables on multiple timescales is unclear. We constructed a hierarchical model to extract these signatures separately from rat OFC neurons during decision-making in a temporal-wagering task. This hierarchical model describes neural activity using latent state space models which capture fast within trial dynamics that are themselves modulated on the slower time scale of trials by another set of low dimensional latents. We used Expectation-Maximization to learn the parameters of this hierarchical model and extract the corresponding latents at the two timescales. We found that the model provides a robust characterization of single neuron responses, with slow latents that reflect rats’ inferred reward states over trials, and fast latents that reflect individual reward offers. These results provide a computational framework for extracting dynamical signatures of cognitive computations spanning multiple timescales. Effectively, model fits reflect intermediate latents that evolve according to with-in trial dynamics including cue perception and reward retrieval, while contextual adaptations are represented on the trial-to-trial timescale. This tool represents a relatively new tool for understanding computation across multiple time scales through the lens of circuit dynamics.
Authors: Danilo T. Perez-Rivera1, Shannon Schiereck1, Cristina Savin1,2, Christine Constantinople1
1Center for Neural Science, New York, USA
2Center for Data Science, New York University, New York, USA
Abstract Author(s): (see above authors)