Novel neural network models of multiscale brain network dynamics and behavior with real-time inference and control capability

University Of Southern California, Los Angeles CA

Investigators

Abstract

PROJECT SUMMARY Our motor, cognitive, and emotional functions arise from the temporal dynamics in the activity of large populations of neurons across brain networks. Dynamic models of neural population activity describe it in terms of low-dimensional latent factors. Recent neural network models of population activity have shown the importance of capturing its nonlinearity. But despite their power, these state-of-the-art nonlinear neural networks have a non-causal inference structure and do not directly address missing neural data, which can happen in wireless naturalistic recordings. Also, these neural networks don’t enable closed-loop control. These issues have limited the utility of these neural network models to correlational validations of hypotheses/models and to experiments with constrained behaviors. But a major emerging goal in neuroscience is to enable causal validations of hypotheses/models with real-time and/or closed-loop perturbation experiments and to study naturalistic behaviors in unconstrained or long-term experiments. We will develop dynamical models that not only capture nonlinearity, but also enable two novel critical capabilities: flexible inference and closed-loop control of latent factors and behavior. Flexible inference is defined as the model’s ability to allow for all the following: a) causal/real-time inference, b) non-causal inference, c) accounting for missing neural data. Also, to extend our models for diverse neuroscience investigations, we will develop multiple novel learning methods to address 3 additional major attributes of neural population dynamics that remain challenging to capture in nonlinear models: i) these dynamics happen at multiple spatiotemporal scales, ii) relate to a mixture of behaviors and internal states that co-occur, and iii) are driven by both intrinsic dynamics and inputs. Thus, this program will provide novel neural network models of neural population activity that can capture nonlinearity, flexible inference, and closed- loop control, and can also admit multiscale data, dissociate behaviorally relevant vs. irrelevant dynamics, and disentangle input vs. intrinsic neural dynamics. We will also develop and share extensive software and documentation for these models and methods. We will comprehensively validate the models and methods on diverse existing nonhuman primate and human datasets including public datasets with multiscale neuronal spiking, field potential, and intracranial recordings, from different brain regions, with measured sensory, neural and neurostimulation input, and during various cognitive and motor behaviors. Our program includes a diverse network of investigators and end-users and outreach activities to enhance diverse perspectives. The developed models and methods will greatly impact diverse neuroscience domains by providing novel tools to causally drive and test new hypotheses about how multiscale neural dynamics control behavior, inform experimental paradigms and data-collection (e.g., closed-loop perturbations), and enhance neurotechnology for decoding/modulation.