Bayesian Decoder Models with a Discriminative Observation Process

Abstract

The Bayesian state-space neural encoder-decoder modeling framework is an established solution to reveal how changes in brain dynamics encode physiological covariates like movement or cognition. Although the framework is increasingly being applied to progress the field of neuroscience, its application to modeling high-dimensional neural data continues to be a challenge. Here, we propose a novel solution that avoids the complexity of encoder models that characterize high-dimensional data as a function of the underlying state processes. We build a discriminative model to estimate state processes as a function of current and previous observations of neural activity. We then develop the filter and parameter estimation solutions for this new class of state-space modeling framework called the "direct decoder" model. We applied the model to decode movement trajectories of a rat in a W-shaped maze from the ensemble spiking activity of place cells and achieve comparable performance to modern decoding solutions, without needing an encoding step in the model development. We further demonstrate how a dynamical auto-encoder can be built using the direct decoder model; where the underlying state process links the high-dimensional neural activity to the behavioral readout. We applied the dynamical auto-encoder model in estimating the intention to verbally communicate of an epileptic participant and their companions. The result shows that the dynamical auto-encoder can optimally estimate the low-dimensional dynamical manifold which represents the relationship between the brain and behavior.