Disentangling input dynamics from intrinsic neural dynamics in modeling of neural-behavioral data

Neural dynamics in a given brain region are generated either due to recurrent connections in the region (intrinsic dynamic), or due to temporally structured external inputs (input dynamics). These inputs can be sensory stimuli or excitations from upstream brain regions. While measuring all inputs is infeasible experimentally, measurements of sensory inputs such as task instructions or partial measurements of neural inputs into a brain region are often possible. When learning models of neural population dynamics, current learning methods can either account for measured input but not behavior, or vice versa. Here, we show that accounting for both measured input and behavior during model learning is critical for correct interpretation of neural computations underlying a specific behavior. We thus develop a new analytical learning method that can incorporate measured inputs into a combined neural-behavioral dynamical model. By simultaneously accounting for neural activity, behavior, and measured input during learning, this method dissociates intrinsic behaviorally relevant neural dynamics from both measured input dynamics and other intrinsic neural dynamics and prioritizes their learning so that they are not missed/masked by other intrinsic neural dynamics. First, in simulations, we show that accounting for input but not behavior or vice versa can lead to inaccuracies when learning dynamical models of neural activity. Second, we show that without their prioritized learning enabled by our method, intrinsic behaviorally relevant neural dynamics are learned much less accurately even when inputs are considered. Third, we analyze motor cortical population activity from two public datasets in which monkeys make random reaching movements following task instruction sensory inputs. We find that our method reveals intrinsic behaviorally relevant neural dynamics that are more predictive of neural-behavioral data and remarkably similar across the two datasets/animals; these consistent dynamics are missed by existing methods, which account for either measured input or behavior but not both.