Inferring Dynamics in Neural Recordings using LFADS with Per-trial Contextual Bias

Dynamical systems theory provides a promising framework for understanding neural activity. LFADS approximates neural recordings using a recurrent neural network (RNN) generator, but analyzing its dynamics has been challenging. We propose a simple modification to LFADS, adding a per-trial inferred bias, that dramatically improves interpretability by replacing complex multi-stable dynamics with simpler contextual dynamics. Both models were applied to neural activity recorded from macaque motor cortex during reaching movements. While the kinematic decoding was similar, they implemented contrasting dynamics. With standard LFADS, multiple dynamical systems, each with a distinct fixed point, exist simultaneously across all trials. The initial condition then utilizes this multi-stable system to drive the oscillation. With a per-trial bias, there was only a single fixed point, which is “contextually adapted” to approximate the activity for a given trial. The per-trial bias was modified based on the reach direction, resulting in reach-specific changes in the center of neural oscillations (fixed point locations), oscillation planes (eigen-planes), and for the first time, subtle alterations in oscillation frequencies (eigenvalues) around a dominant mode. Critically, this was easier to understand than the original LFADS model’s multistable solution, which required a much deeper analysis to understand how the fixed point structure leads to the per-trial dynamics. The per-trial bias captures variability commonly encountered in BCI studies. In Utah array recordings from BrainGate2 participants during attempted finger movements, one direction in the per-trial bias space accounted for non-stationarities over time. By synthesizing neural trajectories after correcting for this drift and augmenting training data, finger position decoding accuracy for the final 25\% of trials improved significantly (R2=0.74) compared to either ignoring earlier trials (0.66) or combining them without augmentation (0.67). For attempted speech, another direction captured variations in neural dynamics speed, allowing us to sort trials by behavioral variability (speech rate).