Behavior-dLDS: A decomposed linear dynamical systems model for neural activity partially constrained by behavior

A team of researchers from institutions including Janelia Research Campus has published a new machine learning model called Behavior-decomposed Linear Dynamical Systems (b-dLDS). The model is designed to solve a fundamental problem in neuroscience: brain-wide recordings capture the activity of tens of thousands of neurons simultaneously, but this activity is a messy mix of signals directly driving observable behavior and a vast amount of unrelated 'internal computation' or background chatter. Traditional models often try to force all neural dynamics to correlate with behavior, which can obscure the true, distributed networks responsible for actions.

The b-dLDS model addresses this by treating behavior as a coarse-grained, lower-dimensional output of specific latent neural subsystems. It decomposes the full neural population activity to identify distinct dynamical subsystems, only partially constraining them by the observed behavior. This allows it to disentangle and highlight the specific connectivity networks that generate behavior, separate from other ongoing brain processes. The team validated b-dLDS on controlled simulations, showing it outperforms state-of-the-art behavior-supervised models. Crucially, they demonstrated its real-world scalability by applying it to large-scale recordings of a zebrafish hindbrain during a complex positional homeostasis behavior, successfully mapping the behavior-related dynamic networks.