Neural dynamics outside task-coding dimensions drive decision trajectories through transient amplification

Linking neural activity to behavior typically involves identifying activity subspaces that encode task information, such as stimuli, memory, and choices. However, it is unclear whether activity in these "coding dimensions" drives behavior or merely reflects the underlying computations. Neural activity in other "residual dimensions" is typically ignored. We developed a recurrent neural network that explicitly models interactions between coding and residual subspaces. Applied to multi-regional Neuropixels recordings from mice performing a delayed movement task, the model reveals that perturbations of residual dimensions reliably alter behavior, whereas perturbations of the choice dimension, which encodes the animals decision, are less effective. These effects arise because residual dimensions drive amplification across an intermediate number ([~]10) of dimensions, before the dynamics settle into discrete attractors corresponding to choice. Our findings show that neural activity previously considered task-irrelevant can have critical roles in driving behavior.