Preferential subspace identification (PSID)

PSID (Sani, Shanechi et al.) is a system-identification method that learns a latent state-space model of neural activity while explicitly prioritizing the part of the dynamics relevant to a measured behavior. Standard identification finds the latent dynamics that best reconstruct the neural data, in which behaviorally relevant activity can be swamped by larger behavior-irrelevant components; PSID instead uses the behavior signal during identification (via projection/regression onto behavior before and during subspace estimation) to isolate a behaviorally relevant latent subspace and its dynamics, optionally alongside the remaining neural-only dynamics.

The practical payoff is more accurate, lower-dimensional behavior decoding and a cleaner dissociation of what the population is doing that relates to behavior from what it is doing for other reasons. It is a linear method in its original form (with nonlinear and preferential deep variants such as DPAD following the same principle), and, like all supervised dissociations, its conclusions are conditioned on the specific behavioral variable chosen — activity labeled irrelevant may simply be relevant to a behavior that was not measured.