Dissociating behaviorally-relevant dynamics

A cluster of methods addresses a single problem: at any moment most population activity is not about the behavior an experiment measures, so latent models that simply capture the largest variance can bury the behaviorally relevant dynamics. The aim is to factor the latent state into a behaviorally relevant part and a behaviorally irrelevant (but still dynamically structured) part, so each can be modeled and interpreted on its own terms. Approaches include preferential subspace identification and its deep successor DPAD, targeted / disentangling seq-VAEs (e.g. TNDM), and priors or contrastive objectives that push behavior information into designated latent dimensions.

This matters for interfaces because a decoder built only on the behaviorally relevant subspace can be both more accurate and more stable, and because the irrelevant-but-structured component is itself of scientific interest (ongoing internal states, other movements, arousal). The unavoidable caveat is definitional: relevance is defined relative to the measured behavior and the model class, so the split is a lens, not a ground truth — a dimension deemed irrelevant to reach kinematics may be highly relevant to force, timing, or an unmeasured variable.