Stabilized latent-space decoder

A stabilized latent-space decoder maintains performance across days by aligning each day's neural activity to a fixed low-dimensional latent space, rather than by re-collecting labeled calibration data and refitting from scratch. The underlying premise is that although individual electrodes are nonstationary, the population's activity lives near a stable low-dimensional manifold whose geometry encodes the behavior and changes little over time. If one can find the transformation that maps today's recordings onto that reference manifold, the original decoder — trained once on the latent state — can be reused without new labels.

Alignment is typically achieved by matching the statistical structure of the new day's activity to the reference distribution using only unlabeled data (for example, matching latent covariance or distribution via distribution-alignment or adversarial methods), so no instructed calibration block is needed. This turns the recalibration problem from a supervised re-fit into an unsupervised alignment, promising near-zero-effort daily use. The approach depends on the manifold actually being stable and on enough neural coverage surviving to reconstruct it; when representations genuinely reorganize, alignment alone is insufficient and some supervised update is still required.