Latent-space stabilization

A family of methods that fights drift not at the level of individual neurons but in the low-dimensional latent space the population traces out. The insight is that even as single-unit tuning drifts and units come and go, the neural manifold — the small set of population activity patterns actually used for the task — is comparatively stable in its geometry. If you can find, on each new day, the transformation that maps today's raw recording back onto a stored reference latent space, a decoder trained once on that reference keeps working without being retrained.

The influential demonstration (Degenhart and colleagues' 'stabilizer') aligned the low-dimensional subspaces of neural activity across days so that a fixed decoder held its performance despite substantial turnover of recorded units, using only brief unlabeled data to estimate each day's alignment. The approach reframes recalibration as a small, unsupervised geometry problem (find one alignment map) rather than a full supervised refit of the decoder, which is both cheaper and more stable.