Distribution-matching alignment

Latent alignment that matches the distribution of today's neural data to a reference without needing paired, moment-by-moment correspondence between them. Rather than lining up specific samples, these methods ask that the two point clouds, viewed as distributions, look the same after alignment. Adversarial versions (an ADAN-style domain-adaptation network) train a discriminator that tries to tell reference latents from realigned test latents while the aligner tries to fool it; distribution-matching methods (NoMAD, nonlinear manifold alignment via distribution matching) match latent dynamics directly; cycle-consistent schemes require that mapping test to reference and back returns the original.

This subclass of latent-space stabilization is powerful precisely because it is fully unsupervised and does not require the two days to have visited the same behavioral states in the same order. Its weaknesses are the ones inherent to matching distributions rather than points: a symmetry in the data (a rotation that leaves the distribution invariant but swaps output labels) can be aligned away incorrectly, adversarial training can be unstable, and success is hard to certify online because there is no label to check against.