Alignment across recording sessions

The sub-problem of placing recordings from different days or arrays into a common representational space even though the observed units change. Classic 'stitching' learns a per-session input and output layer that map each session's idiosyncratic channels into and out of a single shared latent (as in multi-session LFADS). Modern foundation-style models instead learn per-session or per-unit identity embeddings — sometimes called unit identification — so that adding a new session means learning only its small set of embeddings while the backbone stays fixed. A third family aligns latent manifolds directly with Procrustes/CCA, adversarial domain adaptation, or cycle-consistency.

This overlaps with drift-robust decoding but is framed differently: there, the goal is to keep one decoder stable over time; here, the goal is to build one model that spans many sessions at once so that a shared, session-invariant representation exists in the first place. Good alignment is what makes cross-session transfer and pooled pretraining possible.