The no-canonical-coordinate problem
Transfer is hard for a structural reason. Two subjects have different electrode placements sampling different neurons; there is no correspondence that says 'unit 7 here equals unit 7 there'. Worse, even within one subject, the recorded population changes day to day as the array shifts and neurons come and go — the everyday non-stationarity that defeats a fixed decoder. A foundation model must be built so that neither problem is fatal. See cross-subject transfer and alignment across sessions.
Channel-agnostic architectures
The cleanest answer is a channel-agnostic architecture: make the model indifferent to which physical channel a token came from. POYO tokenizes individual spikes and attaches a learned unit embedding and a session context vector, then uses a permutation-invariant attention backbone. Because units are identified by learned latent codes rather than by array index, the model can ingest an arbitrary, variable set of channels — and a brand-new session is just new unit embeddings over the same shared dynamics.
Alignment and stabilization
The complementary strategy assumes a stable low-dimensional neural manifold underneath and only re-aligns a new session onto it. This is latent-space stabilization: find a transform mapping today's embedding onto the reference geometry. Methods range from linear Procrustes rotations, to distribution matching, to Riemannian recentering and manifold transfer — the same geometric toolkit you met for covariance decoders, now applied to learned latents.
Orthogonal-Procrustes alignment: rotate a new session's latent embedding Z_s onto a reference Z_ref, recovering the shared coordinate frame without new labels — the linear core of manifold stabilization.
Does transfer actually help?
Honestly weighed, the evidence is real but bounded. Multi-session and multi-subject pretraining does reduce the labelled data a new session needs and improves few-shot decoding — the effect is reproducible across spikes, EEG, and fMRI models. But the largest, cleanest gains are still within the same lab, task, and hardware. Transfer across very different tasks, species, or electrode types is where returns thin out, and where much of the frontier's real work remains.