A generative picture of drift
Model the recorded population as a stable, low-dimensional intent signal read out through a slowly-changing lens. Let \mathbf z_t be the latent neural state the user actually controls, and \mathbf W_t the mapping from that state to the measured features — units, band powers, threshold crossings.
A stable latent read out through a drifting map. Here the readout W_t is written as a slow rotation R(theta_t) of a day-0 map W_0 — a stylised stand-in for tuning that reorganises while the underlying signal z_t does not.
This factorisation is the whole game. If \mathbf z_t (the manifold coordinates) is stable and only \mathbf W_t drifts, then drift is a coordinate problem: recover the same latent through a changing basis and a fixed decoder on \mathbf z still works. If \mathbf z_t itself drifts, the problem is far deeper. Much of the field's optimism rests on evidence that, for well-practised movements, the former is often closer to the truth.
The stable manifold
The pivotal empirical result behind drift-robust decoding is that the low-dimensional dynamics of motor cortex during a stereotyped task can be remarkably stable over months to years, even as single-unit tuning appears to change. Work aligning the neural manifold across sessions, via its latent axes, recovered near-constant neural trajectories and let a decoder generalise across time. Manifold stability is the empirical foundation the rest of this track builds on.
Why would tuning drift while the manifold holds? One reading: the brain cares about the computation — the trajectory — not which neurons implement it. Redundant population codes let individual cells wander within a null space that leaves the read-out-relevant subspace intact. Within- versus outside-manifold perturbation studies support this: changes inside the manifold are easy to compensate, changes outside it are hard.
Watching a fixed decoder decay
The clearest way to feel the problem is to freeze a decoder and let the world drift under it.
A stylised model, not a measured law: if error is set by the accumulated misalignment angle theta_t, decode quality falls like its cosine, and under a diffusion model of drift the angle grows like the square root of time. It captures the shape, not the numbers.
The qualitative lesson survives the toy model: a fixed decoder degrades gracefully at first and then sharply, because error grows with the accumulated misalignment. It also means a little correction applied early is worth a lot of correction applied late — a theme that will recur through every adaptive method in this track.
Distinguishing drift from instability in practice
Practically, you diagnose the source before you treat it. The signatures differ. Recording instability shows up as abrupt changes tied to specific channels — bad channels, lost units — often correlated with impedance changes. Representational drift is slower, smoother, and distributed across the population. User-state change tracks task difficulty, fatigue, and time of day.