The ledger: demonstrated vs emerging vs open
An honest state-of-the-field, sorted into three columns. Demonstrated: self-recalibrating cursor and typing over months; manifold-stabilised control across sessions in animals and limited human data; ReFIT and CLDA as robust, deployed slow-drift trackers. Emerging: cross-subject foundation models and contrastive embeddings; unsupervised zero-shot transfer on benchmarks; continual learning that adds sessions without full retraining. Open: calibration-free, high-degree-of-freedom decoding over years through major recording changes; provably safe autonomous adaptation; principled handling of activity that leaves the known manifold.
The stability–plasticity dilemma, sharpened
Every lifelong decoder must both remember old, still-valid structure and adapt to new drift — and these pull against each other. Push adaptation and you risk catastrophic forgetting: yesterday's competence overwritten by today's update. Freeze to preserve it and you cannot track drift at all. Continual learning is the sub-field aimed squarely at this tension.
An elastic-weight-style penalty: keep parameters important to past sessions (high Fisher information F_i) from moving far, while still fitting today's loss. One principled line of defence against catastrophic forgetting.
Regularisation that protects important parameters, rehearsal or replay that re-trains on stored or generated old data, and modular or expandable architectures are the main strategies. None is a settled winner for neural data — and all inherit an extra pain unique to BCI: you often cannot store raw neural data for long, for mental-privacy reasons. Most continual-learning methods were never designed under that constraint.
The co-adaptation ceiling and evaluation traps
When both user and decoder adapt, there is a joint optimum they can settle into — but also a ceiling beyond which pushing decoder adaptation merely fights user learning. Where that ceiling sits, and whether current systems are near it, is not well characterised. It is one of the genuinely open questions this whole track circles.
Beware two traps when reading claims. First, results tuned per-day are not calibration-free, however the abstract phrases it. Second, a method that shines on a stable motor task may say nothing about a drifting cognitive one — motor, speech, memory, and mood drift very differently. Always match the claim to the setting.
Where it's going
A plausible reading of the next decade: stabilisation and self-supervised or foundation models converge, so a new session mostly aligns rather than retrains; on-implant edge adaptation moves the loop onto the device for privacy and latency; and evaluation shifts to multi-month, multi-site held-out benchmarks that reward robustness over peak accuracy. The prize is a decoder the user never has to think about — one that quietly re-aligns overnight and simply works in the morning.
The deeper open question is theoretical: *what is stabilisable in principle?* Drift within a stable manifold is compensable; drift that adds genuinely new structure may not be, without new information from somewhere. Knowing which regime a given signal lives in — motor, speech, memory, mood — is the research that will decide how close each application can get to 'never recalibrate.' That, honestly stated, is the frontier this track leaves you standing on.