The instability problem
The single largest barrier to daily, unsupervised use is that the recording is not stationary. Across hours and days, individual units are gained and lost, their tuning shifts, and baseline firing drifts — neural nonstationarity. A decoder tuned yesterday degrades today, and the naive fix — recalibrate every morning — burns the user's time and needs supervision they may not have at home.
Stabilizing through the manifold
The key discovery: although single neurons come and go, the low-dimensional neural manifold that the population collectively traces during movement is far more stable than any individual unit. This manifold stability gives a fix. Learn a reference set of latent factors once; on each new day, find an alignment that maps today's neural activity onto that same latent geometry, and decode from the stabilized latents. The decoder above the latents never has to be retrained.
\mathbf{u}_t = R\,\mathbf{l}_t \quad \text{s.t.} \quad p(\mathbf{u}_t) \approx p_{\mathrm{ref}}(\mathbf{u})A stabilized latent decoder: learn an alignment R that maps today's latent factors \mathbf{l}_t so their distribution matches a fixed reference — no labels, no daily retraining of the decoder itself.
Instead of retraining the decoder every day, learn a small alignment that maps today's latent brain activity so its distribution matches a fixed reference from before. The old decoder then still works — no labels, no daily retraining.
- \mathbf{l}_t
- Today's latent factors — the low-dimensional neural state.
- R
- The alignment map applied to them.
- \mathbf{u}_t = R\,\mathbf{l}_t
- The realigned factors fed to the fixed decoder.
- p(\mathbf{u}_t) \approx p_{\mathrm{ref}}(\mathbf{u})
- The constraint that their distribution matches the reference.
If today's activity is rotated relative to last month's, R rotates it back into register — a stabilized latent decoder.
This is powerful because the alignment can be found unsupervised — matching distributions, not requiring the user to perform a labeled calibration task — which is exactly what home use demands. It is one of the most promising routes to a stabilized latent decoder that just works when you switch it on.
Self-recalibration
The other route keeps recalibrating, but without asking the user. Self-recalibration manufactures its own labels: assume that during ordinary use the user intends to move toward whatever target they are acquiring (the same intention trick from ReFIT), and use those pseudo-labels to nudge the decoder continuously. Push further and you can use the brain's own error signals: an error-related potential or reward-modulated signal tells the decoder when it just did well or badly, turning control into a reinforcement-learning loop with no explicit calibration at all.
Plasticity as an ally, not just a nuisance
The user is not a passive signal source — their cortex learns the decoder, and you can design for that. A landmark result: users readily learn remappings that stay within the existing neural manifold, but learn much more slowly when the required activity pattern lies outside it. This within- vs outside-manifold asymmetry says the manifold is a genuine constraint on plasticity, and it argues for a decoder the user can master and then keep: a fixed decoder plus consolidated BMI skill learning can be more stable than one that keeps adapting under the user.
Where performance actually stands
An honest snapshot. High-performance intracortical cursor and handwriting demos are real and genuinely fluent — but achieved in a small number of implanted participants, with expert setup, and typically through a percutaneous connector while fully wireless systems are only now maturing. Performance drifts over sessions; generalization across tasks and across days is partial; and channel counts, though rising fast with probes like Neuropixels, are still tiny next to the cortex.