The recalibration tax
Classic state-space decoders assume the parameters C, Q fit today still hold tomorrow. They do not. Electrodes shift, tuning drifts, and units appear and vanish across days — neural nonstationarity and signal drift. In practice this forces a fresh calibration block at the start of most sessions: minutes of the user's time, every day, before the device is usable. Reducing or eliminating that recalibration tax is the organizing goal of much current motor-BCI research.
Latent dynamics: the state-space model, learned
Instead of hand-specifying cosine tuning and linear dynamics, let the data specify them. Models like LFADS fit a low-dimensional latent state with nonlinear, learned dynamics and a Poisson readout, inferring a smooth neural trajectory that explains the population's spiking. This is the dynamical-systems view of motor cortex made into a decoder.
Stabilized latent-space decoders
The key empirical insight attacking the recalibration tax: while individual neurons drift, the low-dimensional manifold the population moves on is far more stable over time. A stabilized latent-space decoder aligns each day's activity to a fixed reference latent space (an unsupervised, seconds-long alignment) so a single frozen decoder keeps working across months without a fresh supervised calibration.
Discrete and switching states
Not all intent is continuous. Click versus idle, rest versus reach, one phoneme versus the next — these are discrete modes. A hidden Markov model decodes such categorical states, and a switching state-space model does both at once: a discrete mode variable that selects which continuous dynamics are active, so the decoder can, say, hold still crisply and then reach smoothly. This hybrid is increasingly common in high-performance systems.
The offline gold standard: smoothing
A filter uses only the past. For offline analysis — labelling training data, benchmarking, or building intention targets for ReFIT — you can also use the future. The RTS smoother runs the Kalman filter forward, then a backward pass that blends each estimate with what happened next, giving a lower-variance estimate. It is not real-time (it needs future samples), but it is the gold-standard state estimate for anything not in the control loop.
\hat{x}_{k \mid N} = \hat{x}_{k \mid k} + J_k\big(\hat{x}_{k+1 \mid N} - \hat{x}_{k+1 \mid k}\big), \qquad J_k = P_{k \mid k}\,A^{\top} P_{k+1 \mid k}^{-1}The RTS backward recursion corrects each filtered estimate using the smoothed estimate one step ahead.
Offline you can look at the whole recording, so refine each moment's estimate using information from the future. Sweep backward, correcting every filtered estimate with the smoothed one just ahead. This gives the best possible reconstruction.
- \hat{x}_{k \mid k}
- The forward-filtered estimate, using only past and present.
- \hat{x}_{k \mid N}
- The smoothed estimate, using the entire recording.
- J_k
- The smoother gain linking the present to the future estimate.
- \hat{x}_{k+1 \mid N} - \hat{x}_{k+1 \mid k}
- The future correction fed back one step.
The RTS smoother is the offline gold standard — unusable in real time because it needs data from ahead.
Honest open problems
The field has genuinely hard, unsolved questions. Can recalibration go to zero across months and hardware failures, robustly, for every user? How should uncertainty be used for safety — when should a decoder that is unsure refuse to move an arm? How do we unify continuous kinematics, discrete intent, and language-model priors into one estimator (the speech-BCI track leans hard on this)? And how far does any of this generalize across people, tasks, and days without per-subject retraining? State-space thinking gives the vocabulary for these questions; answering them is the current research frontier.