Dynamics make better decoders
The whole track pays off here. A decoder that models the latent state and its dynamics denoises single trials and interpolates through dropped spikes far better than an instantaneous read-out. That is the through-line of motor BCI's history: from population vector, to the velocity Kalman filter, to intention-corrected ReFIT — each a better model of a smooth, dynamical latent state rather than a better guess at each neuron's momentary rate.
Stability: decoding a manifold that (mostly) doesn't move
Single-neuron tuning drifts and electrodes come and go, which forces daily recalibration if you decode neurons directly. But the manifold itself is comparatively stable: the low-dimensional structure persists even as individual cells change. Stabilized latent-space decoders exploit this by aligning each day's activity to a reference latent space, so a decoder trained once keeps working with little or no recalibration.
LFADS and inferring dynamics per trial
LFADS is where deep learning meets the dynamical-systems hypothesis. It is a sequential autoencoder: a recurrent generator embodies the assumed dynamical system, and inference recovers, for each trial, the initial condition and inputs that best reproduce the observed spikes. The payoff is remarkably clean single-trial trajectories and, often, improved kinematic decoding — a learned prior that the population is a smooth dynamical system doing much of the denoising for you.
BCI as a causal probe of dynamics
Population dynamics is not only something BCI exploits; BCI is also how we test it causally. Perturb the decoder mapping and subjects learn fast when the new mapping stays within the existing manifold, but slowly — or not at all — when it demands off-manifold activity the network cannot easily produce. That is direct causal evidence that the manifold constrains learning. Alongside it, BMI skill learning shows cortex forming stable, reproducible new activity patterns dedicated to control — the manifold is not a passive readout but something the brain actively shapes.
Open problems (an honest list)
The dynamical view reshaped the field, but plenty is unresolved. The frontier is not the polished result on averaged data; it is single-trial, causal, closed-loop performance that holds for months.
- Mechanism or description? Whether internal dynamics are the causal mechanism or an elegant re-description of the same data is still genuinely debated.
- Nonlinear geometry. Real manifolds may be curved; linear tools (PCA/FA/jPCA) can mislead, and better nonlinear latent models are needed.
- Cross-area and cross-subject transfer. Manifolds differ across regions and people, so a decoder trained on one rarely transfers to another.
- Dynamics versus black-box deep decoders. End-to-end networks sometimes match or beat dynamics-based decoders — is the dynamical prior necessary, or just a good regularizer?
- Low-latency single-trial inference. The richest dynamical models remain hard to run causally, in real time, at the latencies closed-loop control demands.