Skip to content

Classical Decoding Algorithms

Before deep learning took over BCI, classical linear and probabilistic methods had already built a well-functioning pipeline. These methods remain benchmarks for clinical BCI today: they are computationally cheap, require little training data, and run easily in real time. Understanding them is a prerequisite for understanding modern decoders.

Why study them still. Since 2018, the deep-learning decoders of Chapter 05 have largely won on paper benchmarks, yet the clinical BCIs actually deployed in patients (Pitt arm, BrainGate, Walk Again) almost all still run variants of Kalman + ReFIT. Three reasons: first, clinical sessions only yield tens of minutes of training data per day, where deep models easily overfit; second, FDA approval demands models that are "interpretable, verifiable, and tunable on-site," which linear-Gaussian models satisfy out of the box; third, closed-loop latency budgets are under 50 ms, beyond the reach of neural foundation-model inference. The methods in Chapter 05 stack on top of this foundation, not in place of it.

Recommended reading order. Read the four sections as "read → calibrate → compute": start with Population Vector Algorithm to absorb the foundational insight of Georgopoulos 1984 (broadly tuned single neurons + population voting = precise intent); then move to the linear-Gaussian paradigm of Wiener and Kalman Filtering; next take in ReFIT and Online Calibration and understand why Gilja 2012 is the cornerstone of closed-loop BCI; finally, Linear Discriminant Analysis and Feature Selection extends the same line into the EEG-BCI side via CSP / FBCSP / Riemannian geometry — still the algorithmic mainstream of non-invasive BCI today.

In this chapter:

评论 #