From filters to deep learning

Goal of this page. nSTAT’s decoders — the point-process adaptive filter and its relatives — are linear, model-based estimators with a long, proven track record. Modern brain–computer interfaces increasingly add deep-learning decoders on top. This page draws the line from one to the other: what carries over, what changes, and where to go to learn the rest. It deliberately does not reimplement those methods — it places nSTAT in the arc.

Glossary jumps: PPAF · CIF · Kalman filter / smoother · hybrid PP filter (PPHF) · encoding vs decoding

What nSTAT’s decoders already give you

The decoding page builds up the point-process adaptive filter (PPAF) and hybrid filter (PPHF). Stripped to essentials, every one of these decoders is a predict/correct loop:

  1. Predict the latent state forward with a dynamics model (a random walk, or linear kinematics).

  2. Correct that prediction using the spikes just observed, weighted by each neuron’s fitted tuning (its conditional intensity).

This is the point-process cousin of the Kalman filter, and it has real virtues: it is interpretable (every term is a tuning curve or a dynamics coefficient), data-efficient, and it reports calibrated uncertainty (the credible band). For many BCIs a linear filter is still a strong, hard-to-beat baseline.

Its limits are equally clear. The encoding is linear in the state (through the link function), the dynamics are linear, and the model is specified by hand. When tuning is strongly nonlinear, non-stationary, or high-dimensional, a learned function can do better.

What deep-learning decoders change

Deep-learning decoders keep the goal — map spikes to a stimulus, intention, or movement — but replace hand-specified pieces with learned ones:

Piece

Classical (nSTAT)

Deep-learning decoder

Encoding / tuning

fitted CIF, linear-in-state

learned nonlinear function

Temporal dynamics

linear (random walk / kinematics)

RNN / LSTM / temporal CNN / transformer

Model specification

written by the analyst

learned from data

Uncertainty

closed-form posterior covariance

needs explicit modeling (often absent)

Data appetite

modest

large

Concretely: recurrent-network decoders and sequence models often outperform a Kalman filter on rich motor BCIs (Glaser et al. 2020); the high-performance handwriting BCI paired a recurrent decoder with intracortical spikes (Willett et al. 2021); and sequential autoencoders such as LFADS infer single-trial population dynamics that a linear filter cannot capture (Pandarinath et al. 2018). The trade is the one in the table: more flexibility and accuracy, at the cost of more data, less interpretability, and uncertainty you must add back deliberately.

How to think about the jump

Three ideas carry all the way from the PPAF to a deep-learning decoder, and keep you oriented:

  • Encoding still underlies decoding. Whether the tuning is a fitted CIF or a learned network, decoding inverts an encoding model. The encoding GLM is the concept that never goes away.

  • Goodness-of-fit still matters. A flexible model can overfit; the discipline of checking the model and holding out data is more important as capacity grows, not less.

  • Uncertainty does not come for free. The PPAF hands you a calibrated band; most deep decoders do not. Knowing how confident a decode is remains essential for a safe BCI — see uncertainty.

Start where nSTAT is strong — interpretable, well-calibrated, model-based decoding — and you will understand exactly what a learned decoder is buying you, and what it is quietly giving up.

Applying nSTAT — why calibration is not optional. The clearest place this matters is a closed-loop clinical device. Adaptive deep brain stimulation gates therapy on a decoded brain state (Little et al. 2013); a read-out that cannot say how confident it is should not drive stimulation. nSTAT’s model-based decoders hand you that calibrated uncertainty for free — see Rhythmic firing and the clinical microelectrode for the applied arc, and uncertainty for the band.

See also