Waking Electrodes Only When the Speller Is Unsure

Wearable brain-computer interfaces (BCIs) that read scalp electrical activity (electroencephalography, or EEG) run on batteries. Every electrode left on draws power for its amplifier, its analog-to-digital converter, and the radio that streams its samples. Yet the signals these devices record are very noisy, so the usual move is to add more electrodes and run bigger decoders. Wireless engineering faced the same tension with large antenna arrays and found something clean. Energy efficiency, measured in bits delivered per Joule, rises and then falls as channels are added. Beyond a "sweet spot," each extra hardware chain costs more power than the information it contributes. This proposal asks whether an EEG speller has the same bell-shaped curve. It also asks whether a speller can sit at the peak of that curve one stimulus at a time: power up extra electrodes only while the decoder's belief about the intended letter is still uncertain, and shut them down once it is confident.

The concrete claim is this. In a recursive Bayesian speller, the user's intended symbol is inferred by accumulating evidence over many brief flashed stimuli. The value of another electrode depends on how uncertain the running belief currently is. A controller that chooses how many electrodes to power for the next stimulus by trading expected information gain against circuit energy should spend fewer Joules per correctly typed symbol than any fixed electrode subset at the same accuracy. This holds even when the fixed subset is the best one a learned selector can find for that energy budget. The new ingredient is not channel selection, and it is not energy-aware selection with a fixed budget. It is closed-loop spatial power gating driven by the decoder's own posterior, evaluated in bits per Joule rather than in accuracy at an arbitrary channel count.

What exists and where the gap is

The wireless side provides the measurement idea and the stopping rule. Designing Multi-User MIMO for Energy Efficiency models circuit power as static power, plus per-chain power, plus digital processing terms. It shows that energy efficiency peaks at a finite array size. Stepwise Transmit Antenna Selection in Downlink Massive Multiuser MIMO adds antennas greedily until the marginal rate gain no longer pays for the chain's power. It finds a peak near 24 of 128 antennas, and the optimal count stops growing as the available pool grows. We borrow both the bits-per-Joule objective and the marginal-gain-versus-cost stopping logic. These papers assume Gaussian noise and selections that stay fixed across a coherence interval. We apply the logic to non-stationary neural evidence and re-decide at every stimulus.

The neural side supplies the decoders and the evidence that channels can be cut. End-to-end learnable EEG channel selection with deep neural networks learns a fixed subset of $K$ electrodes using a differentiable "concrete" selector, a relaxation that lets gradient descent choose discrete channels. With 8 of 128 channels it stays within 5% of full-array motor accuracy. It treats $K$ as an arbitrary budget, has no power model, and never changes the subset at run time. This paper is both our strongest static competitor and a source of channel orderings.

Recursive Estimation of User Intent from Noninvasive Electroencephalography using Discriminative Models supplies the application. A discriminative network scores each flashed stimulus, a recursive Bayes rule updates symbol probabilities, and typing stops once confidence crosses a threshold. On a dataset of over a million stimulus presentations, it reports 78.2% balanced accuracy for a one-dimensional convolutional network and 1.153 bits per symbol for a two-dimensional one. Its posterior is exactly the uncertainty signal our controller needs. That paper never asks what the electrodes cost.

Precise and low-power closed-loop neuromodulation through algorithm-integrated circuit co-design shows that event-driven gating pays off in silicon. A tiny binary network wakes an 8-bit network, and the whole chip averages 142.9 µW. Its gating is temporal and acts only on the digital stage, over a fixed 16-channel front end. Our gating is spatial and acts on the analog and radio stages. ENIGMA motivates the direction from the other end: a 165× smaller network does well on a cheap 32-channel headset. Its reported savings are in parameters, though, not in measured or modeled power.

Core Claim and Mechanism

Train one decoder that works on nested electrode subsets. Electrodes are ordered once by a concrete selector. The decoder is then trained with random truncation of that ordering, so a single network accepts the top $L$ channels for any $L$ from 1 to $M$. For each level $L$, we measure on held-out calibration data how sharply the decoder separates target from non-target stimuli. This gives a lookup of the expected entropy reduction of the symbol posterior for any current posterior and level.

At each stimulus $t$, the controller picks

$$L_t = \arg\max_{L} \; \Big( \mathbb{E}\big[H(p_t) - H(p_{t+1}) \mid L\big] - \lambda\,\tau\,(P_0 + L\,P_{\text{ch}}) \Big),$$

where $p_t$ is the current posterior over symbols and $H$ is its entropy. $\tau$ is the stimulus duration. $P_0$ is always-on power, covering the processor and inference; the 142.9 µW chip above anchors one setting. $P_{\text{ch}}$ is the per-channel cost of amplifier, converter, and radio. $\lambda$ converts Joules into bits. The episode ends when no level has positive net value or the usual confidence threshold is reached. This mirrors the stepwise antenna stopping rule.

Channels switch only at stimulus boundaries. A freshly woken channel throws away its first $s$ milliseconds as settling time, and we sweep $s$. This is the main physical risk, since EEG amplifiers and electrode offsets need time to settle. We predict that adaptive gating gains most when $P_{\text{ch}} L$ dominates $P_0$, and that the gain vanishes when fixed overhead dominates. That makes a falsifiable regime map rather than a single headline number.

The file gives no per-channel power figures for EEG hardware, so we do not assert any. Instead, $P_0$, $P_{\text{ch}}$, and $s$ are swept over ranges taken from component datasheets. Results are reported as functions of the ratio $P_{\text{ch}}/P_0$.

Research Program

The first stage is purely measurement. On the RSVP speller dataset from the recursive-estimation paper, we compute energy per correctly typed symbol, and bits per Joule, for every static subset size under the power model. The goal is to establish whether the neural curve is bell-shaped. We also test whether its peak saturates as the candidate pool grows, as the antenna count did. To do this, we restrict the pool to random spatial subsets of growing size and re-find the optimum. Either outcome is informative. No reported bits-per-Joule result for any EEG decoder was found on Lacuna.

The second stage builds the nested decoder and the controller and compares them against the baselines below. The third stage tests generality on BCI Competition IV dataset 2a, a 22-channel, four-class motor imagery set. Here the "posterior" is the class belief accumulated over sliding windows within a trial, using Filter Bank Common Spatial Pattern as a classical decoder alongside a small convolutional one.

Baselines

The naive reference is the existing recursive speller with all electrodes always on and confidence-threshold stopping, exactly as in Recursive Estimation of User Intent. We rerun it with the 2D network that reported 1.153 bits per symbol, matching that figure on our splits before attaching the power model. This point anchors the accuracy axis.

The strongest competitor is the best static subset per energy budget. We train the concrete selector of End-to-end learnable EEG channel selection separately at every $K$, each with its own decoder and the same confidence stopping. We then take the lower envelope of energy per symbol versus symbol accuracy. This frontier is competitive because the selector already captures synergistic electrode pairs, and per-$K$ retraining gives it more capacity than our single nested network. Every method gets the same subjects, splits, decoder architecture family, and hyperparameter search budget. Beating it means lower energy per symbol at matched symbol accuracy and matched average typing time, with confidence intervals across subjects. The gain must hold across at least the regime where per-channel cost exceeds fixed cost.

A greedy stepwise wrapper adapted from Stepwise Transmit Antenna Selection adds channels until the marginal accuracy gain per Joule falls below cost. It represents the direct wireless transplant.

A temporal-only gate powers the full array during each stimulus's informative window and turns it off between windows, in the spirit of the co-design chip. This isolates whether spatial adaptivity adds anything beyond duty-cycling.

Two ablations isolate the key ingredient, the posterior-driven choice. The first keeps the nested decoder but uses a fixed schedule of $L$ by stimulus index, ignoring the posterior. The second uses the controller with a random rather than learned channel ordering.

Why not just…?

Why not just pick the best fixed subset for the energy budget? This is the main threat. If the value of an electrode barely depends on how uncertain the posterior is, the static frontier will match us. The comparison against the per-$K$ retrained concrete selector settles it directly. We will report a tie as a negative result, since knowing that static selection suffices is itself useful for hardware design.

Why not just stop typing earlier, using the confidence threshold already in the speller? Energy per symbol is power multiplied by time. Lowering the stimulus count already saves energy, and the recursive speller does it for free. Every one of our baselines therefore includes optimized confidence stopping, with the threshold tuned per method on the same budget. The question becomes whether varying the electrode count adds savings on top of varying the stimulus count. The fixed-schedule ablation shows whether any gain comes from posterior awareness or merely from using fewer channels on average.

Why not just compress the decoder and run it on a low-power chip? The co-design chip shows the digital side can fall to roughly a hundred microwatts. Once that happens, the per-channel analog and radio costs make up most of what is left, and compression does nothing to them. Our regime sweep shows this explicitly. When $P_0$ dominates, we expect no advantage for gating, and we will say so.

Why not just duty-cycle the whole array between stimuli? The temporal-only baseline tests this. It is complementary rather than competing, so the final system combines both. The claim stands only if the spatial gate lowers energy below the temporal gate alone.

Proposed Experiments and Decisive Checks

On the RSVP dataset, the primary metrics are energy per correctly typed symbol, bits per Joule, symbol accuracy, and typing time, reported across the $P_{\text{ch}}/P_0$ and settling-time sweeps. The secondary check is BCI IV 2a, where the reference kappa reported for filter-bank spatial patterns is 0.569. Here we report energy per correct trial decision at matched kappa.

Three results are decisive. First, the static curve must show an interior peak. If energy efficiency rises monotonically with channel count, there is no sweet spot and the whole premise fails. Second, the adaptive controller must sit strictly below the best static frontier and the temporal-only gate at matched accuracy. Third, the posterior-blind schedule must lose most of that advantage. If the advantage survives realistic settling times, the result argues for wearable front ends with per-channel power gates. If it dies once settling exceeds a few tens of milliseconds, that tells hardware designers exactly which amplifier property to improve.