Asking for labels only when the room isn't to blame: active learning for ambient electric field wearables

A wrist or ear sensor that measures the body's electric potential can recognize steps, gestures and activities while drawing almost no power, because it listens to fields that are already present. These include the 50/60 Hz radiation from wiring behind the walls, interference from appliances, and static charge built up by footsteps. The same physics that makes these sensors cheap also makes them fragile. The signal changes when the user swaps shoes, walks into a different room, stands near a wall, or when someone switches on a dimmer or a hairdryer. A model trained in one setting degrades in another. The obvious fix is active learning, meaning the device asks the user to label the few windows it is least sure about. In this setting, though, the most uncertain windows will usually be interference bursts that no label can fix.

This note proposes a physical way out. The sensor's own electrode also records the powerline band and broadband interference. Those out-of-band measurements are a free, directly observed description of the electrical environment and of how the body is coupled to it. We use them to split the model's uncertainty into two parts. One part is explained by measured interference, and those windows are never worth asking about. The other part is left over once interference is accounted for, or falls in an electrical coupling state for which no labels exist yet. Only windows of the second kind trigger a label request. The rule is computed in a single forward pass on a microcontroller. The claim is that this rule recovers accuracy after a shift with fewer user labels than entropy-based querying, random querying, or learned uncertainty decomposition. We also claim it does so in the cases where a simple "skip noisy windows" filter fails: shifts that are electrically quiet but new, such as new footwear or a new room.

What the two fields give us, and where the gap is

Ambient electric field sensing has shown strong single-task results with hand-built pipelines. Earable and Wrist-worn Setup for Accurate Step Counting Utilizing Body-Area Electrostatic Sensing counts steps with a bandpass filter and a peak detector on a Qvar electrostatic sensor, reaching 93–94%. Towards Touch-to-Access Device Authentication Using Induced Body Electric Potentials verifies that two devices touch the same body by correlating raw powerline-induced potentials, reaching 98.9% true acceptance at 2% false acceptance. Richer tasks need learned classifiers. Humantenna classifies 12 whole-body gestures at 92.7% across homes. Platypus identifies walkers from ceiling-mounted sensors, but accuracy drops from 94% to 75% as the number of users grows, and the authors note that shoe changes force retraining and that wall coupling creates blind spots. The field survey Body-Area Capacitive or Electric Field Sensing… names subject-dependent continual learning and data scarcity as open problems. These papers motivate the work: they establish that the modality is useful and that it breaks under exactly the shifts we target.

On the active learning side, three lines of work supply methods and one supplies a constraint. ActiveHARNet runs Monte Carlo dropout on an embedded board to personalize inertial activity recognition. Monte Carlo dropout means averaging many randomized forward passes to estimate how unsure the model is. ActiveHARNet raises one failing user from 25% to about 70% with 62 queries, but each uncertainty estimate takes about 14 s. That cost cancels the energy advantage of passive electric sensing. TActiLE makes streaming querying cheap on microcontrollers by combining an entropy threshold with a diversity check, but it was tested only on image benchmarks. Active Domain Adaptation for mmWave-Based HAR separates "this sample looks unlike the training domain" from "the classifier is torn" using evidential networks and Rényi entropy on active radar. It also reports that cross-person shifts are harder to fix by querying than cross-environment shifts. The constraint comes from Privacy-preserving and Uncertainty-aware Federated Trajectory Prediction, which found that picking the maximum-uncertainty sources hurts training because those sources are dominated by irreducible noise.

The closest capacitive work, Agile gesture recognition for capacitive sensing devices, adapts a PCA+KNN gesture model on an ESP32 by adding linear "patches" for error clusters. These patches are triggered after the fact, with no uncertainty model. The authors report that some confusions remain inseparable, and their worst participant group improves only from near 0% to 48.3%. PhyAug uses a physical model to transform source data into a new domain without queries, but it was evaluated on acoustic and seismic sensing. None of these works uses the fact that an electric field sensor measures its own interference and coupling conditions. That fact is the ingredient this proposal adds.

Core Claim

Every window $x$ from the electrode is split into two bands. The low band (below roughly 20 Hz) feeds the activity classifier. The side band yields a small vector $z(x)$ describing the electrical state: powerline amplitude, harmonic ratio, broadband interference energy above the powerline band, and powerline phase stability. From $z(x)$ we take a coupling state $s(x)$, the normalized powerline amplitude and phase stability, which moves with grounding, footwear and distance to wiring. The acquisition score is

$$a(x) = \big[u(x) - \hat g(z(x))\big]_+ + \lambda \, \min_{x' \in L} \lVert s(x) - s(x') \rVert .$$

Here $u(x)$ is the classifier's predictive entropy from a single pass. $\hat g$ is a monotone regressor fit online on unlabeled windows, predicting the entropy expected from interference alone. $[\cdot]_+$ keeps only the positive part. $L$ is the set of already-labeled windows, and $\lambda$ weights the coverage term. The first term is uncertainty that measured interference does not explain. The second term flags coupling states that have no labels, even when they look clean. Windows whose $a(x)$ exceeds a running threshold pass through TActiLE's diversity filter before the user is asked.

The key design choice is that the "noise" part of uncertainty comes from physical measurement, not from a learned aleatoric head that would itself be miscalibrated under shift. Everything above costs one forward pass, a few band-energy computations, and a nearest-neighbor search over a small labeled set.

Research Program

We will build a wrist-worn and an ear-worn prototype. Each pairs a Qvar-style electrostatic front end with a single high-impedance ADC electrode sampled at about 1 kHz, in the manner of TouchAuth, on a Cortex-M4-class microcontroller. The classifier is a small temporal CNN of roughly 30k parameters, matching ActiveHARNet's scale, together with the PCA+KNN pipeline from the Agile gesture paper so that both model families are tested.

Because the file lists no public ambient electric field dataset with active learning results, we will collect one. The target is about 16 participants, 4 rooms with different flooring and wiring density, 2–3 footwear conditions per person, and 8 activity or gesture classes. During recording, a logged relay switches interfering appliances (a dimmer, a hairdryer, a switching power supply) on and off at random. This gives ground-truth labels for which windows are interference-contaminated. The same recording includes the naturally occurring shifts (new room, new shoes, new person) that we expect the coverage term to catch. If the Humantenna, Platypus, or etee capacitive datasets can be obtained, we will replay them as streams to test generality. We make no assumption that they are available.

The analysis has three parts. First, we test whether $\hat g$ actually explains a large share of entropy on the injected-interference windows. Second, we test whether $s(x)$ separates rooms and footwear better than it separates people. Following the mmADA finding, we predict that the coverage term helps environment and footwear shifts much more than person shifts. Third, we measure the energy and latency cost of the full score on the device.

Baselines

The naive reference is the source-trained model with no adaptation, together with uniform random querying at the same label budget. These fix the floor and show whether any selection is worth doing. TActiLE reports only that its methods beat random on image data. No reported result found on Lacuna for random-versus-active querying on ambient electric field data.

The standard active learning competitor is maximum single-pass entropy, run through TActiLE's Info RV and Dual RV streaming rules, with the same thresholds, label budgets and model. Dual RV is the strongest cheap method because it already guards against redundant queries. For latency, the file reports 0.15 s per selection for Info RV on Fashion-MNIST, versus more than 2.5 s for a preemption baseline.

The strongest uncertainty-decomposition competitor is Bayesian querying with learned epistemic uncertainty. We implement BALD over 10 Monte Carlo dropout passes as in ActiveHARNet, and a Rényi-entropy evidential variant following the mmWave active adaptation work. Both get the same label budget and the same hyperparameter search budget as our method. They may run offline, so that we can compare label efficiency even though they do not fit on the microcontroller. The reported reference points come from other modalities: ActiveHARNet raised one HHAR user from 25% to about 70% after 62 queries, and mmADA reached 96.88% on MM-Fi with a 5% target budget versus 89.90% for EADA. No reported result found on Lacuna for either method on electric field sensing.

For the PCA+KNN model family, we compare against the error-patching approach of Agile gesture recognition, triggered with the same number of labeled windows. Its own reported gain, near 0% to 48.3% on the worst group, is on the etee dataset.

Three ablations isolate the key ingredient: the residual term alone ($\lambda=0$), the coverage term alone, and a hard gate that discards windows whose interference energy exceeds a tuned threshold before applying maximum entropy.

Beating the baselines means a higher area under the accuracy-versus-queries curve at budgets of 10–100 labels, averaged over held-out shift conditions. It also means fewer queries to recover within 3 points of in-domain accuracy, at under 5% of the Monte Carlo dropout energy per decision. At minimum we must match BALD's label efficiency while being on-device. Winning on label efficiency is the stronger outcome.

Why not just…?

Why not just throw away noisy windows with a threshold and then use entropy? This is the most serious threat, and it is included as the hard-gate ablation. A gate handles injected appliance interference, but it cannot tell a clean window from a familiar coupling state apart from a clean window from an unseen one, such as new shoes or a different floor. In the latter case the model may even be confidently wrong. If the gate matches the full score on footwear and room shifts, the coverage term adds nothing. We would then report the contribution as a validated physical filter, not a new acquisition rule.

Why not just learn the aleatoric/epistemic split with ensembles or Monte Carlo dropout? Learned decomposition is what ActiveHARNet and the mmWave active adaptation work do. It costs many forward passes, about 14 s per estimate on embedded hardware. Its noise estimate is also learned from source data, so it can be miscalibrated exactly when the environment shifts. The comparison against BALD and the evidential variant measures whether the physically measured split matches or exceeds the learned one on label efficiency. The injected-interference ground truth measures which of the two better avoids querying interference windows.

Why not just correct the shift with physics, as PhyAug does, and skip the queries? PhyAug works when a compact physical transfer model exists. Coupling through footwear, flooring and posture has no such model, and person-specific differences are not physical transforms of the source data. PhyAug's reliance on a few calibration pairs is compatible with our approach, and a PhyAug-style pre-adaptation followed by our querying is a reasonable follow-up. If a simple amplitude-renormalization transform alone closes most of the room gap, we will report it and restrict our claim to footwear and person shifts.

Proposed Experiments

The main experiment uses leave-one-condition-out streams. The model is trained on a set of rooms, users and shoes, then streamed a held-out condition while each method queries under budgets of 10, 25, 50 and 100 labels. We report accuracy curves, queries to recovery, energy and latency per decision on the device, and, most diagnostic, the fraction of queries that land on relay-logged interference windows.

Three results decide the claim. First, our score should send far fewer queries into interference windows than maximum entropy, while matching or beating BALD on label efficiency at a small fraction of its energy cost. Second, on footwear and room shifts the full score should beat the hard gate. This checks that the coverage term, not merely noise rejection, carries the gain. Third, as mmADA's result predicts, the advantage should shrink on new-person shifts. If it does not shrink, the coupling-state interpretation is wrong. If it shrinks to zero, that points to person-level calibration, such as a short enrollment, as the needed complement. Any of these outcomes is informative. The first two together would show that the physics of passive electric sensing can decide which uncertainty is worth a user's label.