Wearables save battery by keeping a cheap sensor, such as an accelerometer, always on and waking a power-hungry sensor (a gyroscope, heart-rate optics, a microphone or a camera) only when the cheap signal suggests the extra data will matter. The rule that decides when to wake the expensive sensor, which we call the gate, is trained once on pooled data from many people. Yet whether the expensive sensor adds anything depends on the person. One user's tremor makes accelerometer-only recognition ambiguous. Another user's stair climbing is already obvious from the wrist. A population gate spends energy on the second user and starves the first.
We propose to keep the population gate fixed and add a small per-user correction to it. The correction is fitted on the device by a single closed-form ridge regression, a least-squares fit with a penalty that shrinks the weights toward zero, so it needs no backpropagation. Its training targets require no labels: on the occasions when the expensive sensor is on, the device records how much the fused prediction disagrees with the cheap-sensor-only prediction. The battery budget then enters as a cutoff on the corrected score, so the operating point can move with the state of charge without retraining anything.
Core Claim
The claim is narrow and testable. Personalization is valuable for gating because it changes which windows of signal receive the expensive sensor. Shifting a per-user threshold, which only changes how many windows get it, cannot do this. At a matched duty fraction for the expensive sensor (the share of time it is on), a population gate plus a closed-form per-user residual should beat both the population gate and a population gate with a per-user tuned threshold. The gains should concentrate in the users the population gate serves worst. If per-user threshold tuning matches the residual, the claim is false, and the experiments are built to show that cleanly.
Positioning in Prior Work
Cross-modal wake-up is well established, and it motivates this work. Wu et al. use a reinforcement-learned policy on wrist motion to trigger a camera. They process 29% of visual frames while keeping 90.85–97.56% intention accuracy, but they evaluate on three subjects and name cross-user generalization of the policy as unsolved. CoughTrigger wakes an acoustic model from an earbud motion sensor. EgoTrigger wakes a camera from audio cues. All three gates are global and make hard on/off decisions.
Ardywibowo et al. gate sensors using predictive entropy, meaning how uncertain the model's prediction is, together with population-fitted duration priors. DWFS learns per-window sampling rates. Both fix the energy–accuracy weight as a constant scalar, and both are population-trained. We reuse their framing of the energy cost but make the gate user-specific and let the budget vary at run time.
AHAR is an important constraint on our design. Its lightweight decision tree outperforms routing on classifier confidence, which warns us that confidence alone is a weak gating signal. We therefore benchmark against an AHAR-style statistical-feature gate rather than a softmax threshold. CondConv HAR adapts the computation per input but leaves sensing fixed. It also shows that adding expert weights inflates on-device memory, which argues against heavy per-user models.
The method itself comes from adaptive preference fusion in recommender systems. RESUS splits prediction into a shared predictor plus a per-user ridge residual. It solves the residual in closed form through the Woodbury identity, a matrix shortcut that makes the solve cheap when a user has few examples, and reports a 10x speedup over gradient-based meta-learning. Liao et al. and Xu et al. shift weight from the population model to the individual as that person's data accumulates. Every one of these methods personalizes a prediction of preference. We instead personalize a prediction of the value of acquiring a signal, on streaming sensor data under an energy budget.
PRISM separates information that each modality carries alone from information that only appears when modalities are combined. It assumes all modalities are free to obtain. Our disagreement target is a cheap, label-free proxy for "what the expensive sensor adds beyond the cheap one." We do not attempt a full information decomposition on device.
Research Program
Let $z_t$ be a compact feature vector computed from the cheap sensor in window $t$. Let $f_{\text{pop}}(z_t)$ be the population gate's predicted gain from waking the expensive sensor. That gain is trained offline as the reduction in classification loss when the expensive modality is fused in. For user $u$, the personalized score is
$$ g_u(z_t) = f_{\text{pop}}(z_t) + w_u^\top \phi(z_t), \qquad w_u = (\Phi_u^\top \Phi_u + \beta I)^{-1} \Phi_u^\top r_u, $$
where the terms are defined as follows:
- $\phi$ is a fixed feature map of 16 to 32 dimensions.
- $\Phi_u$ stacks $\phi(z_t)$ over the user's windows in which the expensive sensor was on.
- $r_u$ holds the residuals between the observed gain proxy and $f_{\text{pop}}$.
- $\beta$ is the ridge penalty, which pulls a new user toward the population gate.
The observed gain proxy is the divergence between the fused prediction and the cheap-only prediction. The expensive sensor is woken when $g_u(z_t)$ exceeds a cutoff $\tau_B$. That cutoff is set as the running quantile of recent scores matching the duty fraction $B$ that the current battery state allows.
The gate only observes gains when it chose to wake the sensor, so its own decisions bias its training data. We counter this with a small fraction of random wake-ups and reweight those samples by inverse propensity, that is, by the inverse of the probability that each window was sampled. Ablations will report what happens without this correction.
We will build the full pipeline in simulation on multi-sensor datasets with datasheet-derived per-sensor costs. We will then measure the cost of the closed-form solve and update on a microcontroller-class target. The motivation is that DWFS is criticized for offline-only evaluation, and we want to avoid the same gap.
Baselines
The naive reference points bracket the problem. One is always-on fusion, which gives the accuracy ceiling at maximum energy. Another is cheap-sensor-only, which gives the floor. Between them sit fixed duty cycling and uniform downsampling at matched average duty, which Servizi and Stankoski et al. document as losing fine temporal detail.
The strongest existing family is the population-trained gate. We implement three versions on identical classifiers and identical cheap-sensor features:
- Entropy-triggered gate in the style of Ardywibowo et al.. Its reported reference is 79.26% accuracy using 73.42% of sensor data at λ = 0.1 on UCI HAR.
- Learned motion-to-modality policy in the style of Wu et al., with its reported 29% visual data at 90.85–97.56% accuracy on its own datasets.
- AHAR-style decision tree over statistical features, repurposed from deciding when to exit early to deciding when to wake the sensor. AHAR reports 91.57% weighted F1 on Opportunity for early exit, not for sensor wake-up. No reported result was found on Lacuna for population sensor gating on PAMAP2 or DSA, so those numbers will be produced here.
The baseline we consider most dangerous is a population gate with a per-user threshold. It uses the same calibration data as our method but uses it only to tune that user's cutoff. A second personalized baseline fine-tunes the gate per user by gradient descent with the same data and a capped compute budget.
Every method receives the same calibration minutes, the same random-wake fraction, and the same hyperparameter search budget on held-out training subjects. We compare at matched duty fractions.
We will also run two ablations. The first replaces $\phi$ with random features, which tests whether the residual simply adds capacity. The second applies the residual to the classifier instead of the gate.
Beating the baselines means two things. First, higher macro-F1 than the best population gate and than per-user thresholding at duty fractions of 20–50%. Second, a larger improvement on the worst-served quartile of users, with paired tests across held-out subjects.
Why not just…?
Why not just tune a per-user threshold on the population gate? This is the real threat. If users differ mainly in how much they benefit from the expensive sensor, not in when, a scalar shift captures almost everything. We test this directly at matched duty. We also measure rank correlation between the population and personalized scores within each user: values near one would mean the residual only shifts the threshold. If thresholding ties our method, we will report that personalization should be a calibration step, not a model.
Why not just personalize the classifier instead of the gate? A better cheap-only classifier for a given user might remove the need to wake the expensive sensor at all. Our classifier-residual ablation applies the same closed-form residual to the classifier. Both arms are evaluated alone and combined, so the experiment shows which location gives more accuracy per joule.
Why not just fine-tune the gate on-device? RESUS reports a 10x speed advantage for closed-form residuals over gradient-based meta-learning. With only minutes of calibration data, fine-tuning also risks overfitting. We include fine-tuning under a capped budget. If it wins on accuracy, we report its energy and latency cost on the microcontroller target so the trade is explicit.
Why not just train a budget-conditioned policy, feeding the battery level to a DWFS-style agent? That addresses the fixed-λ problem but not individual differences. Our quantile cutoff handles the budget independently of personalization. We will cross a budget-conditioned population policy with our residual to separate the two effects.
Proposed Experiments
Datasets and sensor tiers.
- UCI HAR (30 subjects): accelerometer as the cheap sensor, gyroscope as the expensive one. This gives enough subjects to study the tail of the user distribution.
- PAMAP2: wrist IMU as the cheap sensor; chest and ankle IMUs plus heart rate as expensive.
- DSA: one inertial unit as the cheap sensor; the rest as expensive.
- OPPORTUNITY: cheap and expensive body sensors on complex gestures.
- Wu et al.'s Daily Intention data: accelerometer as cheap, camera as expensive. This is a small-cohort check that also matches a published operating point.
Protocol. We use leave-one-subject-out evaluation. Each held-out subject's first calibration window, from 2 to 15 minutes swept, provides $\Phi_u$ and $r_u$, with the remainder used for testing.
Metrics.
- Macro-F1 against expensive-sensor duty fraction, summarized as the area under that curve.
- Per-user gain over the population gate.
- Error in the windows just after an activity transition, the lag that Ardywibowo et al. identify.
- How far the realized duty fraction deviates from the battery-driven target.
- Measured energy and time for the ridge update on the microcontroller.
Decisive checks. Three results settle the claim:
- The residual beats per-user thresholding at matched duty on at least three of the four multi-subject datasets.
- Its gains grow as the population gate's per-user error grows.
- The label-free disagreement target recovers most of the benefit of an oracle target computed from true labels. If it does not, the method needs labels in practice and the claim narrows.
References
- Anticipating Daily Intention Using On-wrist Motion Triggered Sensing
- Coughtrigger: Earbuds IMU Based Cough Detection Activator Using An Energy-Efficient Sensitivity-Prioritized Time Series Classifier
- EgoTrigger: Toward Audio-Driven Image Capture for Human Memory Enhancement in All-Day Energy-Efficient Smart Glasses
- Adaptive Activity Monitoring with Uncertainty Quantification in Switching Gaussian Process Models
- Learning Datum-Wise Sampling Frequency for Energy-Efficient Human Activity Recognition
- AHAR: Adaptive CNN for Energy-efficient Human Activity Recognition in Low-power Edge Devices
- Real-time Human Activity Recognition Using Conditionally Parametrized Convolutions on Mobile and Wearable Devices
- RESUS: Warm-up Cold Users via Meta-learning Residual User Preferences in CTR Prediction
- Continual Human-in-the-Loop Optimization
- A User-Oriented Webpage Ranking Algorithm Based on User Attention Time
- PRISM: Personalized Recommendation via Information Synergy Module
- Transport behavior-mining from smartphones: a review
- Recognizing Activities of Daily Living using Multi-sensor Smart Glasses