Waking the Gyroscope Only When It Will Change the Answer

Wearable activity and gait monitors usually record every sensor all the time: accelerometer, gyroscope, magnetometer, sometimes pressure insoles, all sampled continuously and fed together into one classifier. That is wasteful. The cheapest sensor often settles the question by itself. In the classic smartphone benchmark, adding the gyroscope to the accelerometer bought about 7 points of accuracy overall, and that gain sits in the dynamic, rotational activities, while lying down was already recognized perfectly (A Public Domain Dataset for Human Activity Recognition using Smartphones). Early exiting is the deep-learning technique of attaching small prediction heads partway through a network and stopping once one of them is confident enough. It looks like the obvious tool here: run the accelerometer stage, and power up the gyroscope or insole only when that stage is unsure.

This note argues that the obvious version asks the wrong question. Low confidence from the accelerometer does not mean the gyroscope will help. Sitting versus standing is the least reliable distinction in that same benchmark, at 88% recall (Anguita et al.), and a waist gyroscope sees almost no rotation in either posture. A confidence threshold would therefore keep waking the gyroscope for exactly the windows where it is useless. We propose to train a tiny controller that predicts the marginal benefit of acquiring the next modality, meaning the probability that the fused model is right where the cheap stage is wrong. The controller reads the cheap stage's predictions over the last few windows. Its decision is whether to switch the expensive sensor on for the next window, and the quantity it saves is sensor-on time, not just floating-point operations.

Core Claim

In streaming multi-sensor wearables, a stopping rule that predicts whether an additional modality will change the prediction from wrong to right, using the history of cheap-stage outputs across overlapping windows, reaches the accuracy of always-on fusion with substantially less expensive-sensor duty cycle than a temperature-calibrated confidence threshold at the same accuracy. The gap should be largest on classes that are ambiguous to every modality, which confidence rules treat as hard and benefit rules treat as hopeless. The gap should also be largest on segments with sensor dropout.

Two results from the early-exit literature motivate this framing, and one result from the wearables literature constrains it. Rethinking Calibration for Early-Exit Neural Networks shows that a calibrated exit only measures its own accuracy and knows nothing about whether later stages can succeed. As a result, compute is wasted on "futile" inputs that no exit gets right. The paper's fix is a two-layer meta-classifier over top-k confidences and the exit trajectory, and we reuse that technique. Our changes are that the "later stage" is a different physical sensor, that the cost is sensor power, and that the trajectory runs across time rather than across depth. Hold That Exit proves that irreversible threshold policies cannot guarantee a constant-factor approximation of the best accuracy-latency trade-off, and that policies allowed to recall earlier outputs can. We read the streaming setting as a natural form of recall: a fused prediction made a few overlapping windows ago can be held instead of recomputed.

The constraining result is AHAR. It explicitly abandoned in-network confidence exits for wearable activity recognition because intermediate confidence was unreliable. In their place it uses a decision tree on raw-signal statistics that routes each window to a one-block or two-block CNN. We take that negative result seriously. Our controller is also external and cheap, but it predicts benefit from model outputs over time rather than difficulty from raw statistics. It also chooses which sensors to power, not how many convolution blocks to run.

Research Program

The model is a staged fusion network. Stage one encodes only the accelerometer stream of a window and has its own classification head. Stage two adds the gyroscope (and magnetometer where available) to the stage-one features. Stage three, where the dataset has it, adds pressure insoles or additional body sites. Later heads are trained with the self-distillation losses of Be Your Own Teacher: cross-entropy, a divergence toward the final head's distribution, and feature hints. The point is that the cheap heads stay competitive and their confidences carry information about the full model. Each stage is also trained with random modality dropout so that later stages tolerate missing inputs.

The controller is trained on held-out subjects. For window $t$ and stage $k$, its target is

$$g_{t,k} = \mathbb{1}[\hat y^{(k)}_t = y_t] \cdot \mathbb{1}[\hat y^{(k-1)}_t \neq y_t],$$

which equals one exactly when acquiring modality $k$ turns a wrong prediction $\hat y^{(k-1)}_t$ into a correct one $\hat y^{(k)}_t$. The controller's input covers the stage-$(k-1)$ outputs for windows $t-H$ through $t-1$, the most recent fused output and how many windows ago it was computed, and a small change statistic on the accelerometer features. It outputs a probability $\hat g_{t,k}$, and modality $k$ is powered for window $t$ when $\hat g_{t,k}$ exceeds a cost-dependent threshold. Otherwise the system returns either the stage-$(k-1)$ prediction or the held fused prediction, and a second controller output chooses between the two.

The decision for window $t$ is made before window $t$'s expensive data exists, so the setup respects the fact that a gyroscope cannot be read retroactively. The 50% window overlap standard in these benchmarks makes the previous window an informative proxy.

Energy is reported two ways. The hardware-agnostic metric is the fraction of time each expensive sensor is on. The second is a parameterized energy model that takes per-sensor active power from the datasheets of the devices used in each dataset and adds the controller and network compute. No sensor power figures appear in the file, so none are asserted here.

Baselines

The naive reference points are always-on full fusion, which sets the accuracy ceiling at full sensor cost, and accelerometer-only, which sets the floor at minimum cost. Before any other experiment, we measure the gap between the two on every dataset. If that gap is within seed noise on a dataset, there is nothing to save there, and we will say so.

The strongest competitor on the core question is a temperature-calibrated confidence cascade in the style of The Right Tool for the Job, which tunes post-hoc temperatures on intermediate heads so that thresholds track accuracy. It is the strongest because it is simple, well understood, and exactly what a practitioner would deploy. We run it on the identical staged, self-distilled backbone, so that the only difference is the stopping signal, with thresholds swept over the same validation subjects. A second variant combines it with the multi-window ensembling of Confidence-Calibrated Human Activity Recognition. That paper reports at least a 40% cut in expected calibration error across PAMAP2, UCI HAR, WISDM and Skoda, so this variant tests whether better calibration alone closes the gap. No reported early-exit or sensor-duty result for such a cascade on these benchmarks was found on Lacuna.

The strongest existing wearable method is AHAR's output block predictor. On Opportunity it reports weighted F1 of 91.57% for the adaptive model versus 91.79% for its static baseline, with 12.32% energy savings relative to that baseline (AHAR). That saving is compute energy with all sensors on. We reimplement its statistical-feature decision tree, retarget it to choose sensor sets rather than blocks, and give it the same training subjects and tuning budget.

The learned-gating representative is DFTerNet's Bernoulli sensor-fusion weights. DFTerNet reports up to about 5% accuracy gains on Opportunity and PAMAP2 but learns per-task rather than per-window gates. We adapt it by thresholding the gates per window to obtain duty cycles.

We also run a hold-last-prediction baseline that reuses the last fused output until an accelerometer change detector fires.

Three ablations isolate the key ingredients. The first replaces the benefit target $g$ with plain stage-$(k-1)$ correctness, which turns the controller into the exit-failure predictor of Rethinking Calibration. The second removes the temporal history ($H=0$). The third removes the hold option.

Beating a baseline means matching always-on weighted F1 within one standard deviation across five seeds and leave-subjects-out folds, at a lower expensive-sensor duty cycle. On Opportunity the target is to equal or exceed AHAR's 91.57% weighted F1 while the gyroscope and extra body sites are off for a majority of windows.

Why not just…?

Why not just threshold temperature-calibrated confidence on the accelerometer head? This is a real threat, and it may tie on datasets where every accelerometer mistake is fixable by the gyroscope. The argument against it is structural. Temperature scaling rescales confidence monotonically, and Rethinking Calibration shows such calibration cannot see whether later stages will succeed. The UCI HAR sit/stand confusion (Anguita et al.) is a concrete case where low confidence and low benefit coincide. The decisive measurement is per-class duty cycle at matched accuracy. If the confidence cascade does not spend disproportionate gyroscope time on sit/stand and on dropout segments, our premise is wrong.

Why not just hold the last prediction until the accelerometer signal changes? Daily activities persist for long stretches, so this trivial rule may capture most of the savings. We treat it as a component rather than a rival: the controller already sees a change statistic and can hold. If the hold-plus-change-detector baseline matches the full controller, the honest conclusion is that temporal persistence, not benefit prediction, is the useful ingredient. The ablation without the benefit target is designed to expose exactly this.

Why not just use AHAR's external statistical gate, or a classical model like XGBoost? AHAR shows such a gate preserves F1, and classical models are competitive on some wearable tasks (Human Activity Recognition Dataset for Pedestrians with Mobility Disabilities). However, a gate on raw-signal statistics estimates difficulty, not whether a particular extra sensor would fix the error. It also cannot reuse past fused outputs. Hierarchical, Multi-Sensor Based Classification of Daily Life Activities further notes the cost of hand-crafting many features per window on low-power nodes. The retargeted AHAR gate run under identical conditions answers this directly.

Why not just quantize or use dynamic kernels? Methods like DFTerNet's 2-bit weights or CondConv cut compute per window but leave every sensor sampling continuously. They are orthogonal to this proposal, and we expect them to compose with it rather than compete.

Proposed Experiments

The main study uses three datasets. UCI HAR provides a clean accelerometer-then-gyroscope staging with a known 7-point gyroscope gain. PAMAP2 stages across body sites (wrist first, then ankle and chest). Opportunity allows direct comparison with AHAR and DFTerNet. The Leutheuser four-node SHIMMER data serves as a further site-staging check, against its reported 89.6% mean classification rate. WearGait-PD, with 13 IMUs, pressure insoles, and an 18% sensor dropout rate, is the stress test for futile windows. No reported classification result for it was found on Lacuna, so there we compare only against our own always-on model.

The main figure plots accuracy against expensive-sensor duty cycle for every method. The decisive checks are three. First, at matched accuracy, the benefit controller should use the gyroscope substantially less than the calibrated confidence cascade on sit/stand windows. Second, on WearGait-PD windows with injected or natural dropout, the controller should decline to wake additional sensors while the confidence cascade wakes them. Third, the temporal-history ablation should account for a measurable share of the savings on streaming data but not all of it.

If the benefit target shows no advantage over correctness-based failure prediction, or if hold-plus-change-detection matches the full method, we will report that. Either outcome would tell practitioners which simpler rule to deploy.