acceptodds
Under review as a conference paper at ICLR 2027

Benchmark-Relative Temporal-Resolution Sufficiency for Wearable Time-Series Learning

Abstract

A wearable system must choose temporal resolution before a downstream representation can exploit the signal. Conventional resolution sweeps rank tested configurations, but they do not define when a cheaper rate–representation pair preserves enough utility to remain acceptable. We formulate this as *benchmark-relative temporal-resolution sufficiency*. A declared tolerance first defines an admissible set relative to the best tested reference; deployment costs are considered only inside that set. This yields a single decision chain from *admissibility* to *preference* to *validity*. We characterize how the admissible set changes with tolerance, reference strength, target environments, and budgets; give an exact finite-grid extraction algorithm; separate oracle information loss from fitted-model reversals; and derive a uniform-error condition under which the empirical sufficient set and frontier are stable. Across UCI HAR, HAR70+, and Daphnet, a one-point macro-F1 tolerance changes the minimum tested rate from best-score references at 10, 10, and 64 Hz to sufficient rates at 5, 10, and 32 Hz. Class-specific, cross-cohort, perturbation, data-volume, and cost analyses then identify when those decisions cease to be adequate. The contribution is therefore not a universal sampling threshold, but a formal account of when a finite resolution comparison supports a lower-cost design decision and what evidence is required for that decision to remain valid.

Primary area:
reinforcement learning
Keywords:
Wearable Time-Series Learning, Representation Learning, Representation Sufficiency, Temporal Resolution, Cross-Cohort Robustness, Domain Shift, Edge Computing, Digital Phenotyping, Human Activity Recognition, Health Monitoring

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_w5zMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper formulates wearable temporal-resolution selection as a benchmark-relative sufficiency problem. Given a tolerance ϵ\epsilon, it first identifies configurations whose predictive utility is sufficiently close to the best tested reference, and then selects among these admissible configurations according to deployment costs such as sampling rate, latency, and model size. The paper further studies how this decision changes with the reference set, target environments, information loss, and estimation uncertainty.

Critical strengths
  • The problem formulation is clear and usefully distinguishes the configuration with the highest predictive score from configurations that may already be sufficiently accurate for lower-cost deployment.
  • The theory, algorithm, and experiments are organized around a consistent decision-making framework rather than a collection of loosely connected ablations.
  • The experimental analysis goes beyond a simple sampling-rate sweep and includes class-specific behavior, distribution shift, corruption, and computational cost.
  • The paper is careful not to overclaim a universal sampling threshold and consistently frames the conclusions as benchmark-relative.
Critical weaknesses
  • The main novelty is limited. The trade-off between sensing cost and downstream predictive utility has been studied extensively in wearable and ubiquitous sensing. HiMAE (ICLR 2026) explicitly investigates resolution-dependent representations in wearable time series, while SleepSMC (ICLR 2025) studies maintaining reliable sleep staging under constrained sensing through multimodal coordination; related work such as CCAM (IEEE Internet of Things Journal 2025) also considers reducing sensing requirements from a cross-channel perspective. Although these works optimize different aspects of the sensing pipeline, they suggest that preserving task utility under reduced sensing resources is already a well-established research direction. The main contribution of this paper is therefore the specific formalization based on ϵ\epsilon-sufficiency followed by cost-aware Pareto filtering, rather than the sensing-efficiency problem itself.
  • Many theoretical results follow fairly directly from the definitions, including monotonicity of the admissible set and simple error-propagation bounds. The theoretical contribution is therefore more organizational than technically deep.
  • The empirical study relies mainly on handcrafted features, random convolutions, and Ridge classifiers. Evaluation with stronger modern wearable representation models or foundation models would provide a more convincing validation of the proposed framework.
  • The main analysis uses a fixed ϵ=0.01\epsilon=0.01, but the experiments do not provide a sufficiently strong statistical certificate to justify treating very small score gaps as practically equivalent.
  • The deployment-cost analysis does not directly measure several important real-world quantities, such as sensor acquisition energy, communication cost, or device-level power consumption.
Questions for authors

Please see Weaknesses

1 comment

As the author has not provided a rebuttal, I am maintaining my previous score.

Reviewer_G523OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces a sufficiency criterion for deciding when a lower sampling rate is enough for wearable sensing. A (rate, representation) configuration counts as sufficient if its macro-F1 is within a fixed tolerance of the best configuration in a finite tested grid. Among the sufficient configurations, a Pareto filter then compares deployment costs. The authors state six propositions, a theorem, and a corollary about this construction. In the experiments, four fixed feature families go through a Ridge probe on UCI HAR, HAR70+, and Daphnet. With a tolerance of 0.01 macro-F1, the lowest sufficient rate is 5 Hz on UCI HAR, 10 Hz on HAR70+, and 32 Hz on Daphnet, while the best-scoring rates are 10, 10, and 64 Hz.

Critical strengths

The paper aim to address a practical efficiency problem in wearable sensing, namely how to choose the lowest sampling rate and representation that preserve downstream task performance on resource-limited devices.

Critical weaknesses

A sweep over trained models is not needed to choose the sampling rate. A spectral analysis of the raw signal, combined with the Nyquist criterion, gives the required sampling rate before any model is trained. The paper’s empirical findings therefore restate what this simple analysis predicts and add little new value to the field.

Questions for authors

Sea Weakness

Reviewer_ZKS6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper shows an approach on choosing a sampling rate for wearable time series systems. The motivation is that you don't need to choose these sampling rates a priori but rather be determined downstream across other optimization functions. Instead of using conventional resolution sweeps that just pick the single highest-scoring configuration, the authors introduce a two-stage decision framework: first filtering candidate rate-representation pairs into an admissible "sufficient set" based on a user-declared predictive utility tolerance relative to the best tested reference, and then evaluating deployment costs (such as latency, model size, and sampling rate) only among those admissible candidates. They apply framework across various benchmarks to show best scores across various sampling rates. The authors also try to establish theoretical guarantees.

Critical strengths

None

Overall, this is an interesting problem domain for the research community but the paper doesn't have many strength outside of that.

Critical weaknesses

This paper was very difficult to read and follow. The authors would have to do a major rewrite to make it much easier to understand. This difficulty in the paper presentation makes it really hard to provide constructive feedback to the authors.

The empirical contribution is very thin relative to the theoretical depth the authors go into. Three small wearable datasets are evaluated with a single simple pipeline and the results are not very surprising at all. More importantly, the motivating deployment of cost is never measured so the paper never demonstrates which decision rule would change a real deployment decision. It isn't clear how well this framework would even work in practice.

Questions for authors

Since sensor energy/radio is the motivating cost but was not measured. Is there any rate pair where including an energy coordinate would change the frontier or can the measured coordinate analysis provably bound the unmeasured one?

For the UCI 50Hz→10Hz fitted reversal can any analysis attribute it to fitting error vs. approximation error?

In the AI use statement the authors state that mathematical additions "require author verification before submission" Does this imply that not of the mathematical foundations were verified prior to submission and it just AI generated? This is very worrisome.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.