{"id":"697cd118-4889-4e60-918c-b4693a7a621b","arxiv_id":"2608.05572","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Viveka cuts wearable sensing energy by gating per-activity sensor and rate policies behind a confidence and stability check.","lead":"Viveka is a context-aware sensing framework for smart wearables that picks a small set of sensors and sampling rates per activity, and applies that aggressive policy only when an on-device controller is confident and the context is stable. In simulations on the MHEALTH and PAMAP2 datasets, it reports up to 75% energy savings and 78% data reduction over always-on baselines, while keeping classification accuracy within a few points.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Empirical headline numbers rest on test-set-tuned thresholds and lack variance reporting; the central efficiency claim needs a proper nested leave-one-subject-out evaluation.","rationale":"The reader identified the NP-hardness proof in Appendix VIII as the weakest assumption. That proof is indeed flawed: it constructs the error function C(x) to be satisfied iff a set-cover condition holds, without demonstrating that such a function can be realized as the expected classification error of any actual inference model or data distribution. This makes the reduction vacuous as stated. However, the NP-hardness theorem is a secondary contribution; even if the proof were corrected or removed, Viveka's heuristic framework and its empirical evaluation would stand or fall on their own. The single most load-bearing concern for the paper's central claim is the validity of the reported efficiency numbers. The paper's own footnote reveals that the headline figures were produced under a best-case threshold-selection protocol, while a rigorous leave-one-subject-out evaluation yields much lower savings. This directly undermines the abstract's quantitative claims, which are what a practitioner would take away. The lack of error bars compounds the problem: the accuracy maintenance claim ('within 3-5% of baselines') is asserted without any statistical support on datasets with fewer than a dozen subjects. The reader's rationale did mention absent error bars and best-case numbers, but the formal 'weakest_assumption' focused on the hardness proof, which I do not regard as the load-bearing issue. Verdict should remain CONDITIONAL because the gating framework is plausible and potentially useful, but the empirical evidence as reported is not sufficient to accept the headline performance figures; a nested cross-validation and variance reporting are needed before the central claim can be verified.","tokens_in":18454,"tokens_out":5448,"duration_ms":49837,"concrete_test":"Run a nested leave-one-subject-out evaluation: for each held-out subject, tune all five parameters (theta_PFI, theta_FFT, theta_conf, beta_conf, T_interval) only on the remaining subjects using inner cross-validation, then evaluate on the held-out subject. Report the mean and standard deviation of energy savings, data reduction, and F1 across subjects, and perform a paired statistical test (e.g., Wilcoxon signed-rank) comparing Viveka's F1 to the AlwaysOn baseline and to GS-VS. If the mean savings are close to the footnote's 64%/49% rather than 75%/78%, the abstract's headline is a misleading best-case; if the F1 differences are not statistically significant, the 'within 3-5%' accuracy claim is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (75% energy savings, 78% data reduction, accuracy within 3-5%) is presented in the abstract and conclusion as the outcome of Viveka. However, the footnote in Section I admits these figures are a 'best-case configuration in which policy thresholds were selected on a single subject-disjoint split.' Section V-E describes sequential tuning of the four thresholds and decision interval without stating that tuning was restricted to a held-out validation set, and the footnote further discloses that leave-one-subject-out validation with thresholds selected only on training subjects yields materially lower savings (approximately 64% on MHEALTH and 49% on PAMAP2). Thus, the headline numbers likely reflect selection bias from tuning on test subjects, and the reported improvements may not generalize to a realistic deployment where thresholds must be fixed before seeing a new user. Additionally, the paper reports no error bars, confidence intervals, or per-subject variance for any metric; with only 10 (MHEALTH) and 9 (PAMAP2) subjects, the claimed F1 differences between Viveka and baselines may be within noise. Without a statistically sound evaluation protocol, the central contribution's magnitude is unsupported.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formulates joint sensor selection and sampling-rate selection as an optimization problem (APSSE) under a classification-error constraint, claims it is NP-hard, and proposes Viveka, a lightweight context-aware framework. Viveka uses a low-power controller to estimate activity and confidence, a stability/confidence-gated policy with three tiers (optimized per-context policy, recovery union policy, conservative fallback), and offline per-context sensor/rate policies derived from permutation feature importance and FFT-based spectral analysis. The framework is evaluated on the MHEALTH and PAMAP2 datasets, with reported best-case energy savings up to 75% and data reduction up to 78% while keeping F1 within 3-5% of baselines.","tokens_in":18656,"tokens_out":4228,"duration_ms":40170,"significance":"If the reported results hold under a properly validated protocol, the paper presents a practically relevant and well-motivated approach to energy-efficient wearable sensing. The tiered gating idea is a genuine contribution over static selection and duty-cycling baselines: it directly addresses the risk of applying an aggressive sensing policy to a misidentified context. The paper also provides a concrete, datasheet-grounded energy model, a sensitivity analysis over the four thresholds and decision interval, and a designed co-training step with rate-degraded augmentation. These are strengths that go beyond a purely algorithmic contribution. However, the central quantitative claim currently rests on a best-case configuration whose thresholds were selected on a single subject-disjoint split, as disclosed in the Section I footnote, and the paper reports no variance or per-subject statistics. The contribution is therefore promising but not yet supported at the level claimed in the abstract and conclusion.","major_comments":[{"comment":"The headline numbers (75% energy savings and 78% data reduction in the abstract and conclusion) are attributed to a 'best-case configuration' in which policy thresholds were selected on a single subject-disjoint split (footnote 2 in Section I). Section V-E describes parameter sweeps but does not state that threshold selection was restricted to a held-out validation set; instead, Section V-D reports the resulting dataset-specific hyperparameters as though they are fixed. The footnote's own leave-one-subject-out numbers (≈64% on MHEALTH and ≈49% on PAMAP2) are materially lower than the headline, so the central empirical claim is not yet protected against test-set selection bias. This is load-bearing because the paper's main contribution is the claimed energy/accuracy trade-off. Please re-run the evaluation under a nested leave-one-subject-out protocol in which all four thresholds and T_interval are chosen only on training subjects, and report the resulting Figure 7, Table III, and F1 comparisons.","section":"§I (footnote 2), §V-D, §V-E, Fig. 7, Table III"},{"comment":"The NP-hardness reduction constructs the classifier error function C(x) so that C(x) ≤ C_threshold holds exactly when the selected active sensors cover all elements of U. This is a by-construction equivalence, not a derivation from a concrete classifier, data distribution, and accuracy model. Since C(x) in the original formulation (Eq. (5)) is the expected classification error of the inference model, the reduction should show that a coverage-indicator error function is realizable in that setting; otherwise the proof only establishes hardness for a specially constructed variant of APSSE. This affects the stated contribution 'we prove it is NP-hard.' Please either instantiate the reduction with a concrete classifier and data distribution, or explicitly restate the theorem for a class of problems with arbitrary error functions C(x).","section":"Appendix VIII, Theorem 1"},{"comment":"No variance, confidence interval, or per-subject statistic is reported for any energy, data-reduction, or F1 metric. With only 10 (MHEALTH) and 9 (PAMAP2) subjects, the reported differences between Viveka and baselines, e.g., F1 0.758 vs 0.683 and energy savings 74% vs 72% against GS-VS on MHEALTH, may be within subject-level noise. The projected battery-life numbers in Table III are deterministic point estimates. A per-subject breakdown or confidence intervals are needed to support the comparative claims and to let readers assess the stability of the savings across users.","section":"§V-D, §V-E, Table III"}],"minor_comments":[{"comment":"The abstract and conclusion repeat the best-case savings figures without the caveat disclosed in the Section I footnote; please either add the qualification in these locations or report the leave-one-subject-out numbers as the primary results.","section":"Abstract, Section VII"},{"comment":"The claim that 'Transient periods are captured with distinct sensor configurations' is qualitative; please provide a quantitative measure of transition detection latency or the fraction of transitions correctly detected.","section":"Section V-G, Figure 10"},{"comment":"The WISDM PFI experiment reports accuracy drops without error bars or repeated-run variance; adding this would increase confidence in the claim that sensor importance is activity-dependent.","section":"Section II-B, Figures 3 and 4"},{"comment":"The CoSS reference appears incomplete: no venue or publication year is given.","section":"Section VI, Reference [18]"},{"comment":"There are several typographical and grammatical errors, e.g., 'The nameVivekais' (missing space), 'uncertainity', and 'upto'; a careful proofread is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a sensible systems idea and a weak empirical case. The best part is the three-tier gating policy—apply the aggressive per-context configuration only when context is stable and confident, widen to a union of plausible sensors otherwise, and fall back to safe mode when uncertain. That is a genuinely reasonable way to handle context misclassification, and the co-design step (training the edge model on rate-degraded, reconstructed signals) is a nice touch that closes a real deployment gap. The authors also deserve credit for disclosing, in a footnote, that the headline 75%/78% figures come from a best-case configuration.\n\nThe problems are in the evaluation. First, the thresholds were tuned on a single subject-disjoint split, effectively on the test set, and the parameters are swept in isolation with the best value locked in before the next sweep. That is selection bias, and the authors' own footnote shows what it costs: leave-one-subject-out with thresholds chosen only on training subjects drops savings to about 64% on MHEALTH and 49% on PAMAP2. The 78% data reduction on PAMAP2 falls to 49%—that is not a small change. Second, there are no error bars or per-subject variance anywhere. With 10 and 9 subjects, the F1 differences between Viveka and baselines could easily be within noise. Third, the NP-hardness proof in Appendix VIII defines the error function C(x) so that the accuracy constraint is exactly a set-cover coverage condition, with no argument that any real classifier realizes such an error function. That is a by-construction equivalence, so the hardness result is unproven for actual HAR models. It is an appendix and not the main contribution, but it should be either fixed or removed.\n\nThe conclusion repeats the best-case numbers without the qualifier, which is misleading, though the footnote mitigates it. No code or data is released, which is disappointing for a simulation-only evaluation.\n\nBottom line: the gating mechanism is a reasonable contribution and the paper is clearly written, but the central efficiency claims are not yet supported. It deserves a serious referee, not a desk reject, because the design is worth discussing and the evaluation can be fixed. I would tell the authors to run a proper nested leave-one-subject-out protocol, report variance and per-subject results, release the simulation and the tuned threshold files, and revisit or remove the NP-hardness appendix. If the LOSO numbers hold at the 64%/49% level, that is still a good result, just a more honest one.","headline":"Sensible gating design, but the headline savings are best-case; needs nested cross-validation and a cleaner NP-hardness proof before the numbers can be trusted.","tokens_in":19213,"tokens_out":2564,"would_cite":false,"duration_ms":21136,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Viveka claims a context-gated sensing policy can cut wearable energy use by up to 75% while keeping activity-recognition accuracy within a few percent of always-on baselines.","keywords":["wearable sensing","energy efficiency","context-aware sensing","sensor selection","adaptive sampling","human activity recognition","permutation feature importance","spectral energy analysis"],"falsifier":"Deploy the full policy on a real wearable instrumented to measure per-sensor current draw and total battery life while a subject follows the MHEALTH and PAMAP2 protocols, and check whether measured energy savings and F1 fall within the reported envelope (up to 75% energy, 78% data reduction, accuracy within 3-5%); separately, enumerate all configurations of a small APSSE instance with a real trained classifier to test whether the minimal-energy configuration matches the coverage equivalence used in the NP-hardness reduction.","tokens_in":18183,"feed_emoji":"🔋","tokens_out":5393,"duration_ms":43279,"temperature":0.7,"pith_summary":"The paper tries to show that joint sensor selection and sampling-rate adaptation can be made both safe and practical for wearables by separating context estimation from trust estimation. A light, always-on controller decides not just what activity is happening but how much to trust that guess, and a three-tier gating policy applies aggressive per-context sensor and rate settings only when confidence is high. This turns a risky optimization into a closed-loop decision problem, and the reported evaluations on MHEALTH and PAMAP2 indicate large energy and data savings with only minor accuracy loss. If correct, the framework addresses a practical blocker for smart wearables: battery life and the need to avoid acting on a misidentified context.","feed_headline":"Context-gated sensing cuts wearable energy up to 75%","feed_subtitle":"Viveka applies aggressive low-power sensor settings only when it trusts the activity guess, preserving accuracy within a few percent.","key_machinery":"The central mechanism is the stability-and-confidence gate: a three-tier decision rule that applies an aggressive per-context sensor and sampling-rate configuration only when the lightweight controller's recent activity predictions are both repeated and above a confidence threshold, and otherwise widens the sensor set or falls back to a conservative default. This gate is supported by an always-on controller that estimates both context and its own trustworthiness, offline policy maps built by permutation feature importance for sensor sets and FFT-based spectral energy analysis for per-sensor rates, a rate-degradation augmented training scheme that matches online reconstruction, and a datasheet-grounded energy model spanning sensing, computation, and BLE communication.","core_discovery":"Viveka's central claim is that applying an aggressive, context-specific sensing policy is safe when the on-device context estimate is both stable and confident. The system runs a lightweight controller that continuously outputs an activity prediction and a confidence value; a stability-and-confidence gate then chooses one of three tiers: the fully optimized per-activity configuration when recent predictions are consistent and confident, a recovery configuration that unions the sensor requirements of plausible activities under moderate uncertainty, and a conservative high-fidelity fallback when context is unresolved. Per-activity configurations are computed offline using permutation feature importance to choose minimal sensor sets and spectral (Nyquist-based) energy analysis to choose per-sensor sampling rates, and the recognition model is trained with rate-degraded, reconstructed-signal augmentation so it stays accurate on the subsampled input it receives at runtime. On MHEALTH and PAMAP2, the paper reports up to 75% energy savings and 78% data reduction over standard baselines while keeping F1 within 3-5% of the always-on baseline, with projected battery life extending from 8.7 to 34.6 days on a 300 mAh battery for MHEALTH.","pith_inferences":["Editorial inference: the headline 75% energy and 78% data figures reflect a best-case threshold configuration selected on a single subject-disjoint split, while the paper's leave-one-subject-out numbers are more conservative (approximately 64% on MHEALTH and 49% on PAMAP2), so the practical savings depend on how robustly the thresholds are chosen without tuning on test subjects.","Editorial inference: if the confidence gate works as described, a natural extension is to make the threshold itself adaptive to battery level or user history, trading a little accuracy for longer operation when the battery is low — a testable design the paper does not pursue.","Editorial inference: the offline per-activity sensor importance assumes the activity-conditional feature ranking is stable across users and deployment conditions; personalized or online-updated importance maps could either improve savings or expose sensitivity to distribution shift.","Editorial inference: the NP-hardness reduction uses a coverage-based placeholder for the classifier error function; the heuristic's value does not depend on this proof, but a more realistic error model could change the problem's complexity and should be examined separately."],"forward_implications":["If the reported results hold, wearable devices can extend operational battery life substantially, with the paper's projections roughly quadrupling runtime on MHEALTH under a 300 mAh battery.","Data volume transmitted to the edge drops by roughly three quarters, reducing BLE energy use and on-device storage pressure while still feeding the recognition model reconstructed signals.","The gating idea generalizes to other energy-constrained multi-sensor IoT settings where acting on a misidentified context is risky, since the controller is decoupled from the recognition model.","Because joint sensor and rate selection is NP-hard, a runtime-tractable heuristic like Viveka is the practical route; future exact or approximate solvers would need a different basis.","Training the recognition model on rate-degraded, reconstructed signals closes the gap between offline policy computation and online execution, making sub-sampled sensing viable for deployment.","The three-tier fallback behavior avoids sensor thrashing, addressing the energy overhead of rapidly toggling sensors on and off during activity transitions."],"supporting_citations":[{"why":"Supplies the permutation feature importance method used to compute the minimal per-activity sensor sets in Algorithm 1.","marker":"[23]"},{"why":"Provides the MHEALTH dataset used as one of the two evaluation benchmarks for the reported energy and accuracy results.","marker":"[24]"},{"why":"Provides the PAMAP2 dataset used as the second evaluation benchmark with its larger sensor suite and higher sampling rate.","marker":"[26]"},{"why":"Supplies the TinierHAR model architecture that is split into a lightweight wearable model and a heavier edge model in the evaluation.","marker":"[27]"},{"why":"Defines the variance-based adaptive sampling baseline that Viveka is compared against as a representative adaptive-rate strategy.","marker":"[14]"},{"why":"Provides the activity-based dynamic sensor selection approach that the AS and AS-VS baselines are adapted from.","marker":"[17]"},{"why":"Sources the Nordic nRF52840 BLE datasheet specifications that ground the communication energy model in the evaluation.","marker":"[33]"},{"why":"Provides the BMI160 accelerometer and gyroscope current consumption datasheet values used in the sensing energy model.","marker":"[29]"}],"fun_headline_variants":["Viveka gates sensing on context to save 75% energy","Trust-aware sensing: 75% energy cut, 5% accuracy loss","Context certainty unlocks 75% energy savings in wearables","Stable context? Aggressive sensing saves 75% on wearables","75% less energy, 78% less data with Viveka's context gate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The NP-hardness proof assumes that the classifier error function can be defined so that meeting the accuracy constraint is equivalent to covering a set of activity types with the active sensors, but no concrete classifier, data distribution, or accuracy model is supplied to realize that error function.","fun_headline_variants_meta":{"raw":{"variants":["Viveka gates sensing on context to save 75% energy","Trust-aware sensing: 75% energy cut, 5% accuracy loss","Context certainty unlocks 75% energy savings in wearables","Stable context? Aggressive sensing saves 75% on wearables","75% less energy, 78% less data with Viveka's context gate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000793,"raw_usage":{"total_tokens":3534,"prompt_tokens":1025,"completion_tokens":2509,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":2414}},"tokens_in":641,"tokens_out":2509,"duration_ms":39423,"temperature":1.0,"reasoning_tokens":2414,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T10:15:41.116488+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy the full policy on a real wearable instrumented to measure per-sensor current draw and total battery life while a subject follows the MHEALTH and PAMAP2 protocols, and check whether measured energy savings and F1 fall within the reported envelope (up to 75% energy, 78% data reduction, accuracy within 3-5%); separately, enumerate all configurations of a small APSSE instance with a real trained classifier to test whether the minimal-energy configuration matches the coverage equivalence used in the NP-hardness reduction.","supporting_citations":[{"cited_title":"mhealthdroid: a novel framework for agile development of mobile health applications,","cited_arxiv_id":null,"evidence_quote":"Provides the MHEALTH dataset used as one of the two evaluation benchmarks for the reported energy and accuracy results."},{"cited_title":"TinierHAR: Towards Ultra-Lightweight Deep Learning Models for Efficient Human Activity Recognition on Edge Devices","cited_arxiv_id":"2507.07949","evidence_quote":"Supplies the TinierHAR model architecture that is split into a lightweight wearable model and a heavier edge model in the evaluation."},{"cited_title":"On adaptive sampling algorithms for iot devices,","cited_arxiv_id":null,"evidence_quote":"Defines the variance-based adaptive sampling baseline that Viveka is compared against as a representative adaptive-rate strategy."},{"cited_title":"Instance-wise dynamic sensor selection for human activity recognition,","cited_arxiv_id":null,"evidence_quote":"Provides the activity-based dynamic sensor selection approach that the AS and AS-VS baselines are adapted from."},{"cited_title":"nRF52840 Product Specification,","cited_arxiv_id":null,"evidence_quote":"Sources the Nordic nRF52840 BLE datasheet specifications that ground the communication energy model in the evaluation."},{"cited_title":"BMI160 Data sheet,","cited_arxiv_id":null,"evidence_quote":"Provides the BMI160 accelerometer and gyroscope current consumption datasheet values used in the sensing energy model."}],"review_version":1}