{"id":"ac3c724a-3fb1-4a20-9eda-cc406a15fd03","arxiv_id":"2505.20739","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Adding squeeze-and-excitation style channel weighting and SiLU activation to a 1D ActionFormer improves activity localization mAP on WEAR inertial data.","lead":"An ActionFormer model for wearable sensors is modified with a channel-attention module that weights the most informative feature channels, reporting a 16.01% average mAP gain over the baseline on the WEAR dataset. The result matters because better sensor-based activity localization could improve health and sports monitoring on watches and phones.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 16.01% gain rests on a single unseeded run with test-set hyperparameter selection; the evaluation protocol must be reproduced before the claim is credible.","rationale":"The paper proposes a plausible SE-style enhancement to 1D-ActionFormer, and the mechanism itself is not internally inconsistent. The only strong evidence for the central claim is the CL=0.5 mAP improvement in Tables 1 and 2. The weakest point is the evaluation protocol: the paper does not report the data split, number of seeds, or error bars, and Section 3.2 explicitly says a key hyperparameter was selected based on test results. This matters because the 16.01% headline number is a single point estimate; if the test set was used for model selection, the number is optimistically biased, and without seeds the entire claim could be noise. I also noticed a secondary inconsistency: the TriDet baseline at CL=0.5 is 55.56 in Table 1 but 63.98 in Table 3, which suggests the experimental conditions are not described precisely enough to reproduce even the baseline tables. This supports the reader's concern rather than replacing it. The appropriate verdict remains CONDITIONAL: the claim is plausible but should not be relied upon until the authors provide code, the exact split, multi-seed results, and a validation-based hyperparameter selection protocol.","tokens_in":15491,"tokens_out":5755,"duration_ms":59995,"concrete_test":"Obtain or reconstruct the exact subject-wise train/test split used by Bock et al. for WEAR inertial data; run 1D-ActionFormer and CE-ActionFormer from the same initialization protocol for at least 5 random seeds, selecting beta (and the reduction factor) on a validation split held out from training, not on the test set. Report mean +/- std of average mAP at CL=0.5, and compute the 95% confidence interval for the difference. If the interval includes zero, or if the mean gap is substantially below 16.01 points, the headline claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the CL=0.5 inertial mAP gap (73.27 vs 57.26, +16.01 points). For this gap to be evidence for the model, the evaluation must be fair and stable. That condition is not met in the paper. Section 3.2 states that the swish beta value was 'finalize[d] to use ... 1.0 based on the test results'; the reduction factor of 16 and the module placement are also hand-chosen. No train/test split, no random seed, and no repeated runs are reported, so the 16.01-point gap could reflect selection bias (tuning on the test set) or a single favorable initialization rather than the module's effect. The comparison is also not controlled: the baseline is not given the same test-set hyperparameter optimization, and Table 1 and Table 3 list two different TriDet baselines at CL=0.5 (55.56 vs 63.98), signaling that the evaluation conditions are not stable enough to verify even adjacent numbers. Without code or a precise protocol, the headline improvement is not independently checkable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a channel-wise enhancement module for ActionFormer-based human activity recognition from inertial sensor signals. The module (CE-ActionFormer) performs adaptive average pooling over the feature sequence, passes the channel descriptors through two convolutions with a swish activation, and rescales the original features; a max-pooling variant is proposed for video features. On the WEAR dataset, the authors report an average mAP of 73.27 for inertial data at CL=0.5, which is 16.01 points above their 1D-ActionFormer baseline, and also report gains for CE-TriDet. The paper includes ablations (SE, swish, SE+swish), an efficiency analysis, confusion-matrix-based error analysis, and an appendix with video experiments.","tokens_in":15659,"tokens_out":4353,"duration_ms":44034,"significance":"If the reported results are reproducible, the proposed module is a simple and parameter-efficient extension (a 7.8% parameter increase) that could meaningfully improve temporal action localization on inertial data, and the video variant shows consistent if smaller gains across several datasets. The paper provides a clear ablation structure and a computational efficiency comparison, which are useful. However, the evaluation protocol is not robust: no train/validation/test split, no random seeds, no repeated runs, explicit test-set hyperparameter selection, and inconsistent baseline numbers across tables. The central quantitative claim is therefore not yet established to the stated precision.","major_comments":[{"comment":"The central claim of a 16.01-point improvement (Section 4.4, Table 1) is not yet established because the evaluation protocol does not support it. Section 3.2 states that the swish beta value was 'finalize[d] to use ... 1.0 based on the test results,' and the reduction factor of 16 and module placement are also hand-chosen. Section 4.3 does not report the train/validation/test split, random seed, or number of repeated runs. The reported single-run numbers could therefore reflect selection bias on the test set or a single favorable initialization rather than the effect of the module. Please report the exact split, select hyperparameters on a validation set only, and report means and standard deviations over multiple seeds.","section":"Section 3.2 and Section 4.3"},{"comment":"The same TriDet baseline at CL=0.5 is reported with an average mAP of 55.56 in Table 1 but 63.98 in Table 3 (and again 63.98 in Table 4). Under an identical protocol these values should coincide; a difference of 8.42 average-mAP points indicates that the evaluation conditions are not stable enough to verify even adjacent numbers. Please reconcile the two baseline values and specify the exact feature extraction, data split, and evaluation protocol used in each table.","section":"Tables 1, 3, and 4"},{"comment":"The error analysis does not support the claim that the module 'decrease[s] these defects' in capturing temporal dynamics and spatiotemporal interdependencies. The confusion matrices for the original and proposed models are presented, but the text only gives error proportions for the proposed model (e.g., label 1 predicted as label 3 with proportion 0.085). No quantitative comparison is made between the baseline and the proposed model's confusion matrices. Please provide a per-class comparison of confusion-matrix entries or an aggregate error-rate table showing the change in misclassification rates for the similar-activity pairs discussed.","section":"Section 4.8"}],"minor_comments":[{"comment":"The phrase 'Sequence-and-Excitation strategy' should be 'Squeeze-and-Excitation strategy'; this typo appears in the abstract, Section 1, and Section 5. In addition, the abstract contains incomplete sentences beginning with 'Based on the report of high temporal dynamics ...' that should be rewritten for clarity.","section":"Abstract and Section 1"},{"comment":"The '↑16.01%' and similar values in Table 2 are absolute percentage-point differences in average mAP (e.g., 73.27 − 57.26 = 16.01 points), not relative percentage improvements. Please label them as percentage points or provide relative improvement percentages to avoid overstating the magnitude.","section":"Table 2"},{"comment":"Section 4.5, 'Verify Our Method on TriDet', contains no text or results; it is followed immediately by Section 4.6. Either remove the empty section or move the CE-TriDet results here with the experimental conditions described.","section":"Section 4.5"},{"comment":"The appendix uses placeholder citations '[?]' for THUMOS14, I3D, TSP, ActivityNet-1.3, and EPIC-Kitchens 100, and the term 'CE-HAR' appears without definition. Please fill in the missing references and define all model names.","section":"Appendix A"},{"comment":"Several rows in Table 1 have missing spaces between CL and the mAP values (e.g., 'AFSESwish (Ours) 0.578.15 76.5971.20 ...'), making the table difficult to read. Please reformat the table entries.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The evaluation-protocol issues are the main gate for this manuscript. The empty section 4.5, missing citations in the appendix, and extensive typos suggest the manuscript needs a careful revision before it can be considered for publication. I would ask for code and data splits to be provided with the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this is a routine SE-style channel-attention variant bolted onto 1D-ActionFormer, and the paper knows it. The central claim—73.27 vs 57.26 average mAP at clip length 0.5 on WEAR inertial data—looks good on the page, but the evaluation is not tight enough to know whether the gain is real or an artifact of test-set tuning.\n\nWhat is actually new: attaching an SE-style channel gate with a 1x1 conv and SiLU after each transformer block in the 1D ActionFormer, and showing it improves the inertial baseline. That is a reasonable extension of known pieces; the paper cites the SE-in-HAR lineage openly, so the novelty claim is modest and fairly stated. The confusion-matrix and signal-visualization error analysis is a genuine extra; it does not prove the method works, but it is the kind of qualitative check a subfield paper should have.\n\nThe soft spots are real but not disqualifying. Section 3.2 says the swish beta was finalized to 1.0 based on the test results, and the reduction factor and module placement are hand-picked. No seeds, no repeated runs, no code, no explicit train/test split. With a 16-point gap that alone would be suspicious; with test-set selection it is on the authors to show stability. Also, TriDet's baseline is 55.56 in Table 1 but 63.98 in Table 3 for the same CL=0.5 setting; that inconsistency suggests the evaluation conditions shift between tables, which makes the comparison less clean. There are scattered typos and unresolved \"?\" citations in the appendix that should have been cleaned.\n\nThe central mechanism is not circular—it is an empirical result, not a derivation—and the parameter overhead is honestly reported as a 7.8% increase. So the paper is not sloppy in its claims, only in its reporting discipline.\n\nFor whom: readers working on wearable HAR localization who want to see whether SE-style attention helps a transformer action localizer. It deserves a serious referee, but the referee should ask for a minimal reproducibility package: seeds, error bars, and a protocol that fixes beta and reduction factor without test-set feedback. That said, I would not cite it until those numbers are reproducible.\n\nMy recommendation: send it to peer review, but with the clear expectation that the evaluation section needs to be redone or heavily revised.","headline":"A modest but honest SE-style extension of ActionFormer for inertial HAR; the reported 16-point gain is plausible but not yet established because the evaluation protocol is under-specified.","tokens_in":16172,"tokens_out":1605,"would_cite":false,"duration_ms":17009,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that reweighting feature channels with a squeeze-and-excitation-style module raises ActionFormer's average mAP on wearable inertial data from 57.26 to 73.27 at clip length 0.5.","keywords":["human activity recognition","temporal action localization","ActionFormer","squeeze-and-excitation","channel attention","inertial sensor data","wearable computing","swish activation"],"falsifier":"Re-run the inertial WEAR experiments with at least five random seeds and a fixed validation split, selecting hyperparameters (swish beta, reduction factor, module placement) only on validation data; if CE-ActionFormer's average mAP advantage over 1D-ActionFormer at clip length 0.5 does not reproduce consistently, or vanishes when hyperparameters are not chosen on the test set, the central claim is refuted.","tokens_in":15277,"feed_emoji":"📈","tokens_out":11825,"duration_ms":93019,"temperature":0.7,"pith_summary":"Human activity recognition from wearable inertial sensors needs to locate activity boundaries as well as labels, and the paper argues that transformer-based detectors like ActionFormer miss subtle signal changes because they treat every feature channel as equally informative. The authors propose the adaptive channel-wise enhancement module, a squeeze-and-excitation-style gate with swish activation, inserted after each transformer block to amplify informative channels while preserving negative-range directional information. On the WEAR dataset, this CE-ActionFormer raises mean average precision (mAP) at clip length 0.5 from 57.26 to 73.27, a 16.01 percentage-point gain over the 1D-ActionFormer baseline, with the largest gains at the strictest overlap threshold. The same module also improves a convolutional baseline, suggesting the mechanism is not specific to transformers.","feed_headline":"Reweighting channels lifts wearable activity detection by 16%","feed_subtitle":"A light channel-weighting module helps transformers localize wearable activities, most at strict overlap thresholds.","key_machinery":"The adaptive channel-wise enhancement module is the mechanism that carries the argument. Given features $X$ of shape (batch, channels, time), it computes channel weights $W_C = \\mathrm{sigmoid}(\\mathrm{conv}_2(\\mathrm{swish}(\\mathrm{conv}_1(\\mathrm{avgPool}(X)))))$, reweights the features, and passes them through a $1\\times1$ convolution; the average pooling squeezes each channel to one scalar, the reduction factor 16 keeps the gate cheap, and swish is chosen over ReLU so that negative values—directional information in acceleration—are retained. A MaxPool variant for video replaces average pooling with local max pooling and interpolates the weights back to the original sequence length. This module is the only architectural change relative to the baseline, so the reported gains are attributed to it.","core_discovery":"The central claim is that on inertial data, the most useful cues for activity recognition and boundary localization are concentrated in a few feature channels, so reweighting channels explicitly improves temporal action localization. The paper shows that replacing the naive equal treatment of channels with a lightweight gate—adaptive average pooling to a per-channel scalar, a two-layer bottleneck with reduction factor 16, swish activation, and a sigmoid gate—yields a 16.01 percentage-point average mAP gain over the 1D-ActionFormer baseline on the WEAR dataset at clip length 0.5, including a 37.23 percentage-point gain at tIoU 0.7. The authors further claim that this module reduces the model's defects in capturing temporal dynamics and spatiotemporal interdependencies, and they verify the approach on a convolutional detector (CE-TriDet) and on video inputs.","pith_inferences":["Because the improvement concentrates at high tIoU, a natural testable extension is whether the channel weights sharpen boundary predictions, e.g., by measuring predicted-to-ground-truth boundary distance with and without the module.","The 'channel' notion maps directly to sensor axes (four IMUs times three axes); combining the module with axis-wise normalization or permutation augmentation—both designed to preserve channel-specific information—could compound the gains.","The module also helps on video features, which hints that suppressing redundant channels in the feature pyramid is the underlying mechanism, so it may complement other temporal action localization improvements such as relative boundary modeling rather than replace them."],"forward_implications":["At CL=0.5 and tIoU=0.7, CE-ActionFormer reaches 68.14 mAP versus the baseline's 30.91, so the gain is concentrated where boundary localization is hardest.","The module adds only 7.8% parameters (26.56M to 28.64M) and about 5% execution time, making the improvement inexpensive.","Applying the same gating to TriDet lifts its average mAP from 63.98 to 70.61 at CL=0.5, showing the mechanism transfers to convolutional detectors.","On WEAR video features, an adaptive-average-pool variant improves ActionFormer from 54.44 to 62.24 average mAP, suggesting the channel-reweighting idea generalizes beyond inertial signals.","Ablations show swish alone slightly hurts (57.11) and SE alone helps (68.07), so the full module's gain comes from combining SE-style gating with swish's retention of negative-range information."],"supporting_citations":[{"why":"Defines ActionFormer, the transformer architecture for temporal action localization that the paper modifies.","marker":"[44]"},{"why":"Provides the WEAR dataset, the inertial features, and the 1D-ActionFormer baseline whose mAP the paper compares against.","marker":"[6]"},{"why":"Supplies the squeeze-and-excitation strategy that the adaptive channel-wise enhancement module adapts.","marker":"[21]"},{"why":"Introduces the swish activation function that the module uses to retain negative-range signal information.","marker":"[34]"},{"why":"Defines TriDet, the convolutional temporal action detection model used for the transfer experiment.","marker":"[35]"}],"fun_headline_variants":["Channel reweighting lifts sensor activity detection by 16%","Gating key channels improves wearable activity localization by 16%","Selective channel weighting yields 16% better HAR on wearables","Channel attention sharpens action boundaries, lifting mAP 16%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported 16.01 percentage-point gain assumes a fair and stable evaluation protocol, since the paper selects the swish beta, reduction factor, and module placement based on test results and reports a single run without multiple seeds, error bars, or a stated train/test split.","fun_headline_variants_meta":{"raw":{"variants":["Channel reweighting lifts sensor activity detection by 16%","Gating key channels improves wearable activity localization by 16%","Selective channel weighting yields 16% better HAR on wearables","Channel attention sharpens action boundaries, lifting mAP 16%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000629,"raw_usage":{"total_tokens":2880,"prompt_tokens":888,"completion_tokens":1992,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":1919}},"tokens_in":504,"tokens_out":1992,"duration_ms":14852,"temperature":1.0,"reasoning_tokens":1919,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:47:17.086806+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the inertial WEAR experiments with at least five random seeds and a fixed validation split, selecting hyperparameters (swish beta, reduction factor, module placement) only on validation data; if CE-ActionFormer's average mAP advantage over 1D-ActionFormer at clip length 0.5 does not reproduce consistently, or vanishes when hyperparameters are not chosen on the test set, the central claim is refuted.","supporting_citations":[{"cited_title":"Actionformer: Local- izing moments of actions with transformers","cited_arxiv_id":null,"evidence_quote":"Defines ActionFormer, the transformer architecture for temporal action localization that the paper modifies."},{"cited_title":"WEAR: An Outdoor Sports Dataset for Wearable and Egocentric Activity Recognition","cited_arxiv_id":"2304.05088","evidence_quote":"Provides the WEAR dataset, the inertial features, and the 1D-ActionFormer baseline whose mAP the paper compares against."},{"cited_title":"Tridet: Temporal action detection with relative boundary modeling","cited_arxiv_id":null,"evidence_quote":"Defines TriDet, the convolutional temporal action detection model used for the transfer experiment."}],"review_version":1}