{"id":"2261012a-9b4c-43b7-91f4-88a5ee6e8a3b","arxiv_id":"2507.02826","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A dual-path ResNet/DenseNet framework with multi-stage contrastive learning and confidence-driven gradient modulation is presented for multimodal human activity recognition, with reported improvements on four public datasets.","lead":"Researchers propose DCDP-HAR, a two-branch neural network for recognizing human activities from wearable sensors, combining ResNet and DenseNet features with contrastive alignment and confidence-based gradient adjustments. The goal is to stop one sensor type from dominating training, improving accuracy on four public activity recognition benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hyperparameters tuned on the OPPORTUNITY dataset and then used for all benchmarks mean the reported accuracies are not independent generalization estimates.","rationale":"The paper's central claim is that DCDP-HAR achieves superior recognition performance on four benchmarks and that each component (dual-path, contrastive learning, gradient modulation) contributes. For that claim to hold, the reported accuracies must be honest estimates of generalization to unseen data. The paper explicitly says hyperparameter optimization was conducted on the OPPORTUNITY dataset and the chosen configuration was then used for all experiments. If the same test split was used for tuning and final evaluation, the OPPORTUNITY result is not a valid generalization estimate, and because the final configuration was used on all other datasets, the other results inherit configuration choices that were made using OPPORTUNITY test information (a form of multi-dataset selection). This is a textbook violation of the no-test-set-in-model-selection principle and is sufficient to reject the empirical claims as presented. The concern is concrete and checkable, unlike the more abstract question of whether the unspecified sensor partition is meaningful. I partially agree with the reader: their formal weakest assumption (unspecified partition) is plausible, but the tuning-on-test issue is more directly fatal to the validity of the numbers; the reader's rationale does list this issue among the internal inconsistencies, so we partially overlap. The undefined 'DA' abbreviation in Table 2 and the inconsistent PAMAP2/OPPORTUNITY numbers are secondary symptoms of the same reporting problem, not separate mechanisms. Therefore, the reader's REJECT verdict stands, and I would not change it.","tokens_in":12233,"tokens_out":3495,"duration_ms":38530,"concrete_test":"Determine whether the OPPORTUNITY split used for hyperparameter tuning in Section 5.1 is the same as the test split used for Table 3. Concretely: (1) check the paper or released code for any statement that tuning was done only on the training or validation portion; if absent, rerun the full DCDP-HAR pipeline with a nested protocol—tune all hyperparameters on a held-out validation split of the OPPORTUNITY training set, freeze the configuration, then evaluate once on the benchmark's standard test split—and compare the resulting OPPORTUNITY accuracy to the reported 88.34%. (2) Re-run the final configuration on PAMAP2 and check whether accuracy is 93.90 (Table 2) or 93.71 (Table 3); a mismatch of more than 0.1% under identical settings indicates unreported variability or inconsistent reporting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is that the paper's headline accuracy numbers are not independent of model selection. Section 5.1 states \"This optimization was conducted on the OPPO dataset\" (i.e., OPPORTUNITY), and Section 5.2 says the resulting configuration \"was subsequently employed for all ablation studies and comparative performance evaluations.\" If the same OPPORTUNITY test split was used both to select hyperparameters (optimizer, architecture, contrastive weight/temperature, CGM alpha) and to report the final 88.34% accuracy in Table 3, then that number is an optimistically biased selection result, not a fair test of the method. The same contamination propagates to the ablation study if the ablation is run with the already-tuned configuration. The concern is reinforced by internal numeric inconsistencies: the best tuned OPPORTUNITY accuracy is reported as 88.62% in Section 5.2 but Table 3 lists DCDP-HAR at 88.34%; for PAMAP2, the ablation table shows 93.90% while the comparison table shows 93.71% for the same method. These mismatches make it difficult to verify which configuration produced which number, further undermining the central empirical claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DCDP-HAR, a multimodal activity-recognition framework combining a dual-path ResNet/DenseNet feature extractor, multi-stage contrastive learning for cross-modal alignment, and a confidence-driven gradient modulation (CGM) mechanism that scales branch gradients during backpropagation. The authors report ablations and comparisons on OPPORTUNITY, PAMAP2, WISDM, and UCI-HAR, claiming consistent improvements over baselines. The core claim is that the three components jointly mitigate modality imbalance and improve cross-modal alignment, leading to higher accuracy.","tokens_in":12507,"tokens_out":6571,"duration_ms":77957,"significance":"If the empirical claims were clean, the paper would be a modest but potentially useful contribution to multimodal sensor-based HAR: the CGM idea is a reasonable extension of gradient-modulation approaches, and multi-stage contrastive alignment is a sensible design choice. The paper also addresses a real problem, modality imbalance, and evaluates on four widely used benchmarks. However, the evaluation as presented is not trustworthy: hyperparameters are tuned on the OPPORTUNITY test set, the ablation table contains an undescribed component (DA), the sensor partition for the dual paths is never specified, and several reported numbers are inconsistent across tables. No code or detailed training protocol is provided. These issues are load-bearing for the central empirical claim, so the manuscript needs major additional work before the results can be accepted.","major_comments":[{"comment":"The headline OPPORTUNITY accuracy is not an independent test result. Section 5.1 states that hyperparameter tuning was conducted on the OPPO (OPPORTUNITY) dataset, and Section 5.2 says the finalized configuration was then used for all ablations and comparative evaluations. Because the tuning process used the test accuracy to select optimizers, architectures, contrastive weight/temperature, and the CGM alpha, the 88.34% reported for DCDP-HAR on OPPORTUNITY in Table 3 is optimistically biased by model selection on the test split. The authors must use a separate validation split or nested cross-validation for tuning and report the final test accuracy only after the configuration is fixed.","section":"Sections 5.1 and 5.2, Table 3"},{"comment":"Reported accuracy for the same model is inconsistent. Table 2 lists the full DCDP-HAR at 93.90% on PAMAP2, while Table 3 lists 93.71% for the same dataset and model. Similarly, Section 5.2 reports the best tuned OPPORTUNITY accuracy as 88.62%, but Table 3 reports 88.34% for DCDP-HAR on OPPORTUNITY. The authors need to clarify which configuration produced each number and which test protocol was used; as written, the discrepancies make it impossible to verify the empirical results.","section":"Tables 2 and 3, Section 5.2"},{"comment":"The ablation study includes a component labeled 'DA' (Data Augmentation), including the row '+ DP + DA', but data augmentation is never described in the methodology. Since the full DCDP-HAR model in Table 2 includes DA, the reported 93.90% cannot be attributed solely to the three described components. The authors must specify the augmentation procedure, include it in the method, and ablate it separately with the same naming convention as the other components.","section":"Table 2, Section 3"},{"comment":"The dual-path partition is not specified for any dataset. The paper defines feature index sets I1 and I2 for splitting the input into two complementary subsets but never states, for any benchmark, which sensor channels or feature dimensions go to the ResNet path versus the DenseNet path. Without this information, the claimed benefit of the dual-path design cannot be distinguished from simply increasing model capacity. This is especially problematic for WISDM, which is a single-sensor (accelerometer-only) dataset; the paper's framing in terms of modality imbalance and cross-modal alignment needs to explain how the dual-path split applies there.","section":"Section 3.1"},{"comment":"The ablation discussion is internally inconsistent with the table. Adding contrastive learning to the dual-path baseline decreases accuracy from 91.71% to 91.24%, yet the text claims that removing any single module causes about a 3% accuracy reduction; in the full model, removing contrastive learning (DCDP-HAR vs +DPFE+CGM+DA) reduces accuracy by only 1.04%. The non-monotonic interaction and the mismatch between the text and the table need to be explained, and the contribution of contrastive learning should be justified in light of the negative result in the pairwise ablation.","section":"Section 5.3, Table 2"},{"comment":"The paper does not report essential experimental details needed to reproduce or assess the results, including train/validation/test split definitions, subject-independence, window size, overlap, normalization, batch size, number of epochs, early stopping criteria, and the number of independent runs. Table 3 reports only single accuracy values without standard deviations or significance tests. The authors should provide a complete protocol and, ideally, release code to make the comparisons verifiable.","section":"Section 4, general experimental protocol"}],"minor_comments":[{"comment":"The dataset name is repeatedly written as 'OPPO' instead of 'OPPORTUNITY'; please correct this throughout.","section":"Sections 5.1, 5.2, 5.4"},{"comment":"The text refers to 'The ELK method [39]', but reference [39] is the DanHAR paper; ELK is not introduced in the related work and has no correct citation. Please fix the reference and describe ELK properly.","section":"References, Table 3"},{"comment":"The row '+ DP + DA' should presumably read '+ DPFE + DA' for consistency with the other rows; please correct the abbreviation.","section":"Table 2"},{"comment":"The text says the ablations were performed on both PAMAP2 and OPPORTUNITY, but Table 2 does not identify the dataset and appears to show only one set of results; please clarify which dataset Table 2 refers to and provide the OPPORTUNITY ablation results if they exist.","section":"Section 5.3"},{"comment":"The caption 'Figure 1: The total process of task' is ungrammatical and uninformative, and the section heading 'T otal Training Objective' has a typo; both should be corrected.","section":"Figure 1 and Section 3.3.1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: DCDP-HAR is a reasonable engineering combination—dual-path ResNet/DenseNet, multi-stage contrastive alignment, and a confidence-driven gradient modulation that adapts ideas from OGM-GE and AGM. If the results were credible, this would be a useful, if incremental, contribution to sensor-based HAR. But the empirical validation as written does not support the central claim.\n\nWhat is genuinely new: the specific integration of these three components in one HAR framework, and the use of prediction confidence (rather than gradient magnitude) to modulate branch gradients. The method section is clear at a high level, and the contrastive loss formulation is standard but clean. The paper also properly cites the gradient-modulation work it builds on, which I appreciate.\n\nThe problems start with the experimental protocol. Section 5.1 says hyperparameter tuning was done on OPPORTUNITY, and Section 5.2 says the finalized configuration was then used for all ablations and comparisons. So the OPPORTUNITY result (88.34% in Table 3) is not an independent test; it is a selection outcome. The same contamination propagates to the ablation study and, to a lesser degree, to the other datasets. This alone is a load-bearing flaw.\n\nThen there are internal inconsistencies that make it impossible to verify anything. Table 2 includes a 'DA' (Data Augmentation) component that never appears in the methodology. The full model's PAMAP2 accuracy is 93.90% in the ablation table but 93.71% in the comparison table. The best tuned OPPORTUNITY accuracy is reported as 88.62% in Section 5.2 but 88.34% in Table 3. Also, the ELK baseline is cited to [39], which is actually the DanHAR paper, so the reader cannot check that comparison.\n\nOne more soft spot: the dual-path architecture is motivated by heterogeneous sensor modalities, but the paper never specifies which sensors go to which branch for any dataset. Without that, the observed gain from the dual-path setup could just be extra capacity.\n\nThe ideas are not bad, and the flaws are fixable in principle: rerun with a proper split for tuning, define DA, fix the numbers, and specify the sensor partition. But as submitted, the results are not reliable. I would not send this to peer review in its current form. Worth a look only if the authors resubmit a corrected version.","headline":"A plausible combination of existing ideas, but the experimental protocol is fundamentally compromised by tuning on the evaluation set and by inconsistent reporting.","tokens_in":12990,"tokens_out":4204,"would_cite":false,"duration_ms":46481,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A dual-path network with confidence-scaled gradients and multi-stage contrastive alignment aims to stop dominant sensors from silencing weaker ones in multimodal human activity recognition, reporting accuracy gains on four public sensor…","keywords":["Human Activity Recognition","Multimodal learning","Cross-modal alignment","Gradient modulation","Contrastive learning","Dual-path network","Sensor-based HAR","Modality imbalance"],"falsifier":"Permute the channel partition randomly over several seeds while keeping everything else fixed and compare accuracy; if a random split matches the authors' chosen split, the dual-path mechanism is not doing modality-specific work. Also track per-branch gradient norms during early training: if the confidence-driven modulation never reduces the dominant branch's gradient relative to an unmodulated run, the claimed balancing effect is not operative.","tokens_in":12061,"feed_emoji":"🏃","tokens_out":18279,"duration_ms":182215,"temperature":0.7,"pith_summary":"This paper claims that the two persistent difficulties of sensor-based activity recognition—one sensor modality dominating the learning and heterogeneous signals failing to align in a common semantic space—can be addressed together in a single end-to-end framework. The proposed Dynamic Contrastive Dual-Path Network splits the sensor channels across two parallel feature extractors, aligns the two branches at multiple depths with contrastive losses, and then scales each branch's backpropagated gradients according to how confident it is on the true activity, so a dominating branch is temporarily damped. On four public activity datasets the full framework reports accuracy gains over plain CNN/LSTM baselines and recent multimodal HAR models, and the paper's ablation studies tie the gains to the dual-path design, the contrastive alignment, and the gradient modulation acting together. These results matter because real deployments combine accelerometers, gyroscopes, and other sensors, and the value of adding a sensor depends on preventing a stronger sensor from silencing a weaker one.","feed_headline":"Gradient scaling keeps strong sensors from drowning weak ones in HAR","feed_subtitle":"Confidence-weighted gradients stop dominant sensor channels from suppressing complementary ones, lifting accuracy on four public datasets.","key_machinery":"The load-bearing mechanism is confidence-driven gradient modulation (CGM): after each batch, the network sums the per-sample Softmax probabilities each branch assigns to the true class, computes a dominance ratio between the branches, and for the branch with ratio above 1 multiplies its classification gradient by $1-\\tanh(\\alpha\\,\\mathrm{ReLU}(R-1))$, so the more dominant branch is damped. The alignment side is carried by a multi-stage bidirectional contrastive loss over cosine-similarity matrices at $L$ intermediate projection layers, plus an MSE loss between the $\\ell^2$-normalized final branch features. The branch structure is a partitioned dual-path feature extractor: input channels are split into two disjoint sets $\\mathcal{I}_1 \\cup \\mathcal{I}_2 = \\{1,\\dots,F\\}$, processed by a residual-connection path and a dense-concatenation path, and globally pooled into fixed-size vectors for classification. The total objective combines three classification losses with the alignment losses as $\\mathcal{L}_{total} = (\\mathcal{L}_{cls,res}+\\mathcal{L}_{cls,dense}+\\mathcal{L}_{cls,fusion}) + \\lambda_{align}(\\mathcal{L}_{contrast}+\\mathcal{L}_{align})$. In the authors' account, this machinery lets the two branches learn complementary features while dynamically equilibrating their contributions to the final classifier.","core_discovery":"On its own terms, the paper's central claim is that DCDP-HAR improves multimodal human activity recognition by coupling three mechanisms: (1) a dual-path backbone in which a ResNet-style branch and a DenseNet-style branch each process a disjoint subset of the input channels, producing complementary representations; (2) a multi-stage contrastive learning scheme that treats the two branches' projected intermediate features from the same sample as positive pairs and other samples' features as negatives, averaging bidirectional contrastive losses over $L$ stages and adding an $\\ell^2$-normalized MSE alignment on the final features; and (3) a confidence-driven gradient modulation that accumulates each branch's true-class Softmax confidence over a batch, forms ratios $R_{\\mathrm{res}} = S_{\\mathrm{res}}/(S_{\\mathrm{dense}}+\\epsilon)$ and $R_{\\mathrm{dense}} = S_{\\mathrm{dense}}/(S_{\\mathrm{res}}+\\epsilon)$, and scales the dominant branch's classification gradient by $1 - \\tanh(\\alpha\\,\\mathrm{ReLU}(R-1))$, with momentum-based gradient accumulation for stability. The authors report that the complete system reaches 88.34% accuracy on OPPORTUNITY, 98.75% on WISDM, 97.79% on UCI-HAR, and 93.71% on PAMAP2, beating the compared CNN, LSTM, and hybrid baselines as well as the dual-attention comparison model, and that removing the gradient modulation from the full system lowers PAMAP2 accuracy from 93.90% to 90.57%.","pith_inferences":["The paper leaves it open which sensor channels feed the ResNet branch versus the DenseNet branch for any dataset; testing random channel partitions against the authors' implicit choice would show whether the dual-path gain is modality-specific or just extra capacity.","The confidence-ratio rule should act most strongly early in training, so tracking per-branch gradient norms across epochs could reveal when the modulation actually changes optimization; the paper does not report that trajectory.","The same ratio-based gradient modulation could be transferred to audio-visual or RGB-depth classification, where one modality also tends to dominate; a transfer test would be a natural extension."],"forward_implications":["If the reported results hold, modality imbalance in HAR can be treated at the gradient level rather than only through loss weighting or per-modality learning-rate schedules.","Enforcing contrastive alignment at several intermediate stages should reduce the information loss that comes from fusing only the final feature vectors.","The reported ablation shows the full framework at 93.90% and the same framework without gradient modulation at 90.57% on PAMAP2, so the balancing mechanism carries a substantial share of the claimed gain.","The comparison numbers place DCDP-HAR ahead of the CNN, LSTM, and hybrid baselines on all four datasets, including the sensor-rich and class-imbalanced OPPORTUNITY set."],"supporting_citations":[{"why":"The on-the-fly gradient modulation framework that the confidence-driven mechanism extends.","marker":"[21]"},{"why":"The contrastive loss formulation the multi-stage alignment adapts.","marker":"[30]"},{"why":"Provides the OPPORTUNITY benchmark used in the main experiments and comparison.","marker":"[32]"},{"why":"Provides the PAMAP2 benchmark used in ablations and the comparison table.","marker":"[33]"},{"why":"Provides the WISDM benchmark used in the comparison table.","marker":"[34]"},{"why":"Provides the UCI-HAR benchmark used in the comparison table.","marker":"[35]"},{"why":"Supplies the CNN baseline that DCDP-HAR is compared against.","marker":"[36]"},{"why":"Supplies the LSTM and LSTM-CNN baselines used in the comparison.","marker":"[37]"},{"why":"Supplies the CNN-GRU baseline used in the comparison.","marker":"[38]"},{"why":"The dual-attention multimodal HAR model that serves as the main recent-method comparison.","marker":"[39]"}],"fun_headline_variants":["Confidence-weighted gradients stop sensor dominance in HAR","Dual-path contrastive learning with gradient balance tops four HAR benchmarks","Modality competition eased by confidence-driven gradient modulation","DCDP-HAR: dynamic contrastive dual-path network for robust multimodal HAR","Gradient scaling based on confidence improves multimodal activity recognition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework's central bet is that splitting the sensor channels into two fixed sets, with one set processed by a ResNet branch and the other by a DenseNet branch, produces complementary features; because the paper never says which channels go in which set, the benefit could come from simply having a bigger model.","fun_headline_variants_meta":{"raw":{"variants":["Confidence-weighted gradients stop sensor dominance in HAR","Dual-path contrastive learning with gradient balance tops four HAR benchmarks","Modality competition eased by confidence-driven gradient modulation","DCDP-HAR: dynamic contrastive dual-path network for robust multimodal HAR","Gradient scaling based on confidence improves multimodal activity recognition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000648,"raw_usage":{"total_tokens":3030,"prompt_tokens":1052,"completion_tokens":1978,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":1909}},"tokens_in":668,"tokens_out":1978,"duration_ms":16970,"temperature":1.0,"reasoning_tokens":1909,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:19:45.073585+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Permute the channel partition randomly over several seeds while keeping everything else fixed and compare accuracy; if a random split matches the authors' chosen split, the dual-path mechanism is not doing modality-specific work. Also track per-branch gradient norms during early training: if the confidence-driven modulation never reduces the dominant branch's gradient relative to an unmodulated run, the claimed balancing effect is not operative.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The on-the-fly gradient modulation framework that the confidence-driven mechanism extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The contrastive loss formulation the multi-stage alignment adapts."},{"cited_title":"Chavarriaga, H","cited_arxiv_id":null,"evidence_quote":"Provides the OPPORTUNITY benchmark used in the main experiments and comparison."},{"cited_title":"Anguita, A","cited_arxiv_id":null,"evidence_quote":"Provides the UCI-HAR benchmark used in the comparison table."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LSTM and LSTM-CNN baselines used in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The dual-attention multimodal HAR model that serves as the main recent-method comparison."}],"review_version":1}