{"id":"c0319a6c-fb2c-4417-b2dc-dcbc030360b4","arxiv_id":"2505.13342","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"D&C-Net detects noisy labels with a Gaussian mixture over per-sample losses and selectively corrects them using an estimated transition matrix.","lead":"This paper introduces D&C-Net, a two-stage method that separates samples with noisy labels from clean ones using a Gaussian mixture on per-sample losses, then applies a learned label-noise transition matrix only to the flagged noisy samples. The method reports accuracy gains on MNIST, CIFAR-10, CIFAR-100, and a single-cell RNA-seq dataset, but its own significance analysis shows the improvement is significant in only three of six settings.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own significance table contradicts the central claim of consistent, significant improvement; the strongest results are significant in only 3 of 6 settings, with one baseline winning significantly.","rationale":"The reader correctly identifies a serious weakness in the transition-matrix initialization (Sec. 3.1), where classifier predictions on flagged noisy samples are treated as true labels. That is a real correctness risk: if classifier errors correlate with the noise process, the estimated T can be biased and the correction could amplify noise. However, the most load-bearing concern about the paper's central claim is even more direct: the paper's own significance analysis contradicts the claim of consistent, significant improvement. Table 6 shows that D&C-Net is significantly better in only half the settings, ties in two, and loses significantly in one. This is an internal inconsistency between the abstract's \"significant improvement\" and the reported p-values, independent of any assumptions about the noise process. The missing transition-matrix update rule (Algorithm 2 line 5) is a further completeness problem that blocks reproduction and makes it impossible to verify the method from the text. Together, these issues justify the reader's REJECT verdict; my analysis does not change that verdict. I set agreement_with_reader to partial because the reader's weakest_assumption focuses on initialization bias, whereas I find the significance contradiction to be the decisive flaw, though the reader's rationale does also mention the overstated claim and the missing update rule.","tokens_in":10348,"tokens_out":3137,"duration_ms":30375,"concrete_test":"Run the released code (https://github.com/yuvalgrin/DetectAndCorrect-Net) to obtain per-seed accuracies for all six settings in Table 6, then recompute paired significance across the three seeds (e.g., paired t-test or Wilcoxon signed-rank test) and apply a Benjamini-Hochberg correction for multiple comparisons. If fewer than 3 of 6 settings remain significant, or if VolMinNet remains significantly better in CIFAR-100/Pair-20% after correction, the abstract's claims of \"significant improvement\" and \"consistent outperformance\" must be downgraded. Additionally, inspect the training loop to identify the missing transition-matrix update rule; if no update exists, the method is incompletely specified and the results are not reproducible from the text alone.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, stated in the abstract and Section 4.3, is that D&C-Net \"consistently outperforms leading baselines\" and shows \"significant improvement\" in accuracy, particularly under high noise. This claim is directly undermined by the paper's own statistical significance analysis in Table 6. Comparing D&C-Net to the strongest baseline VolMinNet using Welch's t-test, D&C-Net is significantly better (p<0.05) in only 3 of 6 settings: CIFAR-10/Sym-50%, CIFAR-100/Sym-50%, and CIFAR-10/Pair-20%. In CIFAR-10/Sym-20% (p=0.100) and MNIST/Sym-50% (p=0.170) the difference is not significant, and in CIFAR-100/Pair-20% VolMinNet is significantly better (p=0.0085). Thus the evidence supports at best a selective advantage, not consistent outperformance. Moreover, because only three seeds were used, the t-test has very low power; p-values near 0.05 could easily flip with additional runs. A second, compounding issue is that Algorithm 2 line 5 instructs to \"Update transition matrix T\" but no update rule appears anywhere in the paper, so the reported results cannot be attributed to a fully specified algorithm. These problems are internal to the manuscript rather than disagreements with external consensus, and they directly affect the truth of the headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes D&C-Net, a two-stage method for learning with noisy labels. In a pre-training stage, the method trains a network with a cyclical learning rate, fits a two-component Gaussian mixture model to the per-sample loss distribution, and selects a threshold to flag likely noisy labels. It then initializes a noise transition matrix by counting, among flagged samples, co-occurrences of the observed noisy label and the classifier's predicted label. In the training stage, the method applies a corrected cross-entropy loss only to the flagged noisy samples, leaving clean samples with the standard loss, and the transition matrix is said to be continuously updated. Experiments are reported on MNIST, CIFAR-10, CIFAR-100, and an scRNA-seq PBMC cell-type annotation dataset, comparing against several baselines. The paper claims consistent and significant accuracy improvements, especially under high noise.","tokens_in":10671,"tokens_out":4362,"duration_ms":42952,"significance":"If the method were fully specified and the claimed improvements held, the selective-correction idea—correcting only detected noisy samples while preserving clean-sample losses—would be a practically useful contribution to the noisy-labels literature, and the scRNA-seq application is a worthwhile extension. The paper provides several strengths: a clear motivation, comparison with eight baselines, an ablation study, and a public code repository. However, the central claim of consistent, significant improvement is not supported by the paper's own statistical analysis, and a core component of the algorithm (the transition-matrix update rule) is left undefined, which limits the reproducibility and interpretability of the reported results. The transition-matrix initialization also rests on an unvalidated self-training assumption. These issues substantially reduce the significance of the contribution as it currently stands.","major_comments":[{"comment":"Algorithm 2 instructs to 'Update transition matrix T' at every iteration, and Section 3.2 states that 'The transition matrix T is continuously updated during training to adapt to evolving noise patterns,' but no update rule, objective, or optimization procedure is provided anywhere in the manuscript. The experimental setup mentions a learning rate of 10^-4 'for the transition matrix,' which suggests a gradient-based update, yet T is not a network parameter and no loss with respect to T is defined. Without this specification, the reported results cannot be reproduced, and the method cannot be separated from untested implementation choices.","section":"Algorithm 2, line 5; Section 3.2"},{"comment":"The claim that D&C-Net 'consistently outperforms leading baselines' (Section 4.3) and shows 'significant improvement' (abstract) is contradicted by the paper's own Welch t-test results in Table 6. D&C-Net is significantly better than VolMinNet (p < 0.05) in only 3 of 6 settings, VolMinNet is significantly better in CIFAR-100 Pair-20% (p = 0.0085), and two settings (CIFAR-10 Sym-20%, p = 0.100; MNIST Sym-50%, p = 0.170) show no significant difference. Moreover, with only three seeds, the t-test has low power, so even the significant findings are fragile. The evidence supports at most a selective advantage, not consistent superiority.","section":"Section 4.3 and Supplementary 6.1"},{"comment":"The transition matrix is initialized by treating the classifier's predictions on flagged noisy samples as the true labels: T_{l,k} = sum_n I(\\tilde{y}_n = k, \\hat{y}_n = l, b_n = 1), followed by row normalization. The paper states 'we assume the classifier is sufficiently accurate to provide a meaningful estimate of the noise structure,' but no validation of this assumption is provided. If the classifier's errors on noisy samples are correlated with the noise process, the estimated T is biased, and the correction step in Section 3.2 may amplify rather than fix the noise. The paper's own Limitations section (5.2) acknowledges reliance on accurate detection but does not address this circularity or offer a diagnostic to check it.","section":"Section 3.1, Transition Matrix Initialization"}],"minor_comments":[{"comment":"The threshold t is defined in Section 3.1 by maximizing the average of sensitivity and specificity, but the caption of Figure 1 states that t is 'determined by the intersection of the two distributions.' These two definitions are not generally equivalent; please clarify which one is used in the experiments and whether the figures depict the actual procedure.","section":"Section 3.1 and Figure 1/Figure 2 captions"},{"comment":"Algorithm 1 returns model parameters θ from pre-training, but Algorithm 2 does not state whether the final training starts from these parameters or from a fresh initialization. This affects the interpretation of the two-stage procedure.","section":"Algorithm 1 and Algorithm 2"},{"comment":"The ablation reports an accuracy decrease of approximately 0.5% on CIFAR-10 when the transition matrix is not initialized, but it does not specify the noise type and level, the number of seeds, or whether this difference is statistically significant.","section":"Section 4.5 Ablation"},{"comment":"The scRNA-seq results are not accompanied by a significance test. For Sym-50%, the difference between D&C-Net (93.93±0.05) and VolMinNet (93.75±0.08) is very small and may not be significant.","section":"Table 5"},{"comment":"Reference [9] is listed as 'T-Reivision' in the text of Section 4.3; this appears to be a typo for 'T-Revision'. Please check all reference names for consistency.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript's own supplementary significance analysis (Table 6) directly undercuts the abstract's claim of consistent significant improvement, and Algorithm 2 lacks the central update rule for the transition matrix. These are not merely presentational issues; they concern the validity and reproducibility of the main contribution. Unless the authors can provide a fully specified algorithm and substantially stronger evidence, the paper is unlikely to meet the standard for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a workmanlike incremental contribution to noisy-label learning, and the central twist—apply a learned transition matrix only to samples flagged as noisy—is genuinely sensible. But the manuscript as submitted is not complete enough to support its abstract, and the authors' own significance table cuts against their headline claim.\n\nWhat's new and good: the selective correction idea is a clean way to sidestep the known downside of global transition-matrix methods, which perturb clean labels. Combining the O2U-Net-style loss trajectory analysis with a GMM clean/noisy split and a transition-matrix correction is a straightforward composition, but the selective application is the kind of small twist that can actually matter in practice. The scRNA-seq experiment is a nice extra, something beyond the usual MNIST/CIFAR grind, and the code link is a real plus.\n\nSoft spots, in proportion:\n\nThe biggest problem is that Algorithm 2 says 'Update transition matrix T' but never defines the update rule. A reader cannot reproduce the method. The learning rate for the transition matrix is mentioned in the setup, so there clearly is an update, but it is not written down. That is a load-bearing omission, not a cosmetic one.\n\nThe significance story is also overstated. The abstract claims 'significant improvement'; Table 6 says D&C-Net wins significantly in 3 of 6 settings, loses significantly in 1, and ties in 2. With only three seeds, the tests have low power. A fair claim would be 'selective advantage at high noise,' not consistent outperformance.\n\nThe transition-matrix initialization assumes the pre-trained classifier's predictions on flagged noisy samples are good enough proxies for true labels. The authors acknowledge this, but they don't give any diagnostic showing the estimated T resembles the real noise process. For symmetric noise this is a mild concern; for pair noise it could bias the correction.\n\nWho should read it: people actively working on noisy labels who want a practical recipe could get something from the selective-correction idea and the code. It doesn't open a new direction. It deserves peer review because the idea is workable and a competent revision could fix the specification and reframe the claims. I'd send it out, but I'd expect major revision.","headline":"A workmanlike incremental method with a sensible selective-correction twist, but the missing transition-matrix update rule and an overstated significance claim keep it from being complete as submitted.","tokens_in":11146,"tokens_out":2086,"would_cite":false,"duration_ms":19835,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"D&C-Net claims that selectively applying a transition-matrix loss correction only to samples flagged as noisy, while leaving clean samples on standard cross-entropy, yields higher accuracy under label noise than global correction or…","keywords":["D&C-Net","Noisy Labels","Transition Matrix","Image Classification","Single-cell RNA Sequencing","Cell Annotation","Label Noise Correction","Selective Correction"],"falsifier":"Run D&C-Net on a synthetic noisy dataset whose true transition matrix is known, record the matrix it estimates at initialization and after training, and compute the entrywise distance to the truth; if the estimated matrix is far from the true one even in the symmetric-noise settings where the paper reports its largest accuracy gains, the claimed correction mechanism is not what drives the improvement.","tokens_in":10173,"feed_emoji":"🎯","tokens_out":10770,"duration_ms":94849,"temperature":0.7,"pith_summary":"The paper aims to establish that the two usual ways of coping with noisy training labels—global noise estimation with a transition matrix and filtering that discards suspicious samples—can be merged. Its method, D&C-Net, first flags likely mislabeled samples by fitting a two-component Gaussian mixture to per-sample losses, then estimates a noise transition matrix from those flagged samples and applies the corrected loss only to them, leaving clean samples on the ordinary loss. Across MNIST, CIFAR-10, CIFAR-100, and a single-cell RNA-seq cell-type annotation dataset, the paper reports accuracy gains over the compared baselines in most settings, with the largest margins at 50% symmetric label noise. If the claim is right, noisy samples are no longer wasted: their information is reused through correction, while clean labels are never unnecessarily perturbed.","feed_headline":"Noisy labels? Correct them selectively instead of deleting them","feed_subtitle":"On CIFAR-10 at 50% symmetric noise it reaches 87.3% accuracy, 3.9 points above the prior best","key_machinery":"The load-bearing pair is the binary noise flag $b_n$ and the row-stochastic transition matrix $T$. The flag comes from a Gaussian-mixture fit to per-sample losses, and $T$ is initialized by counting, among flagged samples, how often the observed noisy label $\\tilde y_n$ differs from the classifier's prediction; this is the step whose reliability the paper explicitly ties to classifier accuracy. During training, $b_n$ decides which loss to use and $T$ supplies the class-to-class flip probabilities inside the corrected loss for $b_n=1$. The paper's distinctive move is the localization: $T$ is updated continuously but applied only to flagged samples, so clean samples see no correction.","core_discovery":"The central claim is that loss correction via a noise transition matrix should be applied selectively, not globally. The paper identifies suspect samples by tracking per-sample losses under a cyclic learning rate and fitting a Gaussian mixture with two components; the threshold that separates clean from noisy is chosen to maximize the average of sensitivity and specificity. Only the flagged samples are used to initialize the transition matrix $T$, and during training a flagged sample $n$ is given the corrected cross-entropy loss $-\\log \\sum_{l=1}^{C} T_{l,\\tilde y_n} P_\\theta(Y=l\\mid x_n)$, while an unflagged sample keeps the standard cross-entropy. This targeted correction preserves the learning signal in noisy examples without altering clean ones. In the reported experiments the largest wins appear under 50% symmetric noise, for example 87.30% versus 83.37% for the strongest baseline on CIFAR-10.","pith_inferences":["The paper's own limitation discussion implies the method will lose ground when the clean and noisy loss distributions overlap, as under instance-dependent or structured noise; a natural stress test is to measure the overlap of the two fitted Gaussians and correlate it with the accuracy gain.","Because $T$ is updated using only flagged samples, the correction could feed the classifier's own mistakes back into the matrix; the paper does not compare the recovered $T$ with the true noise process, which would separate bias in the matrix from the method's overall benefit.","A direct ablation the authors do not run is a version of D&C-Net that applies the same updated $T$ to every sample; the accuracy gap between the selective and global variants would isolate exactly how much the selective application contributes."],"forward_implications":["A pre-training phase with a cyclical learning rate yields both a noise indicator and a transition-matrix initialization, so the method does not need to know the noise rate in advance.","Noisy samples are corrected rather than removed, so the effective training set is larger than under filtering, and correctly labeled samples are never loss-modified as they are under global transition-matrix methods.","If the reported numbers hold, transition-matrix correction becomes usable in high-noise regimes where global correction usually struggles, with CIFAR-10 at 50% symmetric noise improving from 83.37% to 87.30%.","On biological data, the same pipeline improves cell-type annotation accuracy on a PBMC single-cell RNA-seq dataset across all tested noise settings, suggesting label-noise correction can be applied where annotation is costly and noisy."],"supporting_citations":[{"why":"Supplies the Gaussian-mixture separation of clean and noisy samples that D&C-Net adapts to its loss-distribution analysis.","marker":"[18]"},{"why":"Provides the loss-trajectory idea for detecting noisy labels that motivates the pre-training stage.","marker":"[13]"},{"why":"Defines the cyclic learning-rate schedule used to amplify loss differences between clean and noisy samples.","marker":"[35]"},{"why":"Formulates the transition-matrix loss correction that D&C-Net applies selectively to flagged samples.","marker":"[30]"},{"why":"Is a transition-matrix estimation baseline D&C-Net compares against and improves on in the experiments.","marker":"[43]"},{"why":"Is the strongest baseline in most experiments, making it the main comparison point for the reported accuracy gains.","marker":"[21]"},{"why":"Defines the pair-noise setting and represents the filtering approach whose discarded data D&C-Net aims to preserve.","marker":"[8]"},{"why":"Provides the Azimuth PBMC reference dataset used for the single-cell RNA-seq cell-type annotation experiment.","marker":"[27]"},{"why":"Supplies the totalVI integration that maps the scRNA-seq cells into the 20-dimensional space used for classification.","marker":"[7]"}],"fun_headline_variants":["Selective noise correction beats global approaches on noisy labels","Target noisy labels for correction, not deletion, to boost accuracy","Loss-based selection improves noisy-label training by 3.9 points","Fix only the noisy samples: a selective correction method","Selective correction preserves noisy data and lifts accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole correction rides on the pre-trained classifier being accurate enough that its prediction for a flagged noisy sample is usually the true label; if classifier mistakes are correlated with the noise process, the estimated transition matrix is biased and the correction step can amplify the noise instead of fixing it.","fun_headline_variants_meta":{"raw":{"variants":["Selective noise correction beats global approaches on noisy labels","Target noisy labels for correction, not deletion, to boost accuracy","Loss-based selection improves noisy-label training by 3.9 points","Fix only the noisy samples: a selective correction method","Selective correction preserves noisy data and lifts accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000677,"raw_usage":{"total_tokens":3067,"prompt_tokens":920,"completion_tokens":2147,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":2067}},"tokens_in":536,"tokens_out":2147,"duration_ms":14243,"temperature":1.0,"reasoning_tokens":2067,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:14:26.774414+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run D&C-Net on a synthetic noisy dataset whose true transition matrix is known, record the matrix it estimates at initialization and after training, and compute the entrywise distance to the truth; if the estimated matrix is far from the true one even in the symmetric-noise settings where the paper reports its largest accuracy gains, the claimed correction mechanism is not what drives the improvement.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian-mixture separation of clean and noisy samples that D&C-Net adapts to its loss-distribution analysis."},{"cited_title":"O2U-Net: A Simple Noisy Label Detection Approach for Deep Neural Networks","cited_arxiv_id":null,"evidence_quote":"Provides the loss-trajectory idea for detecting noisy labels that motivates the pre-training stage."},{"cited_title":"Cyclical learning rates for training neural networks","cited_arxiv_id":null,"evidence_quote":"Defines the cyclic learning-rate schedule used to amplify loss differences between clean and noisy samples."},{"cited_title":"Making deep neural net- works robust to label noise: A loss correction approach","cited_arxiv_id":null,"evidence_quote":"Formulates the transition-matrix loss correction that D&C-Net applies selectively to flagged samples."},{"cited_title":"Dual t: Reducing estimation error for transition matrix in label-noise learning","cited_arxiv_id":null,"evidence_quote":"Is a transition-matrix estimation baseline D&C-Net compares against and improves on in the experiments."},{"cited_title":"Provably End-to-end Label- Noise Learning without Anchor Points","cited_arxiv_id":null,"evidence_quote":"Is the strongest baseline in most experiments, making it the main comparison point for the reported accuracy gains."},{"cited_title":"Co-teaching: Ro- bust training of deep neural networks with ex- tremely noisy labels","cited_arxiv_id":null,"evidence_quote":"Defines the pair-noise setting and represents the filtering approach whose discarded data D&C-Net aims to preserve."},{"cited_title":"Integrated analysis of multimodal single-cell data","cited_arxiv_id":null,"evidence_quote":"Provides the Azimuth PBMC reference dataset used for the single-cell RNA-seq cell-type annotation experiment."},{"cited_title":"Joint probabilistic modeling of single-cell multi-omic data with totalVI","cited_arxiv_id":null,"evidence_quote":"Supplies the totalVI integration that maps the scRNA-seq cells into the 20-dimensional space used for classification."}],"review_version":1}