{"id":"70bba59e-8dc5-4656-a886-f7e5383112db","arxiv_id":"2502.09993","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A training framework that reweights samples by the gap between ground-truth and nearest-negative prediction scores, plus flip consistency, improves FER accuracy under label noise and class imbalance.","lead":"This paper proposes a training scheme for facial expression recognition that gives each image a loss weight based on how close the model's top two predicted probabilities are, plus a consistency loss between original and mirrored images. It reports accuracy gains over prior methods on three in-the-wild datasets, with the largest gains on minority expressions like fear and disgust.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"NAW's hand-set Gaussian parameters (µ_t=[0.5,0.5], µ_f=[0.3,0.15], σ_11=0.8, 2:1/6:1 axes, scheduler 10) are untested for sensitivity, so the core ambiguity estimator may be overfit to RAF-DB; the SOTA margins in Tables 1–2 are too small to rule this out.","rationale":"The reader's CONDITIONAL verdict is appropriate. My stress-test focuses on the same load-bearing assumption: the hand-set Gaussian parameters in Eq. (4)–(6) are the mechanism that separates clean, ambiguous, and noisy samples, and all downstream gains are attributed to it. The paper reports no grid search, no ablation over these constants, and no evidence that the chosen covariance shapes match the actual score geometry on datasets other than the one used for development. Because Tables 1 and 2 report single runs without variance, the small margins over MEK cannot distinguish genuine transfer from tuning artifacts. I also note that Eq. (7) is not mathematically well-defined as written, since KL divergence requires normalized distributions while raw logits are unnormalized; this should be corrected to softmax probabilities, but I treat it as a reproducibility issue rather than the primary attack because it may be a typo in an otherwise plausible implementation. The paper has real strengths: it targets a genuine gap, the ablation isolates NAW's contribution, and the noisy/imbalanced experiments include variance. Those strengths keep the work within 'conditional accept' territory. The proposed sensitivity sweep would settle whether the Gaussian hyperparameters are universal; if they are not, the paper should reposition the claim as dataset-specific tuning rather than a general ambiguity-navigation principle. I would not change the reader's CONDITIONAL verdict.","tokens_in":11873,"tokens_out":4293,"duration_ms":41641,"concrete_test":"Run a controlled sensitivity sweep on RAF-DB and AffectNet with fixed training budget: vary σ_11 ∈ {0.4, 0.6, 0.8, 1.0}, true/false axis ratios {(2:1,6:1), (1:1,4:1), (3:1,8:1)}, scheduler constant ∈ {5, 10, 20}, and if feasible µ_f ∈ {[0.2,0.1], [0.3,0.15], [0.4,0.2]}. Report overall and mean accuracy with 3 seeds. If the best configuration differs by dataset or changing σ_11 by ±0.2 shifts overall accuracy by more than about 0.5%, the universal-mapping premise fails. Additionally, re-run ablation (f) with Eq. (7) computed on softmax probabilities to verify the consistency term is the one described.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim stands or falls on Eq. (4)–(6): NAW reweights every sample via a fixed multivariate Gaussian over (p_GT, p_NN), with means µ_t=[0.5,0.5] and µ_f=[0.3,0.15], σ_11=0.8, axis ratios 2:1 and 6:1, and scheduler constant 10. The paper says these are 'experimentally set' but gives no sensitivity study, no justification that the same kernel is optimal across RAF-DB, AffectNet, FERPlus, noise levels, and imbalance factors. The headline RAF-DB overall gain over MEK is only 0.16 and the mean gain is 1.43; on AffectNet the overall gain is 1.33. These margins are within the range that hyperparameter tuning on the target benchmark could explain. A secondary reproducibility issue compounds this: Eq. (7) defines consistency regularization as KL divergence between raw logits z_i and (z_i+z'_i)/2, but KL/JSD are only defined for normalized distributions; if the implementation actually uses softmax probabilities, the paper should say so. Neither issue by itself proves the method is wrong, but together they leave the strongest claim—universal robustness to ambiguity, noise, and imbalance—not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NLA, a framework for in-the-wild facial expression recognition that jointly addresses label ambiguity, label noise, and class imbalance. NLA combines a Noise-aware Adaptive Weighting (NAW) scheme, which reweights the cross-entropy loss via a multivariate Gaussian kernel over the prediction scores for the ground-truth class and the nearest negative class, with a Jensen-Shannon consistency regularization between original and horizontally flipped images. Experiments on RAF-DB, AffectNet, and FERPlus compare NLA with prior methods, report robustness under 10–30% label noise and imbalance factors 50–150, and include ablations and visualizations. The central claims are that NLA surpasses the state of the art in overall and mean accuracy and is the first framework to handle both ambiguity-related noise and class imbalance.","tokens_in":12386,"tokens_out":3751,"duration_ms":38352,"significance":"If the results hold, NLA would be a practically useful contribution because it is a single-backbone method that improves minority-class accuracy without sacrificing majority-class accuracy, and the noise/imbalance experiments (Tables 3–4) report five-seed means with standard deviations, which is a strength. The ablation study (Table 5) indicates that both NAW-CE and JSD consistency contribute to the final accuracy, and the paper provides qualitative evidence in Figs. 5–7. However, the load-bearing parts of the method—the hand-specified Gaussian parameters in Eq. (4) and the exact form of the consistency loss in Eq. (7)—are not yet established with enough rigor, and the headline SOTA margins in Tables 1–2 are small. The central idea is plausible and not circular, since gains are measured on held-out test sets, but the evidence for universal transfer across datasets and noise/imbalance levels is incomplete.","major_comments":[{"comment":"The NAW weighting in Eq. (4) depends on several parameters that the paper says are 'experimentally set': μ_t=[0.5,0.5], μ_f=[0.3,0.15], σ_11=0.8, axis ratios 2:1 and 6:1, scheduler constant 10, and λ=0.5. No sensitivity analysis is provided, and the paper does not justify why this fixed kernel transfers across RAF-DB, AffectNet, FERPlus, different noise levels, and different imbalance factors. This is load-bearing because Eq. (4) is the core ambiguity estimator, and the overall gains over MEK in Tables 1–2 are only 0.16 on RAF-DB and 1.33 on AffectNet, margins that could be explained by hyperparameter tuning on the target benchmark. I request a sensitivity study over these parameters (or an argument, ideally with evidence, that the chosen values are not dataset-specific).","section":"Method, Eq. (4); Implementation Details"},{"comment":"The consistency loss is written as D_KL(z_i || (z_i+z'_i)/2) + D_KL(z'_i || (z_i+z'_i)/2), where z_i and z'_i are described as 'output logits'. Kullback-Leibler divergence and Jensen-Shannon divergence are defined for probability distributions, not raw logits, so Eq. (7) as written is undefined. If the implementation applies a softmax before computing the divergence, the paper must state this and write the equation accordingly; if it truly uses logits, the formula is not a valid KL/JSD and needs correction. Since Table 5 shows that the JSD term contributes to the final performance (settings (e) vs. (f)), this is not a purely cosmetic issue.","section":"Eq. (7), Consistency Regularization"},{"comment":"The main SOTA comparisons on RAF-DB and AffectNet report single runs with no standard deviation or repeated-seed statistics, unlike Tables 3–4 which do provide five-seed error bars. Given that the margins over the previous best method are small (0.16 overall on RAF-DB and 1.33 on AffectNet), the claim that NLA 'surpasses' the state of the art is not yet statistically supported. Please report mean and standard deviation over multiple seeds for the main benchmark tables, or at least clarify the run-to-run variability.","section":"Tables 1–2, Comparison with Existing Methods"}],"minor_comments":[{"comment":"If the intended quantity is the standard Jensen-Shannon divergence, the formula is missing the factor 1/2 in front of each KL term; either define the loss as the unnormalized sum explicitly or use the standard JSD definition.","section":"Eq. (7)"},{"comment":"The reference list contains duplicate entries for Cui et al. (2019a) and Cui et al. (2019b), which are the same paper; please merge or disambiguate them.","section":"Related Work / References"},{"comment":"The text states that the best performance is observed at epoch 40 out of a maximum of 60 epochs; please specify whether model selection is performed on a validation split or on the test set, as this affects the interpretation of all reported numbers.","section":"Implementation Details"},{"comment":"The column header for the ablation table is difficult to parse; the table would be clearer if the components were listed vertically and each setting indicated which components are enabled.","section":"Table 5"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is interesting and the experimental setup includes useful robustness experiments, but the small margins in the main tables and the lack of sensitivity analysis for the hand-set Gaussian parameters make the main claims fragile. The undefined consistency loss in Eq. (7) is a concrete reproducibility concern. I believe these can be addressed within the scope of a revision, so I recommend major_revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent, incremental contribution to robust FER training. The combination of sample reweighting based on (p_GT, p_NN) Gaussian kernels plus JSD consistency is new for FER, and the ablation evidence supports both components. But the headline SOTA claims rest on single-run comparisons with small margins, and the hand-set Gaussian parameters are never tested for sensitivity. That is the main soft spot; Eq. (7) writing KL on logits is a presentation bug, not a real flaw.\n\nWhat the paper does well: the NAW idea is simple and clearly motivated—ambiguous samples have p_GT ≈ p_NN, noisy samples have p_GT << p_NN, and the two Gaussian kernels encode that geometry. The covariance scheduler making the kernel elliptical over training is a neat detail. The ablations on RAF-DB show NAW-CE alone adds 0.75 overall / 2.87 mean, and JSD adds another chunk. The noise and imbalance experiments report five-seed means and std, and the gains there are consistent and larger than the main benchmark margins.\n\nWhere it's soft: Tables 1 and 2 show single runs. On RAF-DB the overall margin over MEK is 0.16, and on AffectNet 1.33; the mean accuracy margin on RAF-DB (1.43) is more meaningful because it reflects minority classes. More importantly, the Gaussian means, covariance axes, and scheduler constant are 'experimentally set' with no sensitivity analysis. Given the small margins, one cannot rule out that those parameters are tuned to the specific prediction statistics of RAF-DB and transfer well only by luck. A sensitivity sweep over µ and Σ, or at least the scheduler constant, would settle this. The other issue is Eq. (7): KL divergence on raw logits is undefined; the authors clearly mean softmax probabilities, and should just say so. Minor.\n\nThe claim of being 'first to address both problems simultaneously' is plausible but not deeply supported; the related work section is thin on methods that handle both, but I didn't find a missing citation that would change the conclusion.\n\nBottom line: the paper deserves a serious referee. The core idea is sensible, the ablations are honest, and the noisy/imbalanced experiments with stds suggest the effect is real. What it needs before publication is variance reporting on the main tables and a hyperparameter sensitivity study. If those come back fine, this is a useful addition to the FER toolkit.","headline":"A solid, incremental FER training method whose SOTA claims rest on small single-run margins and hand-tuned Gaussian hyperparameters that need a sensitivity study.","tokens_in":12729,"tokens_out":1910,"would_cite":true,"duration_ms":18751,"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":"This paper claims that a single per-sample weighting mechanism can handle both noisy labels and class imbalance in facial expression recognition, and reports state-of-the-art results on RAF-DB and AffectNet.","keywords":["facial expression recognition","label ambiguity","noisy labels","class imbalance","adaptive weighting","consistency regularization","in-the-wild datasets","multivariate Gaussian kernel"],"falsifier":"On FERPlus, each image has multiple annotator labels, so genuine ambiguity can be measured directly as the entropy of the label distribution; if the weights NLA assigns are not highest for the high-entropy (genuinely ambiguous) images, or if the reported gains vanish when the Gaussian constants are varied across datasets, the ambiguity-tracking premise is falsified.","tokens_in":11670,"feed_emoji":"😀","tokens_out":8225,"duration_ms":79582,"temperature":0.7,"pith_summary":"Facial expression recognition in the wild suffers from two problems at once: labels are noisy because expressions are subjective, and classes are imbalanced because some expressions are rare. The paper claims that both can be handled by one mechanism: dynamically estimating how ambiguous each sample is and weighting the loss accordingly. Its framework, NLA, computes two prediction scores per image (for the given label and for the closest competing label) and uses a Gaussian kernel over these scores to raise the weight of ambiguous samples and lower the weight of noisy ones, plus a consistency regularizer. On RAF-DB it reports 89.93% overall and 83.87% mean accuracy, and on AffectNet 67.06%, beating the previous state of the art in both, with the largest gains in minority classes. If the claim holds, a single reweighting rule can reduce the majority-class bias that standard cross-entropy training produces.","feed_headline":"Ambiguity-aware weighting lifts rare-expression accuracy past 70%","feed_subtitle":"One framework handles noisy labels and class imbalance together, beating prior art on RAF-DB and AffectNet.","key_machinery":"The load-bearing mechanism is the Noise-aware Adaptive Weighting (NAW) kernel: a multivariate Gaussian $w^*(\\mathbf{p}|\\mu,\\Sigma)$ evaluated on the pair $\\mathbf{p}=(p^{GT}, p^{NN})$ and multiplied into the cross-entropy loss as $(1+w^*)\\,L_{CE}$. Separate centers $\\mu_t=[0.5,0.5]^T$ for correct predictions and $\\mu_f=[0.3,0.15]^T$ for incorrect predictions, with covariance shapes set by axis ratios 2:1 and 6:1 and a scheduler $CS(e,E)=1-\\exp(-10e/E)$, determine which samples are treated as ambiguous versus noisy. This kernel is what lets the method emphasize minority-class ambiguous samples late in training; the consistency regularizer (Jensen-Shannon divergence between the logits of original and flipped images) supports it by making the scores on which the kernel depends more stable.","core_discovery":"NLA's central proposal is that label ambiguity can be used as a training signal rather than treated as an obstacle. At every iteration the model extracts two numbers from its current softmax output for each sample: $p^{GT}$, the probability of the annotated class, and $p^{NN}$, the probability of the highest-scoring other class. The Noise-aware Adaptive Weighting (NAW) component places these two numbers into a multivariate Gaussian kernel, with one kernel for samples the model currently predicts correctly and another for samples it predicts incorrectly; the kernels' means and covariance shapes are chosen so that the up-weighted samples are the ambiguous ones, while confident and clearly noisy samples receive low weight. A covariance scheduler changes the Gaussian contour from isotropic to elongated over training, shifting emphasis toward ambiguous minority-class samples in later epochs, and a Jensen-Shannon consistency loss between original and flipped images stabilizes the score estimates. The reported consequence is that minority classes such as disgust, fear, and anger improve substantially, reaching over 70% accuracy on RAF-DB for the first time while majority-class accuracy is not sacrificed, yielding the best overall and mean accuracy on two in-the-wild benchmarks.","pith_inferences":["An extension the paper leaves untested is replacing the fixed Gaussian parameters with a learned estimator of ambiguity; if such an estimator adapts across datasets, the method would shed its main hand-tuned constants.","Because NAW uses only the model's own softmax scores, it should transfer to other long-tailed classification tasks with noisy labels; the paper's conclusion gestures at this but provides no evidence outside FER.","The ablation does not separate how much of the gain comes from the Gaussian shape versus the consistency regularizer; a plausible reading is that consistency stabilizes the scores the weights depend on, so a simpler weighting rule plus stronger augmentation could match NLA's results."],"forward_implications":["If NLA's results hold, in-the-wild FER can be improved on both overall and mean accuracy by per-sample loss reweighting alone, without separate noise-cleaning or resampling stages.","The largest gains land in the rarest expressions (disgust, fear, anger), so class-balanced evaluation becomes the right way to compare FER methods, not just overall accuracy.","The reported robustness under 10–30% label noise and imbalance factors of 50–150 suggests the same loss could be applied directly to newly collected, uncurated face data.","The ablation study indicates the weighting and the consistency regularizer reinforce each other; removing either one reduces both overall and mean accuracy, so the gains are not attributable to the Gaussian weighting alone."],"supporting_citations":[{"why":"Supplies the RAF-DB benchmark, the primary dataset for main results and imbalance experiments.","marker":"Li, Deng, and Du 2017"},{"why":"Supplies the AffectNet benchmark used for the second full comparison and noise tests.","marker":"Mollahosseini, Hasani, and Mahoor 2017"},{"why":"Supplies FERPlus, whose crowd labels are used for the noise-robustness comparison.","marker":"Barsoum et al. 2016"},{"why":"MEK is the state-of-the-art class-imbalance baseline that NLA claims to beat in minority classes.","marker":"Zhang et al. 2024"},{"why":"RUL is the relative-uncertainty baseline for label ambiguity and the protocol source for noise experiments.","marker":"Zhang, Wang, and Deng 2021"},{"why":"EAC supplies the erasing-attention-consistency noisy-label baseline and the flipped-image consistency idea that NLA adapts.","marker":"Zhang et al. 2022a"},{"why":"LA-Net is the landmark-aware noisy-label state of the art against which NLA reports average gains under 10–30% noise.","marker":"Wu and Cui 2023"},{"why":"Provides the small-loss/clean-sample assumption that motivates weighting 'easy' samples less.","marker":"Arpit et al. 2017"}],"fun_headline_variants":["Ambiguity turns from foe to ally for facial expression AI","One framework tames noisy labels and class imbalance in FER","Rare emotions get a boost: ambiguity-aware learning wins","NLA teaches models to embrace label ambiguity in the wild","From ambiguity to accuracy: new FER method beats prior art"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire method rests on the assumption that a fixed, hand-chosen Gaussian formula over two prediction scores separates clean, ambiguous, and noisy samples correctly in every dataset and at every training stage; if those constants only reflect patterns seen in RAF-DB, the reported gains will not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Ambiguity turns from foe to ally for facial expression AI","One framework tames noisy labels and class imbalance in FER","Rare emotions get a boost: ambiguity-aware learning wins","NLA teaches models to embrace label ambiguity in the wild","From ambiguity to accuracy: new FER method beats prior art"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000573,"raw_usage":{"total_tokens":2748,"prompt_tokens":1026,"completion_tokens":1722,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":642,"completion_tokens_details":{"reasoning_tokens":1638}},"tokens_in":642,"tokens_out":1722,"duration_ms":11892,"temperature":1.0,"reasoning_tokens":1638,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T19:47:54.080942+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On FERPlus, each image has multiple annotator labels, so genuine ambiguity can be measured directly as the entropy of the label distribution; if the weights NLA assigns are not highest for the high-entropy (genuinely ambiguous) images, or if the reported gains vanish when the Gaussian constants are varied across datasets, the ambiguity-tracking premise is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the AffectNet benchmark used for the second full comparison and noise tests."},{"cited_title":"C.; and Zhang, Z","cited_arxiv_id":null,"evidence_quote":"Supplies FERPlus, whose crowd labels are used for the noise-robustness comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MEK is the state-of-the-art class-imbalance baseline that NLA claims to beat in minority classes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RUL is the relative-uncertainty baseline for label ambiguity and the protocol source for noise experiments."},{"cited_title":"S.; Maharaj, T.; Fischer, A.; Courville, A.; Bengio, Y.; et al","cited_arxiv_id":null,"evidence_quote":"Provides the small-loss/clean-sample assumption that motivates weighting 'easy' samples less."}],"review_version":1}