{"id":"26958437-50ce-419b-9439-8a6886ffb69e","arxiv_id":"2509.00745","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A skewness-guided pruning method removes skin-tone-related components in skin lesion classifiers, improving fairness and reducing computational cost.","lead":"This paper proposes a pruning method that removes skin-tone-related channels, patches, and attention heads in skin lesion classifiers, guided by the skewness of feature maps. It reports fairness improvements with smaller models, which could make fairer medical imaging AI more practical.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fairness gains are small, unreplicated, and selected post hoc from six configurations; without error bars or multiple seeds, the central claim of improved fairness is not statistically established.","rationale":"The paper proposes a plausible mechanism—negative skewness in feature/attention maps corresponds to skin-tone-focused components—and pruning those components could in principle reduce bias. However, the central claim is an empirical one: that pruning improves fairness. The evidence provided is limited to single-run point estimates with small absolute changes (1–2 percentage points in EOdd/EOpp1), no error bars, no repeated seeds, and post hoc selection of the best ViT configuration. These are textbook warning signs that the observed differences may be noise. The reader identified the unverified skewness assumption as the weakest link; I agree that this is a real gap, but I consider the statistical reliability of the empirical gain to be even more load-bearing. Even if the skewness assumption were rigorously verified, the current data would not convincingly demonstrate that the method improves fairness. Conversely, if the gains are not significant, the method's contribution is questionable regardless of its mechanism. Thus the manuscript needs additional experiments and statistical reporting before the claims can be accepted. The reader's verdict of CONDITIONAL is appropriate; my concern does not change that verdict but sharpens the required conditions for acceptance.","tokens_in":14344,"tokens_out":3824,"duration_ms":49551,"concrete_test":"Run each pruning configuration (VGG SkewPrune and ViT Patterns 2–6) with at least 5 random seeds using identical hyperparameters, and report the mean and 95% confidence interval for EOdd, EOpp0, EOpp1, accuracy, and F1 on the same test split. Then perform a paired bootstrap or Wilcoxon signed-rank test comparing each pruned model against the vanilla model on the same test set, with multiple-testing correction across configurations. If the EOdd or EOpp1 improvement over vanilla is not significant (e.g., corrected p ≥ 0.05) or the confidence interval includes zero, the central claim of fairness improvement is not established and the paper should be revised to temper its conclusions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central conclusion asserts that pruning skin-color-associated components 'improve[s] fairness' while reducing computational cost. The empirical support for this claim is thin: in Table II, VGG EOdd improves from 7.76% to 6.10% (1.66 points) and EOpp1 from 6.66% to 5.11%; in Table III, the best ViT configuration (Pattern 6) improves EOdd from 8.22% to 7.11% (1.11 points). These are single-run point estimates with no error bars, no multiple seeds, and no significance testing. Moreover, the authors test six pruning patterns for ViT and report the best one (Pattern 6) as the headline result—a post hoc selection that inflates the apparent effect. Other patterns show instability: Pattern 3 worsens EOpp1 from 7.42% to 9.95%, and Pattern 5 gives only 0.28% EOpp1 improvement. Given that the differences are on the order of 1–2 percentage points and the evaluation uses a single split of ISIC2019, the improvements could easily be within run-to-run noise. If the fairness gains are not statistically reliable, the central claim fails regardless of whether negative skewness truly identifies skin-tone features. The reader's identified assumption (negative skewness → skin-tone encoding) is also unverified, but it is secondary: even if perfectly true, the reported evidence does not demonstrate that pruning those components improves fairness beyond chance.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a skewness-based pruning method ('SkewPrune') to improve fairness in skin lesion classification. For CNNs (VGG11), it computes the skewness of each channel's feature map after ReLU and max pooling, retaining channels with positive median skewness and pruning negatively skewed channels under the assumption that these encode skin-tone information. For ViT-B16, it applies analogous pruning to patch-embedding channels and attention heads, exploring six pruning/fine-tuning patterns. Fairness is measured by EOpp0 (TNR gap), EOpp1 (TPR gap), and EOdd; the authors report that pruning improves fairness on ISIC2019 while preserving accuracy and reducing computational cost. The strongest reported gains are small: VGG EOdd improves from 7.76% to 6.10%; ViT's best pattern improves EOdd from 8.22% to 7.11%.","tokens_in":14697,"tokens_out":4851,"duration_ms":59502,"significance":"If validated, the method would be an attractive fairness intervention because it avoids explicit skin-tone labels, is simple to implement, and simultaneously yields model compression. The paper also provides useful comparisons against SCP-FairPrune and reports FLOPs, parameter counts, and memory footprints. However, the current empirical support is not sufficient. The fairness gains are single-run point estimates with no error bars, no multiple seeds, and no significance testing; for ViT the authors select the best of six pruning patterns post hoc. The central assumption that negative skewness identifies skin-tone-specific components is asserted but never directly verified. These issues are load-bearing for the paper's main claims, although they are addressable with additional experiments. The concept is worth pursuing, but the manuscript in its present form does not convincingly establish the claimed improvements.","major_comments":[{"comment":"The central fairness claim rests on single-run point estimates. No error bars, multiple seeds, or statistical significance tests are reported. The authors test six ViT pruning patterns and present Pattern 6 as the headline (EOdd 8.22→7.11, EOpp1 7.42→6.24), yet Pattern 3 worsens EOpp1 from 7.42% to 9.95%. With a single ISIC2019 train/validation/test split (Section IV-A), differences of 1–2 percentage points could easily be run-to-run noise. Please report per-seed results with confidence intervals, or a paired significance test, and either pre-specify the pruning pattern or correct for multiple comparisons.","section":"Section V-B, Table III"},{"comment":"The method's load-bearing assumption is that a negatively skewed feature/attention distribution indicates a component that encodes skin-tone information, while a positively skewed distribution indicates lesion focus. This is asserted but never verified via feature visualization, attribution, or skin-tone perturbation. Negative skewness could also arise from broad background patterns, illumination, or lesion-adjacent texture. Please add a direct test: for example, show that pruned components respond more to skin-tone changes than to other image perturbations, or include a randomization/magnitude-pruning control that removes the same number of channels without the skewness criterion.","section":"Algorithm 1 and Section III-A"},{"comment":"The claimed computational reduction for ViT is very small: GFLOPs 11.29→11.22 (0.6%), parameters 57.3M→56.97M (0.6%), memory footprint 327.3M→320.19M (2.2%). Describing these as 'significant' reductions (Section V-B) is not supported by the table. Please report relative reductions and temper the RQ3 claim for ViT. The VGG11 reductions (FLOPs 7.61→6.99, params 128.8M→107.91M, memory 491.33M→411.65M) are more substantial and should be distinguished from the ViT case.","section":"Section V-B, Table III"},{"comment":"There is no control that prunes the same number of channels, patches, or heads using a neutral criterion, such as random pruning, magnitude pruning, or pruning based on activation mean. Such a control is necessary to attribute observed fairness improvements to the skewness criterion rather than to generic structural pruning and fine-tuning. The SCP-FairPrune comparisons in Table II are informative but do not isolate this mechanism, especially since SCP-FairPrune uses a different pruning objective and different prune/fine-tune schedule.","section":"Section IV/V"}],"minor_comments":[{"comment":"The metrics EOpp0 and EOpp1 are called 'Equal Opportunity', but EOpp0 is a TNR gap and EOpp1 is a TPR gap. Standard equal opportunity is usually defined via TPR only. Consider renaming them 'TNR gap' and 'TPR gap' to avoid confusion.","section":"Section III-C"},{"comment":"Typo: 'VCG11' should be 'VGG11'. Also 'FLOPS' and 'FLOPs' are used inconsistently; choose one.","section":"Section IV-B"},{"comment":"Typos: 'Unearning' should be 'Unlearning'; in Related Work, 'during interference from inputs' should be 'during inference from inputs'.","section":"Introduction and Related Work"},{"comment":"The table header contains 'Fine-Turning' (should be 'Fine-Tuning'), and the columns 'Prune1'/'Prune2' are not explained in the table caption; the text should define them explicitly.","section":"Table I"},{"comment":"The text says the input dimension is reduced 'from 786 to 336', but ViT-B16's patch embedding dimension is 768, not 786. Also, Algorithm 3's zero-padding procedure is underspecified: when d > |K| it pads, but the positions k_i are not defined, and the case d < |K| is not handled. Please clarify.","section":"Section III-B"}],"recommendation":"major_revision","confidential_remarks":"The idea is interesting and the paper addresses a practically important gap, but the empirical evidence as presented is too weak for acceptance. The authors should be asked to add multiple seeds with confidence intervals, a random/magnitude-pruning control, and a direct validation of the skewness–skin-tone link. The post hoc selection of ViT Pattern 6 also needs to be addressed explicitly. No code is provided, which further limits reproducibility of the pruning pipeline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe short version: this paper has a real idea—using skewness of feature/attention maps to prune skin-tone-related components, and extending that to ViT patches and heads—but the empirical case is not yet convincing. Single runs, no error bars, and a best-of-six selection for the ViT results mean the reported improvements of 1–2 points could easily be noise.\n\nWhat is actually new: the skewness criterion is a clever, low-cost heuristic that doesn't require skin-tone labels. Applying it to attention heads in ViT is a legitimate extension beyond the CNN-focused FairPrune/SCP-FairPrune line. The algorithms are clearly specified, and the comparison against SCP-FairPrune with matched pruning budgets is a nice touch. On VGG, the method does reduce parameters and FLOPs while slightly improving fairness metrics, which is a plausible result.\n\nThe soft spots are in the evaluation. Table III is the biggest problem: six patterns, and the headline result is the one that worked. Pattern 3 makes fairness worse, Pattern 5 barely moves it. There are no confidence intervals, no multiple seeds, no significance tests. Given the effect sizes, the central claim that pruning improves fairness beyond chance is simply not established. The secondary issue is the unvalidated assumption that negative skewness reflects skin-tone encoding; the authors never show that the pruned filters/heads are actually skin-tone-specific. If the assumption is wrong, the method could prune informative features, which might explain the inconsistent ViT results.\n\nIn short, this is a reasonable early-stage idea that is not yet a solid result. I'd send it to peer review—the topic is important and the novelty is genuine—but it needs multi-seed experiments, proper statistical reporting, and at least one validation of the skewness heuristic (e.g., feature visualizations or attribution) before the claims can be trusted. I wouldn't cite it yet.\n\nBest,","headline":"Skewness-based pruning for fairness is a plausible new idea with a useful ViT extension, but the empirical evidence is too thin to establish the central claim.","tokens_in":15173,"tokens_out":2279,"would_cite":false,"duration_ms":26540,"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":"A purely statistical skewness signal can identify and prune skin-tone bias in skin-lesion classifiers, improving fairness and shrinking the model.","keywords":["skin lesion classification","fairness","pruning","skewness","Vision Transformer","VGG","model compression","dermoscopic images"],"falsifier":"Train the same VGG and ViT models, then prune the same number of components chosen at random, or chosen with positive skewness. If fairness improves as much as with negative-skewness pruning, the skewness criterion is not the causal mechanism. Alternatively, use Grad-CAM or a probe classifier to show that the removed channels and heads were not preferentially selective for skin regions over lesion regions; if they were not, the pruning is removing something other than skin-tone encoding.","tokens_in":14283,"feed_emoji":"⚖️","tokens_out":8058,"duration_ms":84878,"temperature":0.7,"pith_summary":"The paper proposes that internal model components whose activations are negatively skewed — spread widely across an image — are the ones encoding skin-tone information in skin-lesion classifiers, and that pruning these components makes the model fairer across skin tones. The authors test this on VGG11 and ViT-B16 using the ISIC2019 dataset, keeping only channels, patches, and attention heads whose median activation skewness is positive (localized, lesion-like). They report improved equalized-odds fairness metrics with little or no accuracy loss, and smaller models with lower FLOPs and memory. Because the criterion is purely statistical, the method needs no skin-tone labels, no group definitions, and no adversarial training.","feed_headline":"Negative skewness marks skin-tone bias; pruning it helps","feed_subtitle":"A no-labels statistical pruning method improves fairness and shrinks the model in one step.","key_machinery":"The load-bearing object is the skewness statistic computed on per-channel feature maps (CNNs) and per-head attention maps (ViTs). Negative skewness is read as widespread moderate activation over skin; positive skewness as localized high activation over a lesion. Algorithm 1 computes the median skewness of each filter/head over the validation set and discards components with non-positive median; Algorithm 2 physically removes the filters and adjusts the classifier input size; Algorithm 3 keeps ViT dimensions consistent by zero-padding pruned residual connections. This statistic carries the entire argument because it identifies 'skin-tone-related' components without attribute labels.","core_discovery":"The central discovery is that the skewness of feature-map and attention-map distributions is a usable, label-free signal for skin-tone bias in lesion classifiers. In dermoscopic images, skin occupies most of the frame and the lesion a small region, so a component that fires broadly (negative skewness) is plausibly tracking skin tone, while one that fires in a small hotspot (positive skewness) is tracking the lesion. The authors prune all components with negative median skewness and fine-tune the model. This improves EOpp1 and EOdd by about 1.5 percentage points on the VGG and 1.1 points in the best ViT configuration, with accuracy roughly unchanged and computation reduced. On the ViT, the be","pith_inferences":["A random-pruning control with the same pruning budget would isolate whether the fairness gains come from removing skin-tone-specific components or from compression-induced regularization.","Applying the skewness criterion to other medical imaging tasks, where background structure differs, would show whether the positive/negative skewness separation is a general bias signal or specific to dermoscopy.","Evaluating with finer-grained skin tone groups rather than a binary light/dark split would show whether the method reduces bias across the full spectrum.","The same criterion could be tested on other vision backbones, such as ResNet or Swin Transformers, to see if the positive-skewness direction always aligns with lesion-focused attention."],"forward_implications":["Fairness can be improved without skin-tone labels, group definitions, or adversarial debiasing, removing a practical barrier for clinical deployment.","The pruning produces smaller and faster models, so fairer classifiers become lighter and cheaper to run, including on edge devices in low-resource settings.","The same skewness criterion applies to both convolutional and transformer architectures, suggesting a common mechanism for bias removal.","On the ViT, freezing the pruned patch-embedding weights during fine-tuning preserves the fairness gains, implying that retraining can undo bias removal if not constrained."],"supporting_citations":[{"why":"Supplies the fairness metrics (EOpp0, EOpp1, EOdd) and the pruning-for-fairness approach that this method extends.","marker":"[31]"},{"why":"Provides the train/validation/test split and the SCP-FairPrune baseline against which the VGG results are compared.","marker":"[33]"},{"why":"Defines the ViT patch notation used to formalize patch pruning in the paper's Equation (1).","marker":"[43]"},{"why":"Provides the self-attention formula used to compute attention maps and to define head pruning.","marker":"[44]"},{"why":"Supplies evidence that many attention heads are redundant and can be pruned, motivating the head-pruning step.","marker":"[42]"},{"why":"The ISIC2019 dataset sources used for all experiments and fairness evaluation.","marker":"[45]–[47]"}],"fun_headline_variants":["Prune negative skewness to cut skin-tone bias in lesion AI","Label-free skewness pruning improves fairness and shrinks model","Skin-tone bias detected via skewness, pruned for fairer diagnosis","Skip labels, use skewness to prune skin-tone bias in classifiers","Prune negatively skewed features to make lesion AI fairer and leaner"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that a negatively skewed activation distribution reliably marks a component as skin-tone-focused rather than as coding any other broad background structure or lesion-adjacent texture — an assumption the paper does not directly verify with attribution or visualization.","fun_headline_variants_meta":{"raw":{"variants":["Prune negative skewness to cut skin-tone bias in lesion AI","Label-free skewness pruning improves fairness and shrinks model","Skin-tone bias detected via skewness, pruned for fairer diagnosis","Skip labels, use skewness to prune skin-tone bias in classifiers","Prune negatively skewed features to make lesion AI fairer and leaner"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00066,"raw_usage":{"total_tokens":2826,"prompt_tokens":688,"completion_tokens":2138,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":432,"completion_tokens_details":{"reasoning_tokens":2058}},"tokens_in":432,"tokens_out":2138,"duration_ms":16213,"temperature":1.0,"reasoning_tokens":2058,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:13:36.937135+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same VGG and ViT models, then prune the same number of components chosen at random, or chosen with positive skewness. If fairness improves as much as with negative-skewness pruning, the skewness criterion is not the causal mechanism. Alternatively, use Grad-CAM or a probe classifier to show that the removed channels and heads were not preferentially selective for skin regions over lesion regions; if they were not, the pruning is removing something other than skin-tone encoding.","supporting_citations":[{"cited_title":"Assessing bias in skin lesion classifiers with contemporary deep learning and post-hoc explainability techniques,","cited_arxiv_id":null,"evidence_quote":"Supplies the fairness metrics (EOpp0, EOpp1, EOdd) and the pruning-for-fairness approach that this method extends."},{"cited_title":"Evaluating and mitigating bias in image classifiers: A causal perspective using coun- terfactuals,","cited_arxiv_id":null,"evidence_quote":"Provides the train/validation/test split and the SCP-FairPrune baseline against which the VGG results are compared."},{"cited_title":"Achieving fairness through channel pruning for dermatological disease diagnosis,","cited_arxiv_id":null,"evidence_quote":"Defines the ViT patch notation used to formalize patch pruning in the paper's Equation (1)."},{"cited_title":"Fairquantize: Achieving fairness through weight quantization for dermatological disease diagnosis,","cited_arxiv_id":null,"evidence_quote":"Provides the self-attention formula used to compute attention maps and to define head pruning."},{"cited_title":"Toward fairness through fair multi-exit framework for dermatological disease diagnosis,","cited_arxiv_id":null,"evidence_quote":"Supplies evidence that many attention heads are redundant and can be pruned, motivating the head-pruning step."}],"review_version":1}