{"id":"23ce4939-1da8-486d-8025-8fa2b4709818","arxiv_id":"2501.11752","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A VAE trained on dermatology images reconstructs lighter skin better than darker skin even under balanced training, and its uncertainty estimates fail to warn about this disparity.","lead":"This paper trains an image-reconstructing AI on dermatology photos and checks how well it handles light versus dark skin. It finds that darker skin is reconstructed worse even with balanced training data, and the model's uncertainty scores do not flag this disparity.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Condition mix confounds the 'independent of representation' claim: light and dark test sets differ in fine-grained diagnoses, so balanced FST sampling does not isolate skin tone.","rationale":"The reader's weakest assumption already identifies the confound between skin tone and disease difficulty, and reading §5.2 confirms it is the central weak point of the paper. The paper honestly reports the fine-grained diagnosis imbalance in Figure 7c and explicitly asks whether skin tone and condition can be disentangled, so this is not a hidden flaw; it is an acknowledged limitation that caps the strength of the abstract claim. The VAE uncertainty finding is less load-bearing for the headline claim because it is a negative result about a monitoring tool and does not depend on the causal attribution of the performance gap. I therefore agree with the reader that CONDITIONAL remains the right verdict, conditional on a condition-matched or statistically adjusted analysis for the 'independently of representation' claim, or on a corresponding weakening of the claim. The proposed condition-matched recomputation would settle whether the residual gap is attributable to skin tone or to condition mix. The paper does provide ten repeated runs and points to a public implementation, which is useful, but neither resolves the confound.","tokens_in":9061,"tokens_out":3017,"duration_ms":32881,"concrete_test":"Restrict both test sets to fine-grained Fitzpatrick17k condition labels that have at least N ≥ 20 images in both FST1-2 and FST5-6, and recompute Figure 3's likelihood/MSE under the same Dataset B protocol; if no significant light-dark gap remains in the matched subset, the 'independent of representation' claim should be withdrawn or weakened. If fine-grained subsets are too small, use the 9-partition middle labels with a mixed-effects model: condition as random effect, skin tone as fixed effect, and test whether tone still predicts MSE.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline finding is that 'even independently of representation, the VAE performs better for lighter skin tones' (abstract) and that 'lighter skin tones appear to be easier to reconstruct' (§5.1). The support is Figure 3, especially Dataset B with a 50/50 skin-tone training split, where dark MSE remains above light MSE. But 'representation' is balanced only for skin-tone labels, not for disease labels. Section 5.2 and Figure 7c show that fine-grained condition prevalence differs substantially between the FST1-2 and FST5-6 groups, and the authors explicitly leave open whether skin tone and specific skin condition can be disentangled. If darker-skinned test images contain more difficult-to-reconstruct conditions (e.g., diffuse textured rashes versus clearly delineated lesions, as Figures 4 and 5 suggest), the Dataset B gap can be driven by condition difficulty rather than by skin tone per se. The claim 'independent of representation' is therefore not established by the current experiment; the honest limitation in §5.2 makes the conclusion conditional rather than definitive.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies racial bias in a VAE trained on the Fitzpatrick17k dermatology dataset. It varies the proportion of light (FST 1-2) and dark (FST 5-6) skin tones in training sets across three configurations (light-only, 50/50 mixed, dark-only), and measures reconstruction MSE and latent posterior standard deviation on held-out light and dark test sets. The authors report that reconstruction performance tracks representation, that a performance gap favoring light skin persists even under a 50/50 skin-tone split, and that latent standard deviation does not reflect this gap. They conclude that generative models are generally biased against darker skin tones and that uncertainty estimates are not useful fairness indicators.","tokens_in":9313,"tokens_out":6633,"duration_ms":62933,"significance":"If the residual gap under balanced skin-tone representation is confirmed, the paper makes a valuable contribution: it extends fairness analysis from discriminative to generative models, provides a reproducible protocol based on repeated independent runs, and raises a practical warning about uncertainty-based monitoring. The authors are also transparent about the condition-confounding limitation in §5.2. However, because the main abstract claim depends on ruling out condition-mix differences, the current evidence is conditional rather than definitive.","major_comments":[{"comment":"The claim \"even independently of representation, the VAE performs better for lighter skin tones\" (abstract) is not established because the light and dark test sets differ in fine-grained condition distribution. The authors' own Figure 7c shows substantial differences at the fine-grained level, and Figure 8 shows that condition-level MSE is uneven; §5.2 explicitly leaves open the disentanglement of skin tone and condition. The Dataset B comparison therefore confounds skin tone with condition difficulty. To substantiate the claim, please report the gap within condition-matched subgroups or after adjusting for condition prevalence, or restrict the conclusion to \"with balanced skin-tone sampling.\"","section":"§5.2, Figures 7-8"},{"comment":"The phrase \"independent of representation\" overstates the experimental control: only the coarse skin-tone label is balanced, not the joint distribution of skin tone and diagnosis. Since the training sets for Dataset B are 50/50 in skin tone but not in disease mix, the model may have encountered different diseases for each group; this leaves open an alternative explanation for the residual gap. Please either redesign the sampling to balance fine-grained conditions or soften the language throughout.","section":"§3.4, abstract, §5.1"}],"minor_comments":[{"comment":"The sentence \"one with darker skin, comprising FST 4-6\" contradicts the test-set definition in the same paragraph (FST 5-6) and the statement that \"FST 3-4\" images are discarded; please clarify which Fitzpatrick types are actually used.","section":"§3.4"},{"comment":"The subcaptions \"(a) Coarse level... (b) Medium level... (c) Fine-grained level...\" do not match the main caption \"Averaged MSE of skin conditions based on fine-grained labels\" or the actual subplot labels (Dataset A/B/C); please correct the mismatch.","section":"Figure 8"},{"comment":"In the description of the encoder, \"qϕ(x|z)\" should be \"qϕ(z|x)\" in the sentence \"combining an stochastic encoder qϕ(x|z).\"","section":"§3.1"},{"comment":"There is a typo \"widht\" for \"width,\" and the normalization factor would be clearer written as 1/(C^l W^l H^l).","section":"Eq. (2)"},{"comment":"The paper evaluates reconstruction, not generation of new samples; the title and abstract's \"image generation\" should be aligned with the actual experiments or note that generation is not evaluated.","section":"Title and abstract"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest about the confound in §5.2, but the abstract and §5.1 overstate the strength of the evidence. A condition-stratified analysis would substantially strengthen the paper and could turn the major concern into a minor one."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. The paper is a rare controlled study of skin-tone bias in a generative model for dermatology, and the repeated-sampling experiment is done carefully. But the headline claim that the VAE performs better for lighter skin 'even independently of representation' is not supported by the data as presented, because light and dark test sets differ systematically in diagnosis mix. The authors know this and say so in Section 5.2, yet the abstract still overstates it.\n\nWhat is genuinely new: extending Larrazabal et al.'s representation-control setup from classifiers to a perceptual-loss VAE, and the negative result on latent standard deviation as a fairness monitor. That negative result may be the most robust finding. The uncertainty estimates do not track subgroup performance or representation, so anyone planning to use VAE uncertainty as a red-flag mechanism should read this. The paper also gives honest qualitative examples and acknowledges the skin-tone-versus-condition disentanglement problem explicitly.\n\nWhere it is soft, in rough proportion: the 'independent of representation' claim is the central load-bearing assertion, and Section 5.2's own Figure 7c shows fine-grained condition prevalence differs substantially between FST 1-2 and FST 5-6. Figures 4 and 5 are consistent with darker examples being more diffuse and textural. A matched-condition or condition-interaction analysis is needed before claiming skin tone itself drives the gap; otherwise the residual could just be image difficulty. That is a real confound, not a minor caveat. Second, the perceptual loss uses an ImageNet-pretrained VGG19, which may itself have skin-tone-dependent features; the loss function is not neutral. Third, there are no statistical tests on the boxplots, and ten runs is modest power; code would help. The exclusion of FST 3-4 is defensible but should be remembered when generalizing.\n\nWho should read this: people working on fairness in medical generative models, dataset construction, and uncertainty-based bias monitoring. It deserves a serious referee. With the central claim qualified and a bit more evidence or released code, it would be a useful contribution.","headline":"A solid controlled study of skin-tone bias in a dermatology VAE, but the 'independent of representation' claim is confounded by condition mix and needs softening.","tokens_in":9750,"tokens_out":2222,"would_cite":true,"duration_ms":21652,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A VAE trained on clinical dermatology images reconstructs lighter skin tones better than darker ones even when training data is balanced 50/50, and its built-in uncertainty estimates do not flag the gap.","keywords":["racial bias","dermatology","generative models","variational autoencoder","Fitzpatrick skin type","uncertainty quantification","representation bias","fairness"],"falsifier":"Train the same VAE on a balanced 50/50 dataset, then test on light and dark sets that are matched exactly on fine-grained skin-condition labels (for example, only common conditions with nearly equal counts); if the light-minus-dark MSE gap collapses to zero, the representation-independent bias claim fails, while a gap on matched sets would confirm the bias as skin-tone-related.","tokens_in":8889,"feed_emoji":"⚖️","tokens_out":6383,"duration_ms":56288,"temperature":0.7,"pith_summary":"The paper asks whether a deep generative model trained on clinical dermatology images is fair across skin tones, and answers with a controlled VAE study on the Fitzpatrick17k dataset. It establishes that reconstruction quality tracks subgroup representation: more dark-skin images in training lowers dark-skin reconstruction error. It then reports a second, representation-independent effect: even at a 50/50 training mix the VAE reconstructs lighter skin (FST 1-2) better than darker skin (FST 5-6). It also finds that the VAE's averaged latent standard deviation, its built-in uncertainty estimate, does not distinguish the two subgroups and so cannot serve as a warning sign for this bias. The importance is that balancing datasets is necessary but apparently not sufficient for fair generative models, and uncertainty quantification needs rethinking.","feed_headline":"Balanced skin tones still favor lighter skin in a VAE","feed_subtitle":"Dark-skin reconstructions stay worse even with balanced training, and the model's uncertainty measure never signals the gap.","key_machinery":"The central object is a VAE with a Gaussian latent ($q_\\phi(z|x) = \\mathcal{N}(\\mu_\\phi(x), \\mathrm{diag}(\\sigma_\\phi^2(x)))$), a Gaussian decoder, and a perceptual loss based on VGG19 feature maps. The argument is carried by comparing two quantities across three training-set configurations: the reconstruction likelihood (MSE) and the averaged latent standard deviation. The three configurations -- Dataset A (100% light), Dataset B (50/50), Dataset C (100% dark) -- instantiate the representation axis, while separate light and dark test sets isolate subgroup performance.","core_discovery":"The authors train a VAE with a VGG19 perceptual loss on Fitzpatrick17k images under three training configurations -- all light (FST 1-2), 50/50 mixed, and all dark (FST 5-6) -- and test on fixed light and dark test sets. Their central discovery is that darker skin is reconstructed with systematically higher MSE (lower likelihood) than lighter skin across all configurations, including the balanced one, so the performance gap is not explained by representation alone. A second discovery is that the averaged latent standard deviation of the stochastic encoder looks nearly identical for light and dark test sets even where reconstruction performance differs, making the VAE's natural uncertainty estimate useless as a fairness diagnostic. The paper leaves open whether the residual gap is caused by skin tone itself or by differences in the fine-grained skin conditions present in the two groups.","pith_inferences":["If the fine-grained condition mismatch is the true driver, then a matched test set that holds skin-condition distributions equal across skin tones should erase or shrink the residual MSE gap; this is a testable extension the paper does not run.","The failure of latent-variance UQ suggests exploring decoder-side heteroscedastic noise or predictive intervals on reconstructed pixels, which may better reflect subgroup difficulty.","Because the Fitzpatrick scale itself conflates skin tone with other attributes, the 'darker skin' group may be a proxy for correlated image properties such as texture, lighting, and lesion morphology, implying the bias could persist even with perfect skin-tone balance."],"forward_implications":["A balanced 50/50 training set does not equalize VAE reconstruction quality between lighter and darker skin tones.","Latent standard deviation cannot be relied on as a fairness alarm when deploying such a VAE in clinical settings.","The common real-world scenario of light-dominated training data produces the largest dark-skin performance gap, so skin-tone bias should be expected in generative models trained on existing dermatology collections.","Improving dataset diversity may be necessary but is not sufficient; the source of the representation-independent gap needs separate investigation."],"supporting_citations":[{"why":"Supplies the Fitzpatrick17k dataset with Fitzpatrick skin type labels and skin-condition labels that the study rebalances into light and dark training configurations.","marker":"[8]"},{"why":"Provides the experimental design of varying subgroup representation in training to expose how subgroup performance depends on subgroup representation.","marker":"[14]"},{"why":"Defines the VAE formulation with Gaussian encoder, Gaussian decoder, and ELBO objective that the paper adapts with a perceptual loss.","marker":"[12,20]"},{"why":"Introduces the deep feature perceptual loss that gives the VAE high-fidelity reconstructions, which are then evaluated with MSE.","marker":"[9]"},{"why":"Documents previous evidence of skin-tone disparities in dermatology AI, motivating the study of racial bias and providing the context for the observed gap.","marker":"[1,3]"},{"why":"Represents the alternative use of large diffusion models to generate synthetic samples and balance datasets, serving as the contrast point for studying bias inside a generative model.","marker":"[21]"},{"why":"Supplies the notion of hidden stratification in medical imaging that the discussion uses to explain why fine-grained condition differences may complicate the skin-tone attribution.","marker":"[16,17]"},{"why":"Supports the discussion's caution that Fitzpatrick skin type labels are contested and may not measure the right features for racial bias.","marker":"[13,22]"}],"fun_headline_variants":["VAE still favors lighter skin even with balanced training data","Darker skin reconstructions stay worse in balanced VAE setup","Uncertainty estimates fail to flag racial bias in dermatology VAE","Balanced skin tones don't eliminate VAE's racial performance gap","Generative model bias: darker skin worse even when data is balanced"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's key assumption is that the light and dark test groups are otherwise comparable, but the fine-grained skin-condition distributions differ substantially between them, so the residual performance gap could be driven by condition difficulty rather than skin tone itself.","fun_headline_variants_meta":{"raw":{"variants":["VAE still favors lighter skin even with balanced training data","Darker skin reconstructions stay worse in balanced VAE setup","Uncertainty estimates fail to flag racial bias in dermatology VAE","Balanced skin tones don't eliminate VAE's racial performance gap","Generative model bias: darker skin worse even when data is balanced"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000226,"raw_usage":{"total_tokens":1491,"prompt_tokens":994,"completion_tokens":497,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":408}},"tokens_in":610,"tokens_out":497,"duration_ms":5402,"temperature":1.0,"reasoning_tokens":408,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:53:56.962069+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same VAE on a balanced 50/50 dataset, then test on light and dark sets that are matched exactly on fine-grained skin-condition labels (for example, only common conditions with nearly equal counts); if the light-minus-dark MSE gap collapses to zero, the representation-independent bias claim fails, while a gap on matched sets would confirm the bias as skin-tone-related.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the Fitzpatrick17k dataset with Fitzpatrick skin type labels and skin-condition labels that the study rebalances into light and dark training configurations."},{"cited_title":"In: 2017 IEEE winter conference on applications of computer vision (WACV)","cited_arxiv_id":null,"evidence_quote":"Introduces the deep feature perceptual loss that gives the VAE high-fidelity reconstructions, which are then evaluated with MSE."},{"cited_title":"In: NeurIPS 2022 Workshop on Synthetic Data for Empowering ML Research (2022)","cited_arxiv_id":null,"evidence_quote":"Represents the alternative use of large diffusion models to generate synthetic samples and balance datasets, serving as the contrast point for studying bias inside a generative model."}],"review_version":1}