{"id":"446bd44f-4dc9-4f16-81dd-67ab5042c705","arxiv_id":"2505.20512","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A feature-space association test using external probe datasets estimates demographic bias in facial expression recognition models without test-time demographic labels.","lead":"This paper proposes a feature-level method for measuring demographic bias in facial expression recognition models that does not require demographic labels on the test images. It validates the method against human-annotated labels on AffectNet and shows it aligns better with ground truth than existing pseudo-label approaches.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DiA in Eq. (4)-(5) marginalizes over the test set's demographic mix per expression, so test-set imbalance can drive DiA even for a fair model; the validation may be an artifact of this composition confound.","rationale":"The reader's weakest assumption was that DiA reflects TPR disparity without theoretical derivation, with probe expression confounds as an additional worry. My concern is more specific and more load-bearing: DiA as defined in Eq. (4)--(5) is a marginal statistic over the test set, not a conditional one. Even a perfectly fair model, or an untrained encoder, will produce nonzero DiA for expressions whose test images are demographically imbalanced, because the average similarity to a probe group tracks the prevalence of that group in Z_e. The validation against DEO, which is computed per demographic group, can therefore be driven by a common confound: both DiA and DEO depend on the same demographic imbalances in AffectNet, but only DEO actually measures the model's per-group classification behavior. The paper does not test this confound, for example by rebalancing Z_e or by using a model with intentionally removed bias. Because the central claim is that the feature-level framework evaluates model bias more effectively than pseudo-label methods, this unaddressed confound makes the current empirical support inconclusive. However, the issue is testable and the paper's overall structure is otherwise sound; conditional acceptance pending the balanced-subset test is appropriate, so the reader's verdict should remain unchanged.","tokens_in":25026,"tokens_out":9595,"duration_ms":102754,"concrete_test":"Use the human-annotated demographic labels on the AffectNet test set (Hu et al., used as ground truth) to construct, for each expression e and each sensitive attribute, a balanced subset Z_e^bal in which every demographic group has the same number of test images (subsample to the smallest group). Recompute DiA with Z_e^bal and recompute the L1 distances in Tables II, IV, and VI. If the 'ours' methods no longer show consistently lower L1 than the pseudo-label baselines, the original result is an artifact of test-set composition. As a second arm, repeat the computation with an untrained (random-initialized) Swin-B: if DiA remains large and significant, the statistic is not measuring learned model bias.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (4) defines A(e,s_j) as the average cosine similarity between every test embedding z_e in Z_e (all AffectNet test images labeled expression e) and every probe embedding z_sj. Equation (5) then subtracts these averages across demographic groups. Because the sum over z_e runs over the entire expression-specific test set, A(e,s_j) is a marginal statistic: it is inflated whenever expression e is disproportionately represented by group s_j in the test set, regardless of how the FER model treats that group in classification. The ground truth it is validated against, DEO (Eq. (6)), is conditional: TPR is computed separately within each demographic group and then differenced. The paper's central claim is that DiA reflects this classification-level bias, but no derivation supports that bridge, and the empirical agreement in Tables II/IV/VI could be produced by the shared dependence of both DiA and DEO on AffectNet's demographic imbalances within expressions (the same imbalance illustrated for RAF-DB in Fig. 1). A random-initialized encoder that encodes gender/age/race would still yield nonzero DiA on imbalanced Z_e, so the statistic does not isolate model behavior. This is the weakest point in the argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a feature-level bias evaluation framework for FER models when demographic labels are unavailable in the test set. Instead of relying on pseudo-demographic labels, the method uses a probe dataset (UTKFace or FairFace) and computes differential associations (DiA) between expression-specific test embeddings and demographic probe embeddings in the feature space of the frozen FER model (Eqs. 4-5). A permutation-based statistical module is introduced to test the significance of both DiA and the conventional performance disparity (DEO, Eq. 6). The framework is validated on the AffectNet test set using human-annotated demographic labels from Hu et al. as ground truth, comparing L1 distances against pseudo-label baselines across gender, race, and age (Tables II, IV, VI), and is further applied to eight network architectures and a sensitivity analysis of the significance threshold.","tokens_in":25206,"tokens_out":5730,"duration_ms":62233,"significance":"If the central claim holds, the framework would fill a practical gap: evaluating demographic bias in FER models without demographic test labels, while avoiding the distortions of pseudo-label classifiers. The paper also contributes a plug-and-play statistical module and a large-scale bias analysis. Strengths include released code, a reproducibility-focused supplementary with per-sample exclusion lists, validation against human-annotated ground truth, and experiments across multiple architectures and probe datasets. However, the central statistic is marginal over the test set and may be confounded by test-set composition and probe expression content; the validation evidence is therefore not yet conclusive.","major_comments":[{"comment":"The differential association DiA is a marginal statistic: A(e,s_j) averages cosine similarity over all test embeddings Z_e and all probe embeddings Z_{s_j}. On a test set in which expression e is demographically imbalanced, any model whose features encode demographic attributes—including a classifier that is perfectly fair in the TPR sense—will produce a nonzero DiA. The ground truth DEO in Eq. (6) is conditional (TPR is computed within each demographic group), so the agreement in Tables II, IV, and VI may be driven by the shared dependence of both quantities on the same composition imbalance rather than by DiA capturing classification bias. No derivation or control experiment (e.g., a random-initialized encoder or a balanced subset of the test set) rules out this confound; this is the load-bearing weakness of the validation.","section":"IV-A, Eqs. (4)-(5)"},{"comment":"The reference group s^e_max is selected from the test data before permutation testing. Under the null hypothesis, the maximum over groups is a data-dependent statistic, so the one-sided p-values for DiP and DiA computed against that selected maximum are anti-conservative; the selection step is not part of the null distribution. No multiple-comparison correction is applied across the n-1 remaining groups, seven expressions, three attributes, or multiple architectures, so the claim that the statistical module 'ensures statistical significance' (abstract, Section IV-B) is overstated.","section":"IV-B, Eqs. (7)-(9)"},{"comment":"The probe datasets UTKFace and FairFace contain natural facial expressions, but A(e,s_j) averages over all embeddings in the probe group regardless of expression content. Expression-related differences between demographic probe subsets (e.g., female faces smiling more often, infant faces showing particular expressions) can inflate the association for the corresponding target expression independently of FER model bias. The paper does not report expression distributions in the probe groups or any control for this confound; this threat is especially acute for the age results in Table V, where very young faces may systematically display fear or surprise.","section":"V-A and Eq. (4)"},{"comment":"The L1 evaluation metric is computed on statistically zeroed values eV from Eq. (10). Because both the candidate method and the ground truth are zeroed by the same permutation module, the L1 comparison partly measures agreement in significance decisions rather than the magnitude of bias. The paper should also report distances on raw DEO/DiA values and include confidence intervals or effect sizes to ensure that the proposed methods are not favored merely by having larger p-values.","section":"V-D, Eq. (11)"}],"minor_comments":[{"comment":"The indicator function is written as 'l(·)'; this appears to be a typographical remnant (likely 1{·}) and should be corrected.","section":"IV-B, Eq. (9)"},{"comment":"The curves for UTKFace (ours) and FairFace (ours) are nearly indistinguishable from the ground truth in the figure; adding distinct markers or a zoomed inset would make the sensitivity comparison readable.","section":"Figure 4"},{"comment":"The pilot study demonstrates qualitative clustering via t-SNE; adding a quantitative measure (e.g., linear-probe accuracy or silhouette score per attribute) would make the claim 'FER models are good attribute encoders' more rigorous.","section":"V-C"},{"comment":"The symbol '↓' in the bias-direction rows is not defined in the captions; please define it at first use, for example in a footnote to Table V.","section":"Tables V and VI"},{"comment":"The phrase 'without demographic labels in the test set' is not strictly accurate for the validation setting, which uses the human-annotated test set from Hu et al.; consider phrasing 'for deployment settings where such labels are unavailable' to avoid confusion.","section":"Abstract and IV-A"}],"recommendation":"major_revision","confidential_remarks":"The central concern is the composition confound in DiA, which is load-bearing for the claimed validation. If the authors can provide a controlled experiment (e.g., a balanced test-set subset or a random-encoder baseline) and address the statistical testing issues, the paper could be publishable; otherwise the empirical support for the central claim remains inconclusive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful engineering contribution with a real statistical blind spot. The idea—evaluate FER bias via differential associations in feature space rather than pseudo-labels—is worth taking seriously. The reformulation of iEAT to multi-class expressions and multiple demographic groups is new, and the plug-and-play permutation module is a genuinely useful addition to a literature that too often reports raw disparity numbers without significance tests. The validation against human-annotated demographic labels on AffectNet is the right gold standard, and the breadth—seven expressions, three attributes, eight architectures—is impressive.\n\nBut the central statistic has a problem. DiA in Eq. (4)-(5) is a marginal average: A(e, s_j) averages cosine similarity over every test image labeled e against every probe image of group s_j. If the test set contains more male faces labeled 'anger' than female, those male faces pull the mean embedding of anger toward the male probe centroid. A completely fair model would still produce a nonzero DiA on an imbalanced test set. The ground truth DEO is conditional—TPR computed within each group—so the agreement in Tables II, IV, VI could be an artifact of the shared dependence of both quantities on the dataset's demographic mix per expression. The paper offers no derivation connecting the two, and the empirical validation alone can't separate the composition effect from genuine model behavior. This is load-bearing, not a minor caveat.\n\nOther issues: the reference group for the permutation test is selected from the same data before permuting, and with many comparisons (7 expressions × 3 attributes × n groups) there's no multiple-comparison correction. The probe datasets may contain expression content that overlaps with the target expression, which would inflate association for groups that happen to smile or scowl more in the probe set. The subjective data cleaning is disclosed and small, so I only flag it lightly. No repeated training runs for error bars is a minor weakness.\n\nWho should read this: people building bias evaluation tools for FER and affective computing. The statistical module is worth borrowing. The benchmark numbers across architectures are useful even if the interpretation is contested.\n\nRecommendation: send to peer review, but with a major-revision expectation. The authors need to address the composition confound—for example, by reweighting the expression-specific test embeddings to a uniform demographic distribution, or by including a randomly initialized encoder as a negative control, which should yield zero DiA if the test is isolating model behavior. Without that, the central claim doesn't hold.","headline":"Clever adaptation of iEAT to FER, but the test's marginal averaging over the test set means it can flag dataset composition as model bias; the central validation is shakier than the low L1 distances suggest.","tokens_in":25794,"tokens_out":3829,"would_cite":false,"duration_ms":42279,"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":"Demographic bias in facial expression recognition can be measured directly in a model's feature space, without test-set demographic labels, and this measurement tracks human-annotated ground truth more closely than pseudo-label pipelines.","keywords":["facial expression recognition","demographic bias evaluation","algorithmic fairness","feature-space bias","differential association","permutation test","pseudo-demographic labels","AffectNet"],"falsifier":"Construct or find a test set where expression content and demographic appearance are deliberately anti-correlated (for instance, anger shown mostly on female faces and happiness mostly on male faces) with reliable human demographic labels; if the differential-association ranking then disagrees with the TPR-disparity ranking, the bridging assumption fails. A lighter check: recompute the differential associations after cropping probe images to the face region or filtering probe sets to neutral expressions only, and see whether the L1 distances to ground truth degrade; if they shift sharply, the probe-content confound is doing real work.","tokens_in":24757,"feed_emoji":"⚖️","tokens_out":11246,"duration_ms":92102,"temperature":0.7,"pith_summary":"The paper tackles a practical obstacle in fairness auditing of facial expression recognition (FER) models: public FER datasets rarely carry demographic labels, so previous bias studies either restricted their scope or invented pseudo-labels by training attribute classifiers, which can distort the very bias they aim to measure. The authors propose to read bias instead from the feature space of the frozen FER model, comparing how strongly embeddings of each expression associate with face images from a separate probe dataset labelled by perceived gender, age, and race. Against human-annotated demographic labels on the AffectNet test set, they claim this feature-level differential association reproduces true classification-level bias more closely than pseudo-label pipelines, especially for age and race, and it never misidentifies the reference group where pseudo-label methods fail. They also argue that many reported bias numbers may be statistical noise, and supply a permutation-test module that zeros out any disparity not significant at $\\alpha = 0.05$. If right, the paper gives auditors a practical way to check FER models for demographic bias using only a face-attribute dataset and the model's own embeddings.","feed_headline":"Emotion-AI bias read straight from the model's features","feed_subtitle":"A permutation-tested feature-space score matches human-annotated bias on AffectNet better than pseudo-label methods.","key_machinery":"The load-bearing object is the differential association $\\mathrm{DiA}^e_{(j,j')} = A(e, s_j) - A(e, s_{j'})$, where $A(e, s_j)$ is the average of $\\cos(\\cdot)+1$ between embeddings of test images showing expression $e$ and embeddings of probe images labelled with demographic group $s_j$. This is a reformulation of the Image Embedding Association Test (iEAT), an association test originally built for binary target–attribute pairs in self-supervised models, reworked here for multi-class classification with multi-group sensitive attributes. The machinery does its work under a bridging assumption: differential association in the feature space is taken to mirror the true positive rate disparity $\\mathrm{DEO}^e_{(\\max,k)}$ that the model actually exhibits between demographic groups. A plug-and-play statistical module—a permutation test with $B = 10000$ shuffles and one-sided $p$-values—decides which observed values survive; anything failing the $\\alpha = 0.05$ threshold is set to zero, and the reference group is always the one with the highest performance or strongest association.","core_discovery":"The central claim is that demographic bias in an FER model can be evaluated without demographic labels on the test set by computing differential associations in the model's own feature space. For each expression $e$ and each demographic group $s_j$, the mean cosine similarity (shifted by $+1$) between embeddings of test images showing $e$ and embeddings of probe images labelled $s_j$ defines an association $A(e, s_j)$; the difference between two groups, $\\mathrm{DiA}^e_{(j,j')}$, is the feature-level bias estimate. Tables I–VI show that, across seven expressions and three attributes, both UTKFace and FairFace probe versions land closer in average L1 distance to the human-annotated ground-truth bias (true positive rate disparities on AffectNet) than do two pseudo-demographic-label pipelines, and the feature-level method identifies the correct reference group in every case where the pseudo-label methods return NaN. The claim extends across architectures: run on eight network backbones, the feature-level method tracks the ground-truth average bias for both CNNs and Transformers. Alongside this, the permutation-test statistical module is the paper's second contribution: with 10,000 label shuffles and a one-sided threshold $\\alpha = 0.05$, each disparity or association is either retained as statistically significant or set to zero, filtering out noise before any bias claim is made.","pith_inferences":["The paper does not control for emotional content inside the probe images: if a demographic group's probe set happens to display the expression under evaluation, the cosine-similarity association could be inflated, so a natural extension is to restrict probe sets to neutral expressions and re-check whether the L1 rankings survive.","Because the audit operates on embeddings, it could transfer to settings with even scarcer demographic labels, such as fine-tuned downstream tasks or multimodal emotion-recognition systems; the paper names this as future work rather than a demonstrated result.","The ground-truth comparison rests on a single annotation source, so an independent human-annotated FER test set would be the sharpest check on whether the feature-level advantage generalises across annotation protocols.","The permutation test is applied expression by expression without multiple-comparison correction, so a stricter analysis with a family-wise correction across the seven expressions would likely shrink the set of biases reported as significant."],"forward_implications":["FER models can be audited for demographic bias without collecting demographic labels on the test set, and a modest probe dataset (UTKFace, about 24,000 images) is sufficient for the audit.","Pseudo-demographic-label pipelines misidentify the best-performing demographic group for age and race, producing NaN results in the comparison tables, while the feature-level method identifies the correct reference group across all seven expressions.","Bias evaluations that skip statistical testing can report noise as bias; the permutation module makes significance checking plug-and-play for both the new framework and existing label-based pipelines.","On AffectNet, age and race biases are markedly larger than gender biases, transformer backbones show higher average bias than CNNs, and within each architecture family bias declines as model size grows."],"supporting_citations":[{"why":"Supplies the Image Embedding Association Test that the paper reformulates for multi-class FER; the differential-association machinery is adapted from here.","marker":"[47]"},{"why":"Provides the human-annotated demographic labels on the AffectNet test set used as ground truth, plus the data split; every method's L1 distance is measured against this.","marker":"[42]"},{"why":"UTKFace, one of the two probe datasets; its demographic labels anchor the feature-space association computation in Equation (4).","marker":"[64]"},{"why":"FairFace, the larger balanced probe dataset; used to test how sensitive the framework is to probe scale and composition.","marker":"[65]"},{"why":"AffectNet, the large-scale FER dataset on which the eight models are trained and the bias evaluation is run.","marker":"[41]"},{"why":"The pseudo-demographic-label evaluation pipelines that serve as the comparison baselines the paper claims to outperform.","marker":"[35], [36], [39], [43]"},{"why":"Earlier feature-space bias studies applying iEAT to vision models; they motivate the position that bias lives in embeddings rather than the final classification layer.","marker":"[45], [46]"}],"fun_headline_variants":["No labels needed: FER bias from features alone","Feature-space test reveals emotion AI bias","Bias without demographic labels: feature-level check","Statistical filter for unbiased AI emotion checks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the bridging assumption that the feature-space differential association of Equations (4)–(5) tracks the classification-level bias of Equation (6) — a link the paper validates empirically but never derives — and on the probe datasets representing perceived demographic groups without their own expression content contaminating the similarity comparison.","fun_headline_variants_meta":{"raw":{"variants":["No labels needed: FER bias from features alone","Feature-space test reveals emotion AI bias","Bias without demographic labels: feature-level check","Statistical filter for unbiased AI emotion checks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000602,"raw_usage":{"total_tokens":2858,"prompt_tokens":1043,"completion_tokens":1815,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":659,"completion_tokens_details":{"reasoning_tokens":1759}},"tokens_in":659,"tokens_out":1815,"duration_ms":13685,"temperature":1.0,"reasoning_tokens":1759,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:52:15.304623+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct or find a test set where expression content and demographic appearance are deliberately anti-correlated (for instance, anger shown mostly on female faces and happiness mostly on male faces) with reliable human demographic labels; if the differential-association ranking then disagrees with the TPR-disparity ranking, the bridging assumption fails. A lighter check: recompute the differential associations after cropping probe images to the face region or filtering probe sets to neutral expressions only, and see whether the L1 distances to ground truth degrade; if they shift sharply, the probe-content confound is doing real work.","supporting_citations":[{"cited_title":"Image representations learned with unsu- pervised pre-training contain human-like biases,","cited_arxiv_id":null,"evidence_quote":"Supplies the Image Embedding Association Test that the paper reformulates for multi-class FER; the differential-association machinery is adapted from here."},{"cited_title":"Rethinking affect analysis: A protocol for ensuring fairness and consis- tency,","cited_arxiv_id":null,"evidence_quote":"Provides the human-annotated demographic labels on the AffectNet test set used as ground truth, plus the data split; every method's L1 distance is measured against this."},{"cited_title":"Age progression/regression by condi- tional adversarial autoencoder,","cited_arxiv_id":null,"evidence_quote":"UTKFace, one of the two probe datasets; its demographic labels anchor the feature-space association computation in Equation (4)."},{"cited_title":"Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation,","cited_arxiv_id":null,"evidence_quote":"FairFace, the larger balanced probe dataset; used to test how sensitive the framework is to probe scale and composition."}],"review_version":1}