Pith. sign in

REVIEW 3 major objections 4 minor 66 references

Bridging the gap in FER: addressing age bias in deep learning

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims age-related bias in facial expression recognition can be substantially reduced by simple training changes, with Age-weighted Loss raising elderly macro F1 on FACES from 0.7528 to 0.8140 and cutting the adult–elderly gap…

desk verdict Worth refereeing, but the central mitigation claim rests on 114 elderly test images and no significance tests; treat the elderly gains as suggestive, not established. read the letter →

arxiv 2507.07638 v1 pith:NI2GIR2C submitted 2025-07-10 cs.CV

classification cs.CV
keywords facialexpressionrecognitionagebiaselderlymitigationage-weightedlossmulti-tasklearningexplainableAIAffectNet
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that age-related bias in facial expression recognition, especially the poor recognition of elderly faces, can be substantially reduced without redesigning the neural network. Using AffectNet with automatically estimated age labels, the authors identify which expressions suffer most for elderly people (neutral, sadness, anger) and test three simple fixes: reweighting the training loss by age group, adding age prediction as a secondary task, and feeding age to the classifier. The strongest results come from Age-weighted Loss, which raises the elderly macro F1 on the FACES benchmark from 0.7528 to 0.8140 and cuts the adult–elderly performance gap from 0.0917 to 0.0380. The broader point is that even noisy, automatically estimated demographic labels are enough to steer a model toward fairer behavior across age groups.

What carries the argument

The load-bearing mechanism is the Age-weighted Loss, a per-sample weight $w_{s,a} = N_{\min}/N_{s,a}$ that rebalances the joint distribution of expression classes and age groups during training, making underrepresented (expression, age) pairs contribute proportionally more to the gradient. The paper's diagnosis uses saliency heatmaps aggregated per expression and age group, computed with single-pass saliency maps rather than the more expensive LIME, to show that the baseline attends to largely overlapping facial regions for neutral, sadness, and anger—exactly the confused classes in elderly faces—and that the improved model shifts attention for sadness and happiness. The other two strategies, multi-task learning with a parallel age-regression head and multi-modal input that concatenates the age scalar to a downsampled visual feature vector, are alternative routes to injecting age information, but the weighted loss produces the largest elderly gains.

What would settle it

Run the best Age-weighted Loss model on an independent, larger elderly dataset with all seven expressions and manual age labels; if the adult–elderly macro-F1 gap does not shrink by roughly the same margin, the claimed mitigation is specific to the FACES subset rather than to age bias itself.

Watch

Extended reading notes

Core claim

The central claim is that age-related bias in deep FER originates mainly in training-data representation rather than in architecture, so simple training modifications can substantially reduce it. On AffectNet, with age labels automatically estimated by MiVOLO, the paper trains MobileNetV3, Swin Transformer, and ConvNeXt-Base in four configurations and finds that all three mitigation strategies—Age-weighted Loss, Multi-task Learning, and Multi-modal Input—improve recognition for elderly faces overall. The strongest result is ConvNeXt-Base with Age-weighted Loss: elderly macro F1 on the FACES benchmark rises from 0.7528 to 0.8140, and the adult–elderly gap falls from 0.0917 to 0.0380. Confusion-matrix analysis shows the main gains come from resolving elderly neutral–anger–sadness confusions and the sadness–disgust confusion, and saliency heatmaps show the age-aware model shifting attention to more informative facial regions for elderly faces, such as the brow and mouth edges for sadness. The paper reads this as evidence that even approximate, automatically generated demographic labels are enough to make large-scale FER training fairer across age groups.

Load-bearing premise

The entire elderly-side conclusion rests on a single small benchmark: the FACES elderly subset contains 114 images and no surprise expressions, so the measured bias reduction may not transfer to elderly faces more broadly.

Editorial extensions

If this is right

  • Adding Age-weighted Loss to a standard FER training pipeline improves elderly recognition without changing the network architecture, and it combines with both multi-task and multi-modal age-injection strategies.
  • The adult–elderly macro-F1 gap on FACES drops from 0.0917 to 0.0380 for ConvNeXt-Base, and elderly macro F1 rises from 0.7528 to 0.8140, while overall performance across age groups stays roughly stable.
  • The elderly misclassifications that most hurt fairness—neutral read as anger or sadness, and sadness read as disgust—are the ones that shrink most under age-aware training.
  • Models trained on AffectNet with automatically estimated age labels outperform a specialized age-group-aware method [31] on the elderly FACES subset, suggesting that dataset diversity and simple reweighting can rival bespoke architectures.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension the paper leaves open is to feed the multi-task model's own age estimate back as the age input at inference; if the elderly gains persist, the approach stops depending on external age labels entirely.
  • The children's fear–surprise confusion survives all three mitigations, which suggests this particular bias may be driven by how the expressions are labeled or expressed rather than by training-set imbalance; re-annotating children's faces with adult labels would test that.
  • If rebalancing is the active mechanism, the same Age-weighted Loss should shrink performance gaps for other rare demographic groups, such as specific ethnicities or gender groups, on the same benchmark suite.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper studies age-related bias in deep facial expression recognition (FER) models, with a focus on elderly subjects. Using AffectNet as the training set with automatically estimated (MiVOLO) age labels, the authors evaluate three model-agnostic mitigation strategies: Age-weighted Loss (Eq. 1), Multi-task Learning with age regression, and Multi-modal Input with age concatenated to the feature vector. Evaluation is carried out on balanced, lab-controlled test sets for children, adults, and elderly (the latter exclusively from FACES, 114 images). The paper reports that Age-weighted Loss, in particular, raises elderly macro-F1 for ConvNeXt from 0.7528 to 0.8140 and narrows the adult-elderly gap from 0.0917 to 0.0380, with supporting confusion-matrix and saliency-heatmap analyses. The central claim is that simple training modifications can effectively mitigate age-related bias, and that approximate demographic labels are useful for fairness-oriented FER training.

Significance. If the reported elderly improvements were statistically robust and reproducible on other elderly data, this would be a useful contribution: it would show that model-agnostic, architecture-independent training changes can reduce demographic bias without the need for age-conditioned architectural modifications, and that automatically estimated age labels on a large, unlabeled dataset can support bias-aware training. Strengths of the paper include the public code release, the use of three network architectures, the systematic per-expression and per-age-group analysis, the XAI-driven attention comparison, and the comparison against a prior method (Huang et al.) on the same elderly subset. The main weakness is that the entire elderly evidence base is a single, small (114-image) dataset with no surprise class, the reported uncertainties reflect only training-fold variability, and the paper contains no significance testing or correction for multiple comparisons. The consistency claim is also not fully supported by Table 2, since MobileNetV3 with Age-weighted Loss has a lower average F1 than its baseline.

major comments (3)
  1. [Section 4.2, Table 2] The entire elderly evaluation rests on 114 FACES images, with no surprise class, which gives roughly 19 images per expression class. The reported standard deviations (e.g., 0.0182 for the ConvNeXt baseline) are computed over the five training folds on this fixed test set; they do not capture test-set sampling variability. With about 19 samples per class, a small number of prediction changes can shift macro-F1 by several points, so the headline improvement (ConvNeXt baseline 0.7528 to Age-weighted Loss 0.8140) may be within plausible sampling noise. No confidence intervals, bootstrap estimates, or significance tests are reported, and Section 7 concedes that no other elderly dataset was used. To support the abstract's claim that age-related bias is 'effectively mitigated,' the authors should either add a second elderly evaluation set, provide uncertainty estimates that account for test-set sampling, or substantially soften the claim.
  2. [Section 5.2, Figure 6] The 'best approach' for each network is selected on the same test sets used for evaluation, as stated in the caption: 'best-performing approach... based on the average F1 scores reported in Table 2.' This post hoc selection on the test data introduces a multiple-comparisons problem and inflates the apparent success of the chosen method. Moreover, the consistency claim in the abstract ('consistent improvements') is contradicted by Table 2: MobileNetV3 with Age-weighted Loss has an overall average F1 of 0.7740, below its baseline of 0.7804, and its elderly F1 gain is smaller than for the other two networks. The authors should report results for all methods without test-set selection, use a validation split for method choice, or apply multiple-comparison correction when claiming consistency.
  3. [Section 4.2 and Section 6.3] All training-age information comes from MiVOLO's automatic age labels on AffectNet, but the paper provides no quantitative validation of these labels on AffectNet itself. Section 6.3 asserts that the labels were 'sufficiently accurate' and that the results demonstrate this, yet no ground-truth check, error analysis, or sensitivity analysis is reported. Because the age-weighted loss and the age regression task both depend directly on these labels, label noise could bias or attenuate the measured effects of the mitigation strategies. A small manually labeled validation subset, or a perturbation of age-group boundaries, would make the robustness of the conclusions testable.
minor comments (4)
  1. [Table 1] RQ1.2 contains a typo: 'FER??' with a double question mark, which should be corrected.
  2. [Section 6.3] The sentence beginning 'Notably, training on AffectNet allowed our baseline...' reports values 'approximately .7329, .7535 and .7528 accuracy'; these are macro-F1 scores, not accuracy, and should be labeled consistently.
  3. [Sections 5.2 and 6.2] The word 'significant' is used repeatedly (e.g., 'All three strategies significantly reduced this confusion') without any statistical testing; consider replacing it with 'substantial' or reporting appropriate significance measures.
  4. [Figure 2 and Section 4.2] The text 'the image count per expression is approximately: 200+ for children, 500+ for adults, and exactly 114 for elderly' is ambiguous because 114 is the total elderly count, not a per-expression count; please clarify the wording.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Age-weighted loss is a training reweighting evaluated on held-out FACES, and the central mitigation claim does not reduce to its inputs.

full rationale

The paper's derivation chain is not circular. The Age-weighted Loss (Eq. 1) weights each training sample by N_min / N_{s,a}, the inverse frequency of the (expression, age-group) cell in AffectNet; this is a standard reweighting defined entirely from training-set counts, and the elderly gains are measured on held-out FACES images that were never used to set the weights. No fitted parameter is renamed as a prediction, no target quantity is defined in terms of the method, and no uniqueness theorem is imported from the authors' prior work. The multi-task and multi-modal strategies inject estimated age as a secondary target or input, and their held-out improvements are empirical outcomes rather than identities. Self-citations [19,20,21,43] support the XAI pipeline and background claims, but the bias-mitigation result does not reduce to any of them; the citation about AffectNet generalization is offered as an explanation for baseline superiority, not as the load-bearing step for the mitigation conclusion. Concerns about the 114-image elderly test set (Section 4.2) and unvalidated MiVOLO labels (Section 6.3) are statistical-validity issues, not circularity: even if the result is fragile, it is not equivalent to its inputs by construction.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new entities and no fitted constants, but it depends on several unverified domain assumptions. The two most consequential are the sufficiency of automatically estimated age labels and the representativeness of the 114-image FACES elderly set. The unspecified multi-task loss weight and density-weighting bandwidth are unreported hyperparameters that affect the results.

free parameters (2)
  • multi-task loss weight = not reported
    Section 4.5.2 combines age regression MSE with expression cross-entropy using a weighted sum but does not report the weight, which controls how strongly age learning is forced.
  • density-based weighting bandwidth = not reported
    Section 4.5.2 applies density-based weighting for imbalanced regression on the age task; the kernel/bandwidth hyperparameters are unspecified.
assumptions (5)
  • domain assumption MiVOLO automatic age estimates on AffectNet are accurate enough to define age groups and support age-aware training.
    Section 4.2 introduces MiVOLO labels; Section 6.3 asserts they are 'sufficiently accurate' without a ground-truth comparison on AffectNet.
  • domain assumption The FACES elderly subset (114 images, no surprise) is representative of elderly facial expression recognition.
    Section 4.2 selects FACES as the only elderly evaluation dataset; Section 5 uses it for all elderly results.
  • domain assumption Saliency maps faithfully identify the facial regions that drive model predictions.
    Section 4.3 adopts saliency maps for local explanations and uses them to compare attention across age groups and methods.
  • domain assumption Expression labels from AffectNet and from the controlled test datasets denote the same seven categories.
    Section 4.2 merges datasets with different annotation protocols and maps them to a common label set without a consistency check.
  • domain assumption Cross-dataset F1 on controlled frontal-face benchmarks is a valid measure of model generalization and bias.
    All conclusions are based on evaluating AffectNet-trained models on these external datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging the gap in FER: addressing age bias in deep learning." pith.science (2026). https://pith.science/paper/NI2GIR2C

@misc{pith2026250707638,
  author       = {Pith},
  title        = {Pith review of: Bridging the gap in FER: addressing age bias in deep learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NI2GIR2C}},
  note         = {Machine review of arXiv:2507.07638}
}
read the original abstract

Facial Expression Recognition (FER) systems based on deep learning have achieved impressive performance in recent years. However, these models often exhibit demographic biases, particularly with respect to age, which can compromise their fairness and reliability. In this work, we present a comprehensive study of age-related bias in deep FER models, with a particular focus on the elderly population. We first investigate whether recognition performance varies across age groups, which expressions are most affected, and whether model attention differs depending on age. Using Explainable AI (XAI) techniques, we identify systematic disparities in expression recognition and attention patterns, especially for "neutral", "sadness", and "anger" in elderly individuals. Based on these findings, we propose and evaluate three bias mitigation strategies: Multi-task Learning, Multi-modal Input, and Age-weighted Loss. Our models are trained on a large-scale dataset, AffectNet, with automatically estimated age labels and validated on balanced benchmark datasets that include underrepresented age groups. Results show consistent improvements in recognition accuracy for elderly individuals, particularly for the most error-prone expressions. Saliency heatmap analysis reveals that models trained with age-aware strategies attend to more relevant facial regions for each age group, helping to explain the observed improvements. These findings suggest that age-related bias in FER can be effectively mitigated using simple training modifications, and that even approximate demographic labels can be valuable for promoting fairness in large-scale affective computing systems.

Figures

Figures reproduced from arXiv: 2507.07638 by the authors.

Figure 1
Figure 1. (a) Class distribution and (b) age histogram of the AffectNet dataset. • Adults (18-59): FEGA [53], RaFD [38], KDEF [9], WSEFEP [50], and FACES [14] • Elderly (60+): FACES [14] While we collected four and five datasets for the children and adult groups, respectively, only one dataset (FACES) contained elderly individuals. Although the Lifespan dataset [46] includes some elderly faces, it was excluded due to class im… view at source ↗
Figure 2
Figure 2. shows the number of images per expression, dataset, and age group. We included the six basic expressions (happiness, sadness, surprise, fear, anger, and disgust), as defined by Ekman [16], plus the neutral expression. However, not all expressions were present in all datasets. For instance, DEFSS and NIMH-ChEFS lack “disgust” and “surprise,” while FACES lacks “surprise.” As seen in the figure, the image count per exp… view at source ↗
Figure 3
Figure 3. We also computed confusion matrices separately for each age group to better understand which expressions were misclassified ( [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Normalized confusion matrices by age group for the baseline approach. From left to right: children, adults, and elderly. From up to bottom: MobileNetV3, Swin Transformer and ConvNeXt-Base. Interestingly, these confusions also appear in the adult group, albeit to a less…
Figure 5
Figure 5. Figure 5: Explanation heatmaps by age group and expression for the ConvNeXt-Base baseline approach. Among the expressions, “happiness” and “disgust” exhibited the most distinct activation patterns. For “happiness”, the model consistently focused its attention on regions such as …
Figure 6
Figure 6. Figure 6: Performance of the best approach based on macro average F1 score by expression and age group. From left to right: MobileNetV3 Multi-modal Input, Swin Transformer Age-weighted Loss and ConvNeXt-Base Age-weighted Loss. “sadness” as “disgust”, resulting in a substantial i…
Figure 7
Figure 7. Figure 7: Normalized confusion matrices by age group for the ConvNeXt-Base Age-weighted Loss approach. From left to right: children, adults, and elderly [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Explanation heatmaps by age group and expression for the ConvNeXt-Base Age-weighted Loss approach. 6. Discussion To structure our discussion, we revisit the two research questions outlined in Section 3. 6.1. RQ1: Are there significant differences in model performance a…
Figure 9
Figure 9. Figure 9: Explanation heatmaps by age group and expression for MobileNetV3 baseline approach. A.2. Improvement strategies detailed results [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Explanation heatmaps by age group and expression for Swin Transformer baseline approach. confusion between “sadness” and “neutral” expressions. Notably, this particular interaction does not manifest in the adult age group. For the middle-age group, no significant diff…
Figure 11
Figure 11. Figure 11: Per-class performance of the different methods addressing age bias. From left to right: children, adults and elderly. From up to bottom: MobileNetV3, Swin Transformer and ConvNeXt-Base [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: Normalized confusion matrices by age group for the MobileNetV3 multi-modal approach. From left to right: children, adults and elderly. Gaya-Morey et al.: Preprint submitted to Elsevier Page 17 of 21 [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: Normalized confusion matrices by age group for the Swin Transformer age-weighted loss approach. From left to right: children, adults and elderly [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]
Figure 14
Figure 14. Figure 14: Explanation heatmaps by age group and expression for MobileNetV3 Multi-modal Input approach. Gaya-Morey et al.: Preprint submitted to Elsevier Page 18 of 21 [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]
Figure 15
Figure 15. Figure 15: Explanation heatmaps by age group and expression for Swin Transformer age-weighted loss approach. Data availability All code is made publicly available at https://github.com/Xavi3398/FER-for-elderly. References [1] Adadi, A., Berrada, M., 2018. Peeking inside the blac…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

66 extracted references · 39 canonical work pages

  1. [1]

    Peekinginsidetheblack-box:Asurveyonexplainableartificialintelligence(xai)

    Adadi,A.,Berrada,M.,2018. Peekinginsidetheblack-box:Asurveyonexplainableartificialintelligence(xai). IEEEAccess6,52138–52160. doi:10.1109/ACCESS.2018.2870052

  2. [2]

    Fully automated age-weighted expression classification using real and apparent age

    Al-Garaawi, N., Morris, T., Cootes, T.F., 2022. Fully automated age-weighted expression classification using real and apparent age. Pattern Analysis and Applications 25, 451–466. doi:10.1007/s10044-021-01044-1

  3. [3]

    Alvi, M., Zisserman, A., Nellåker, C., 2018. Turning a blind eye: Explicit removal of biases and variation from deep neural network embeddings,in:ComputerVision–ECCV2018Workshops:Munich,Germany,September8-14,2018,Proceedings,PartI,Springer-Verlag. p. 556–572. doi:10.1007/978-3-030-11009-3_34

  4. [4]

    A review study: The effect of face aging at estimating age and face recognition

    Atallah, R.R., Kamsin, A., Ismail, M.A., 2019. A review study: The effect of face aging at estimating age and face recognition. Journal of Physics: Conference Series 1339, 12006. doi:10.1088/1742-6596/1339/1/012006

  5. [5]

    Explainableartificialintelligence(xai):Concepts,taxonomies,opportunitiesandchallengestowardresponsible ai

    BarredoArrieta,A.,Díaz-Rodríguez,N.,DelSer,J.,Bennetot,A.,Tabik,S.,Barbado,A.,Garcia,S.,Gil-Lopez,S.,Molina,D.,Benjamins,R., Chatila,R.,Herrera,F.,2020. Explainableartificialintelligence(xai):Concepts,taxonomies,opportunitiesandchallengestowardresponsible ai. Information Fusion 58, 82–115. doi:10.1016/j.inffus.2019.12.012

  6. [6]

    A survey of predictive modeling on imbalanced domains

    Branco, P., Torgo, L., Ribeiro, R.P., 2016. A survey of predictive modeling on imbalanced domains. ACM Comput. Surv. 49. doi:10.1145/ 2907070

  7. [7]

    Further evidence on preschoolers’ interpretation of facial expressions

    Bullock, M., Russell, J.A., 1985. Further evidence on preschoolers’ interpretation of facial expressions. International Journal of Behavioral Development 8, 15–38. doi:10.1177/016502548500800103

  8. [8]

    Comparisonbetweendeeplearningmodelsandtraditionalmachinelearningapproachesforfacial expressionrecognitioninageingadults

    Caroppo,A.,Leone,A.,Siciliano,P.,2020. Comparisonbetweendeeplearningmodelsandtraditionalmachinelearningapproachesforfacial expressionrecognitioninageingadults. JournalofComputerScienceandTechnology35,1127–1146. doi: 10.1007/s11390-020-9665-4

Show all 66 references
  1. [9]

    Karolinska directed emotional faces

    D., L., A., F., Öhman A., 1998. Karolinska directed emotional faces. PsycTESTS Dataset 91, 630. doi:10.1037/t27732-000

  2. [10]

    Thedartmouthdatabaseofchildren’sfaces:Acquisitionandvalidationofanewfacestimulus set

    Dalrymple,K.A.,Gomez,J.,Duchaine,B.,2013. Thedartmouthdatabaseofchildren’sfaces:Acquisitionandvalidationofanewfacestimulus set. PLoS ONE 8. doi:10.1371/journal.pone.0079131

  3. [11]

    The Expression of the Emotions in Man and Animals

    Darwin, C., Ekman, P., Prodger, P., 1998. The Expression of the Emotions in Man and Animals. Oxford University Press

  4. [12]

    Demographicbiasinbiometrics:Asurveyonanemergingchallenge

    Drozdowski,P.,Rathgeb,C.,Dantcheva,A.,Damer,N.,Busch,C.,2020. Demographicbiasinbiometrics:Asurveyonanemergingchallenge. IEEE Transactions on Technology and Society 1, 89–103. doi:10.1109/TTS.2020.2992344

  5. [13]

    Fairness through awareness, in: Proceedings of the 3rd Innovations in Theoretical Computer Science Conference, p

    Dwork, C., Hardt, M., Pitassi, T., Reingold, O., Zemel, R., 2012. Fairness through awareness, in: Proceedings of the 3rd Innovations in Theoretical Computer Science Conference, p. 214–226. doi:10.1145/2090236.2090255

  6. [14]

    Faces-a database of facial expressions in young, middle-aged, and older women and men: Development and validation

    Ebner, N.C., Riediger, M., Lindenberger, U., 2010. Faces-a database of facial expressions in young, middle-aged, and older women and men: Development and validation. Behavior Research Methods 42, 351–362. doi:10.3758/BRM.42.1.351. Gaya-Morey et al.:Preprint submitted to Elsevi...

  7. [15]

    The nimh child emotional faces picture set (nimh-chefs): A new set of children’s facial emotion stimuli

    Egger, H.L., Pine, D.S., Nelson, E., Leibenluft, E., Ernst, M., Towbin, K.E., Angold, A., 2011. The nimh child emotional faces picture set (nimh-chefs): A new set of children’s facial emotion stimuli. International Journal of Methods in Psychiatric Research 20, 145–156. doi:10...

  8. [16]

    An argument for basic emotions

    Ekman, P., 1992. An argument for basic emotions. Cognition and Emotion 6, 169–200. doi:10.1080/02699939208411068

  9. [17]

    Facial action coding system : investigator’s guide

    Ekman, P., Friesen, W.V., 1978. Facial action coding system : investigator’s guide. volume 3. Palo Alto, Calif

  10. [18]

    Facialageaffectsemotionalexpressiondecoding

    Fölster,M.,Hess,U.,Werheid,K.,2014. Facialageaffectsemotionalexpressiondecoding. FrontiersinPsychology5. doi: 10.3389/fpsyg. 2014.00030

  11. [19]

    Deeplearning-basedfacialexpressionrecognition for the elderly: A systematic review

    Gaya-Morey,F.X.,Buades-Rubio,J.M.,Palanque,P.,Lacuesta,R.,Manresa-Yee,C.,2025a. Deeplearning-basedfacialexpressionrecognition for the elderly: A systematic review. doi:10.48550/arXiv.2502.02618

  12. [20]

    Evaluating facial expression recognition datasets for deep learning: A benchmark study with novel similarity metrics

    Gaya-Morey, F.X., Manresa-Yee, C., Martinie, C., Buades-Rubio, J.M., 2025b. Evaluating facial expression recognition datasets for deep learning: A benchmark study with novel similarity metrics. doi:10.48550/arXiv.2503.20428

  13. [21]

    Unveiling the human-like similarities of automatic facial expression recognition: An empirical exploration through explainable ai

    Gaya-Morey, F.X., Ramis-Guarinos, S., Manresa-Yee, C., Buades-Rubio, J.M., 2024. Unveiling the human-like similarities of automatic facial expression recognition: An empirical exploration through explainable ai. Multimedia Tools and Applications doi: 10.1007/ s11042-024-20090-5

  14. [22]

    Le développement de la reconnaissance des expressions faciales des émotions chez l’enfant

    Gosselin, P., 1995. Le développement de la reconnaissance des expressions faciales des émotions chez l’enfant. Canadian Journal of Behavioural Science/Revue canadienne des sciences du comportement 27, 107

  15. [23]

    Children’sknowledgeoffacialexpressionsofemotions:Distinguishingfearandsurprise

    Gosselin,P.,and,J.S.,1999. Children’sknowledgeoffacialexpressionsofemotions:Distinguishingfearandsurprise. TheJournalofGenetic Psychology 160, 181–193. doi:10.1080/00221329909595391

  16. [24]

    Having difficulties reading the facial expression of older individuals? blame it on the facial muscles, not the wrinkles

    Grondhuis, S.N., Jimmy, A., Teague, C., Brunet, N.M., 2021. Having difficulties reading the facial expression of older individuals? blame it on the facial muscles, not the wrinkles. Frontiers in Psychology 12. doi:10.3389/fpsyg.2021.620768

  17. [25]

    Xai—explainable artificial intelligence

    Gunning, D., Stefik, M., Choi, J., Miller, T., Stumpf, S., Yang, G.Z., 2019. Xai—explainable artificial intelligence. Science Robotics 4, eaay7120. doi:10.1126/scirobotics.aay7120

  18. [26]

    Facial expression recognition influenced by human aging

    Guo, G., Guo, R., Li, X., 2013. Facial expression recognition influenced by human aging. IEEE Transactions on Affective Computing 4, 291–298. doi:10.1109/T-AFFC.2013.13

  19. [27]

    Human age estimation: What is the influence across race and gender?, in: 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Workshops, pp

    Guo, G., Mu, G., 2010. Human age estimation: What is the influence across race and gender?, in: 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Workshops, pp. 71–78. doi:10.1109/CVPRW.2010.5543609

  20. [28]

    Deep residual learning for image recognition, in: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

    He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for image recognition, in: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778. doi:10.1109/CVPR.2016.90

  21. [29]

    Searchingformobilenetv3,in:2019IEEE/CVFInternationalConferenceonComputerVision(ICCV),pp.1314–1324

    Howard, A., Sandler, M., Chen, B., Wang, W., Chen, L.C., Tan, M., Chu, G., Vasudevan, V., Zhu, Y., Pang, R., Adam, H., Le, Q., 2019. Searchingformobilenetv3,in:2019IEEE/CVFInternationalConferenceonComputerVision(ICCV),pp.1314–1324. doi: 10.1109/ICCV. 2019.00140

  22. [30]

    Mobilenets: Efficient convolutional neural networks for mobile vision applications

    Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H., 2017. Mobilenets: Efficient convolutional neural networks for mobile vision applications. doi:10.48550/arXiv.1704.04861

  23. [31]

    Facial expression recognition with age-group expression feature learning, in: 2024 International Joint Conference on Neural Networks (IJCNN), pp

    Huang, Y., Peng, J., Cai, Z., Guo, J., Chen, G., Tan, S., 2024. Facial expression recognition with age-group expression feature learning, in: 2024 International Joint Conference on Neural Networks (IJCNN), pp. 1–8. doi:10.1109/IJCNN60899.2024.10649944

  24. [32]

    Bringinganecologicalperspectivetothestudyofagingandrecognitionofemotionalfacialexpressions: Past, current, and future methods

    Isaacowitz,D.M.,Stanley,J.T.,2011. Bringinganecologicalperspectivetothestudyofagingandrecognitionofemotionalfacialexpressions: Past, current, and future methods. Journal of Nonverbal Behavior 35, 261–278. doi:10.1007/s10919-011-0113-6

  25. [33]

    Expression recognition across age, in: 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021), pp

    Jannat, S.R., Canavan, S., 2021. Expression recognition across age, in: 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021), pp. 1–5. doi:10.1109/FG52635.2021.9667062

  26. [34]

    Ultralytics yolo

    Jocher, G., Chaurasia, A., Qiu, J., 2024. Ultralytics yolo. URL:https://github.com/ultralytics/ultralytics

  27. [35]

    Kim, E., Bryant, D., Srikanth, D., Howard, A., 2021. Age bias in emotion detection: An analysis of facial emotion recognition performance on young, middle-aged, and older adults, in: Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, pp. 638–644. doi:10.11...

  28. [36]

    Changes in computer-analyzed facial expressions with age

    Ko, H., Kim, K., Bae, M., Seo, M.G., Nam, G., Park, S., Park, S., Ihm, J., Lee, J.Y., 2021. Changes in computer-analyzed facial expressions with age. Sensors 21. doi:10.3390/s21144858

  29. [37]

    Mivolo:Multi-inputtransformerforageandgenderestimation,in:AnalysisofImages,SocialNetworks and Texts, Springer Nature Switzerland, Cham

    Kuprashevich,M.,Tolstykh,I.,2024. Mivolo:Multi-inputtransformerforageandgenderestimation,in:AnalysisofImages,SocialNetworks and Texts, Springer Nature Switzerland, Cham. pp. 212–226

  30. [38]

    Presentation and validation of the radboud faces database

    Langner, O., Dotsch, R., Bijlstra, G., Wigboldus, D.H., Hawk, S.T., van Knippenberg, A., 2010. Presentation and validation of the radboud faces database. Cognition and Emotion 24, 1377–1388. doi:10.1080/02699930903485076

  31. [39]

    Deep facial expression recognition: A survey

    Li, S., Deng, W., 2022. Deep facial expression recognition: A survey. IEEE Transactions on Affective Computing 13, 1195–1215. doi:10.1109/TAFFC.2020.2981446

  32. [40]

    Exploring disentangled feature representation beyond face identification, in: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Liu, Y., Wei, F., Shao, J., Sheng, L., Yan, J., Wang, X., 2018. Exploring disentangled feature representation beyond face identification, in: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2080–2089. doi:10.1109/CVPR.2018.00222

  33. [41]

    Swin transformer: Hierarchical vision transformer using shifted windows, in: IEEE/CVF International Conference on Computer Vision (ICCV), pp

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B., 2021. Swin transformer: Hierarchical vision transformer using shifted windows, in: IEEE/CVF International Conference on Computer Vision (ICCV), pp. 10012–10022. doi:10.48550/arXiv.2103.14030

  34. [42]

    A convnet for the 2020s, in: IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S., 2022. A convnet for the 2020s, in: IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11966–11976. doi:10.48550/arXiv.2201.03545

  35. [43]

    Analysisofgenderdifferencesinfacialexpressionrecognitionbasedondeeplearningusing explainable artificial intelligence

    Manresa-Yee,C.,Ramis,S.,Buades,J.M.,2023. Analysisofgenderdifferencesinfacialexpressionrecognitionbasedondeeplearningusing explainable artificial intelligence. International Journal of Interactive Multimedia and Artificial Intelligence In Press, 1–10. doi:10.9781/ ijimai.2023.04.003

  36. [44]

    A survey on bias and fairness in machine learning

    Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., Galstyan, A., 2021. A survey on bias and fairness in machine learning. ACM Comput. Surv. 54. doi:10.1145/3457607. Gaya-Morey et al.:Preprint submitted to Elsevier Page 20 of 21 Bridging the gap in FER: addressing age bias in...

  37. [45]

    Thecreationandvalidationofthedevelopmentalemotionalfacesstimulusset

    Meuwissen,A.S.,Anderson,J.E.,Zelazo,P.D.,2017. Thecreationandvalidationofthedevelopmentalemotionalfacesstimulusset. Behavior Research Methods 49, 960–966. doi:10.3758/s13428-016-0756-7

  38. [46]

    Alifespandatabaseofadultfacialstimuli

    Minear,M.,Park,D.C.,2004. Alifespandatabaseofadultfacialstimuli. BehaviorResearchMethods,Instruments,&Computers36,630–633. doi:10.3758/BF03206543

  39. [47]

    Assessingfidelityinxaipost-hoctechniques:Acomparativestudywithgroundtruth explanations datasets

    Miró-Nicolau,M.,iCapó,A.J.,Moyà-Alcover,G.,2024. Assessingfidelityinxaipost-hoctechniques:Acomparativestudywithgroundtruth explanations datasets. Artificial Intelligence 335, 104179. doi:10.1016/j.artint.2024.104179

  40. [48]

    Affectnet: A database for facial expression, valence, and arousal computing in the wild

    Mollahosseini, A., Hasani, B., Mahoor, M.H., 2017. Affectnet: A database for facial expression, valence, and arousal computing in the wild. IEEE Transactions on Affective Computing 10, 18–31. doi:10.1109/TAFFC.2017.2740923

  41. [49]

    Facial expression recognition software | FaceReader

    Noldus Information Technology b.v, 2021. Facial expression recognition software | FaceReader. URL: https://www.noldus.com/ facereader. Accessed 22nd Nov 2024

  42. [50]

    Warsawsetofemotionalfacialexpression pictures: A validation study of facial display photographs

    Olszanowski,M.,Pochwatko,G.,Kuklinski,K.,Scibor-Rylski,M.,Lewinski,P.,Ohme,R.K.,2014. Warsawsetofemotionalfacialexpression pictures: A validation study of facial display photographs. Frontiers in Psychology 5. doi:10.3389/fpsyg.2014.01516

  43. [51]

    Facial emotion recognition analysis based on age-biased data

    Park, H., Shin, Y., Song, K., Yun, C., Jang, D., 2022. Facial emotion recognition analysis based on age-biased data. Applied Sciences 12. doi:10.3390/app12167992

  44. [52]

    Shape preserving facial landmarks with graph attention networks, in: 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022, BMVA Press

    Prados-Torreblanca, A., Buenaposada, J.M., Baumela, L., 2022. Shape preserving facial landmarks with graph attention networks, in: 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022, BMVA Press. doi:10.48550/arXiv.2210. 07233

  45. [53]

    A novel approach to cross dataset studies in facial expression recognition

    Ramis, S., Buades, J.M., Perales, F.J., Manresa-Yee, C., 2022. A novel approach to cross dataset studies in facial expression recognition. Multimedia Tools and Applications 81, 39507–39544. doi:10.1007/s11042-022-13117-2

  46. [54]

    Explainablefacialexpressionrecognitionforpeoplewith intellectual disabilities, in: Proceedings of the XXIII International Conference on Human Computer Interaction

    RamisGuarinos,S.,ManresaYee,C.,BuadesRubio,J.M.,Gaya-Morey,F.X.,2024. Explainablefacialexpressionrecognitionforpeoplewith intellectual disabilities, in: Proceedings of the XXIII International Conference on Human Computer Interaction. doi:10.1145/3612783. 3612789

  47. [55]

    Whyshoulditrustyou?:Explainingthepredictionsofanyclassifier,in:Proceedingsofthe22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, p

    Ribeiro,M.T.,Singh,S.,Guestrin,C.,2016. Whyshoulditrustyou?:Explainingthepredictionsofanyclassifier,in:Proceedingsofthe22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, p. 1135–1144. doi:10.1145/2939672.2939778

  48. [56]

    Rothe,R.,Timofte,R.,VanGool,L.,2015.Dex:Deepexpectationofapparentagefromasingleimage,in:2015IEEEInternationalConference on Computer Vision Workshop (ICCVW), pp. 252–257. doi:10.1109/ICCVW.2015.41

  49. [57]

    A meta-analytic review of emotion recognition and aging: Implications for neuropsychological models of aging

    Ruffman, T., Henry, J.D., Livingstone, V., Phillips, L.H., 2008. A meta-analytic review of emotion recognition and aging: Implications for neuropsychological models of aging. Neuroscience & Biobehavioral Reviews 32, 863–881. doi:10.1016/j.neubiorev.2008.01.001

  50. [58]

    Simonyan, K., Vedaldi, A., Zisserman, A., 2014. Deep inside convolutional networks: Visualising image classification models and saliency maps, in: 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Workshop Track Proceedings

  51. [59]

    Density-based weighting for imbalanced regression

    Steininger, M., Kobs, K., Davidson, P., Krause, A., Hotho, A., 2021. Density-based weighting for imbalanced regression. Machine Learning 110, 2187–2211. doi:10.1007/s10994-021-06023-5

  52. [60]

    Suresh, H., Guttag, J., 2021. A framework for understanding sources of harm throughout the machine learning life cycle, in: Proceedings of the 1st ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization. doi:10.1145/3465416.3483305

  53. [61]

    A computational study on aging effect for facial expression recognition

    Sönmez, E.B., 2019. A computational study on aging effect for facial expression recognition. Turkish Journal of Electrical Engineering and Computer Sciences 27, 2430–2443. doi:10.3906/elk-1811-70

  54. [62]

    Enhancedfacialexpressionrecognitionbyage,in:201511thIEEEInternationalConferenceandWorkshops on Automatic Face and Gesture Recognition (FG), pp

    Wu,S.,Wang,S.,Wang,J.,2015. Enhancedfacialexpressionrecognitionbyage,in:201511thIEEEInternationalConferenceandWorkshops on Automatic Face and Gesture Recognition (FG), pp. 1–6. doi:10.1109/FG.2015.7163117

  55. [63]

    Investigating bias and fairness in facial expression recognition, in: Computer Vision – ECCV 2020 Workshops, Springer International Publishing, Cham

    Xu, T., White, J., Kalkan, S., Gunes, H., 2020. Investigating bias and fairness in facial expression recognition, in: Computer Vision – ECCV 2020 Workshops, Springer International Publishing, Cham. pp. 506–523

  56. [64]

    Joint estimation of age and expression by combining scattering and convolutional networks

    Yang, H.F., Lin, B.Y., Chang, K.Y., Chen, C.S., 2018. Joint estimation of age and expression by combining scattering and convolutional networks. ACM Trans. Multimedia Comput. Commun. Appl. 14. doi:10.1145/3152118

  57. [65]

    Ageprogression/regressionbyconditionaladversarialautoencoder,in:2017IEEEConferenceonComputer Vision and Pattern Recognition (CVPR), pp

    Zhang,Z.,Song,Y.,Qi,H.,2017. Ageprogression/regressionbyconditionaladversarialautoencoder,in:2017IEEEConferenceonComputer Vision and Pattern Recognition (CVPR), pp. 4352–4360. doi:10.1109/CVPR.2017.463

  58. [66]

    Defining standard values for facereader facial expression software output

    Zhu, A., Boonipat, T., Cherukuri, S., Bite, U., 2024. Defining standard values for facereader facial expression software output. Aesthetic Plastic Surgery 48, 785–792. doi:10.1007/s00266-023-03468-y. Gaya-Morey et al.:Preprint submitted to Elsevier Page 21 of 21

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.