Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

A Shared Latent for Partially-Labeled Multi-Task Facial Affect Recognition

T0 review · 2 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Partially labeled facial affect frames, treated as observations of one shared affect latent, let a frame labeled for a single task shape all three decoders; on the s-Aff-Wild2 validation split this lifts expression macro-F1 from 0.403 to 0.

desk verdict A disciplined, well-controlled benchmark paper whose formulation and empirical care mostly land; the one missing control—per-task latents vs a shared latent—leaves its central coupling claim under-determined, but the paper still deserves a fair referee. read the letter →

arxiv 2607.16285 v1 pith:V3NMYD2G submitted 2026-07-11 cs.CV

Hong Hai Nguyen , Sy Phan Van , Soo-Hyung Kim , Van-Thong Huynh This is my paper
classification cs.CV
keywords facialaffectrecognitionmulti-tasklearningpartiallylabeleddatasharedlatentvariationalboundevidenceloweractionunitsexpression
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

The paper claims that the usual way of handling partially labeled multi-task facial affect data—masking each task's loss on frames where it is unlabeled—throws away the cross-task signal, and that treating the missing labels as quantities to marginalize over a single shared affect latent recovers that signal. The masked loss, far from a heuristic, reappears as the reconstruction term of an evidence lower bound, so a frame annotated for any one task shapes the representation the other task decoders use. On the s-Aff-Wild2 split the shared latent lifts expression macro-F1 from 0.403 to 0.446 over a dedicated specialist, and pairing two near-peer face backbones with decorrelated errors lifts action-unit F1 from 0.543 to 0.556; replacing the latent with the standard masked-loss model costs 0.046 in the combined score entirely in expression. The authors present matched-control negatives to argue that the persistent failure on rare classes is representational, not a matter of loss shaping, and report a small, regime-dependent transfer of the expression gain to AffectNet and RAF-DB as exploratory.

What carries the argument

The shared affect latent: a 96-dimensional stochastic bottleneck z ~ N(mu, sigma^2) encoded linearly from the bidirectional recurrent temporal feature h, with three linear decoders—Gaussian for valence-arousal, categorical for expression, independent Bernoullis for action units—that are conditionally independent given z. The load-bearing identity is that minimizing the evidence lower bound with a beta-weighted KL term reproduces the standard masked loss plus a regularizer, so partial labels are marginalized rather than ignored; the beta weight turns out to control how strongly the bottleneck separates rare expression classes. A second ingredient is the requirement that the second backbone be

What would settle it

On the same backbone and temporal model, replace the shared latent with the masked-loss model and perform a wide search over post-hoc logit adjustments and class-conditional temperatures on the expression head; if anger and fear each reach macro-F1 levels comparable to those attained with the latent assembly, or if expression overall reaches 0.446 or better, the claim that no decision-rule change can recover the rare classes is refuted.

Watch

Extended reading notes

Core claim

The central reduction is that the masked multi-task objective used throughout this benchmark is exactly the reconstruction term of the evidence lower bound for a generative model in which three heterogeneous labels—valence-arousal regression, eight-way expression classification, and twelve-way action-unit multi-label—are conditionally independent given a shared latent drawn from a standard normal prior, with a learned inference network over the temporal feature. A frame annotated for only one task still constrains the posterior the other decoders draw on; unobserved tasks integrate to one, so the marginal likelihood of the observed labels is a product over observed tasks. Empirically this co

Load-bearing premise

The conclusion that the rare-class failure is representational rather than a matter of loss shaping rests on the set of loss-side and decision-rule interventions tried (class weights, focal/BCE, margin and asymmetric losses, logit adjustment, oversampling, post-hoc logit adjustment, and circumplex-gated logits); if some untested loss-side mechanism separates anger from fear, the mechanistic conclusion would need revision even though the positive gains from the shared latent c

Editorial extensions

If this is right

  • If a frame labeled for one task shapes the representation used by all decoders, then the partially labeled majority of the data—63% of frames lacking at least one label—is no longer dead weight; every partially annotated frame contributes to all three tasks.
  • The rare-class failure in facial affect recognition is characterized as a problem of representation rather than of loss weighting, so the standard reweighting, logit-adjustment, and resampling toolbox should not be expected to close the gap on this benchmark.
  • Combining backbones by averaging predictions only pays off when the second model is a near-peer with decorrelated errors; a weaker or larger-but-worse encoder does not help, which gives a cheap pre-training criterion for whether an extra model will earn its cost.
  • The masked loss is not an ad hoc heuristic but a valid variational objective, and the same marginalization construction transfers to any setting where several heterogeneous tasks are annotated on largely disjoint examples, such as medical imaging and multimodal behavior analysis.
  • The assembled score of 1.679 on validation is presented as an in-sample endpoint whose per-task sources are selected on validation; the paper rests its conclusions on the paired, one-factor controlled comparisons rather than the absolute number.

Reading between the lines

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

  • If the representational account is right, the anger–fear confusion that survives every tested intervention is a signature of the two classes occupying the same region of valence-arousal space; an untested extension would be contrastive training that explicitly pushes anger apart from fear in the latent, or auxiliary semantic labels that separate them, rather than another loss reweighting.
  • The near-peer condition on the second backbone could be operationalized before any training: compute per-task error vectors on a small held-out set and require both decorrelation and comparable strength; the paper's results imply a simple yes/no test for whether a candidate backbone will improve the ensemble.
  • The identity that the masked loss is a reconstruction term suggests a cheap probe: if the posterior mean of the latent is a sufficient encoder, linear probes on the 96-dimensional code should match or exceed probes on the raw backbone features for rare classes on other affect benchmarks; the RAF-DB compound-expression transfer is consistent with this and could be tested with a controlled bottlenec
  • A structured prior—for example a mixture of per-expression components—was not tested and might push expression further without hurting valence-arousal, given the paper's finding that the KL weight, not the decoder, governs rare-class separation.
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

2 major / 5 minor

Summary. The paper casts partially-labeled multi-task facial affect recognition on s-Aff-Wild2 (valence–arousal regression, 8-way expression classification, 12-way AU detection) as marginalization over a shared 96-dimensional variational latent. The masked per-task losses are shown to be the reconstruction term of an ELBO (Eq. 4–5), with a β-KL regularizer. On a single AffectNet-initialized ViT-B/16 backbone, the shared latent raises expression macro-F1 from 0.403 (dedicated specialist) to 0.446 (p=0.017); adding a second, near-peer backbone with decorrelated errors raises AU macro-F1 from 0.543 to 0.556 (p=0.005); and an all-else-equal replacement of the latent by the masked-loss model on the same two backbones drops PMTL by 0.046 entirely in expression (p=0.001). The paper reports extensive matched negative controls — loss reweighting, LDAM/ASL, logit adjustment, oversampling, post-hoc logit adjustment, circumplex-gated logits, mean-teacher, external AU data — to argue that the rare-class failure is representational rather than loss-side. The assembled PMTL=1.679 is transparently reported as an in-sample endpoint because task sources are selected on validation; AffectNet and RAF-DB transfer results are presented as exploratory.

Significance. If the central claim holds, the paper contributes a principled way to exploit partial multi-task labels by treating missing labels as marginalized latent variables, and it provides a carefully controlled empirical study on a benchmark with restricted pretraining. Strengths include the correct ELBO derivation under the stated conditional-independence assumption, seed-averaged bootstrap tests against matched controls, a nested protocol bounding source-selection optimism, and explicit negative results. The main weakness is that the distinctive claim — that cross-task coupling through a shared latent is the active ingredient — is not isolated from the effect of a per-task variational bottleneck; the paper itself concedes this in the AffectNet discussion. The contribution is promising but requires one more critical control or a substantially softened claim.

major comments (2)
  1. [Sec. 3, Eq. (4)–(5); Sec. 5, Fig. 10 and Tab. 10 discussion] The paper's central claim is that a single shared latent couples the three tasks, so a frame labeled for one task shapes the representation the other decoders use (Abstract; Sec. 1). The experiments compare this shared-latent ELBO against a masked-loss model with no latent, and ablate β and bottleneck width. This changes two factors simultaneously: the introduction of a variational stochastic bottleneck and the sharing of that bottleneck across tasks. Fig. 10 shows that the β-KL regularizer and capacity matter, but it does not show that sharing is the active ingredient. A control with three independent per-task latents — same ELBO form, same β, same total capacity, same masked losses, same backbones — would isolate the cross-task coupling. Without it, the observed expression gain could be due entirely to a per-task variational regularizer, in which case the 'shared' characterization is u
  2. [Sec. 4, Tab. 6; Sec. 5, circumplex gate and post-hoc logit adjustment] The conclusion that 'the rare-class failure is representational, not a matter of loss shaping' (Abstract; Sec. 7) is an inference from a set of null results on specific loss-side interventions: class-weighted focal/BCE, LDAM, ASL, logit adjustment, oversampling, post-hoc logit adjustment, and a circumplex-gated decoder. Absence of improvement across this sampled family is evidence, but it is not proof that no loss-side mechanism would help; the causal attribution is broader than the tested set. The paper is appropriately cautious in calling the companion result 'corroborated' and the transfer experiments 'not a mechanism proof' (Sec. 5), but the Abstract and Sec. 7 state the causal conclusion more strongly. I recommend either adding the per-task-latent control described above, which would also strengthen the representational interpretation, or explicitly demoting the loss-shaping claim t
minor comments (5)
  1. [Sec. 4, Tab. 3 note] The text says the bootstrap tests were 'few and pre-registered rather than a search.' If 'pre-registered' is meant literally, a registration URL or dated protocol is needed; otherwise the unadjusted one-sided p-values rest on an unverifiable assumption. Please clarify.
  2. [Eq. (5)] As written, L denotes the objective being minimized; it is helpful to state explicitly that this is the negative ELBO (or add a minus sign on the right-hand side) to avoid sign confusion with the usual ELBO convention.
  3. [Sec. 3, Eq. (2)] Please specify boundary handling for the BiGRU window near the start and end of each video (padding, truncation, or valid-only). This affects reproducibility of the temporal-model results.
  4. [Fig. 4 caption] The caption notes that absolute scores sit below the multi-seed 1.509 because of a single seed and fixed stride. It would be useful to state the stride value and to make clear whether the ablation uses the same architecture and recipe except for window size and seed.
  5. [Sec. 5, post-hoc logit adjustment] The statement that 'post-hoc logit adjustment selects a near-zero margin' would be more informative with the selected margin value and the resulting expression macro-F1; a quantitative null is easier to interpret.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the ELBO objective and controlled comparisons stand independently; only a minor, non-load-bearing self-citation to companion work.

full rationale

The paper's central derivation, Eq. (4) to Eq. (5), is a standard variational lower bound: the per-task likelihoods in Eq. (3) are defined so that their negative log-likelihoods are the losses of Eq. (1), missing tasks integrate to one, and the masked loss appears as the reconstruction term plus a KL term. This is a mathematical reformulation, not a fitted equivalence or a parameter renamed as a prediction. The main empirical claim is tested with controlled comparisons: the shared latent is compared against a matched masked-loss model on the same two backbones (Tab. 3: 0.470 vs 0.424, p=0.001), against a mean-teacher external baseline (Tab. 9), and against a dedicated specialist (0.403 to 0.446, p=0.017); these are real held-out validation measurements, not quantities forced by the construction. The 'representational, not loss-shaping' conclusion does cite companion work by overlapping authors ('Our design follows a diagnosis, made in companion work (Huynh et al., 2026)'), but the paper independently re-tests the claim with loss-side ablations (Tab. 6: LDAM, ASL, logit adjustment do not improve) and a shuffled-geometry decoder control in Sec. 5, so the self-citation is not load-bearing. The AffectNet and RAF-DB transfers are explicitly labeled exploratory and use features that did not see those labels. The absence of a per-task-latent control is a potential confound between cross-task sharing and the variational regularizer, but that is an experimental attribution gap, not a reduction of a prediction to its input by construction. The only issue of note is the minor self-citation, which does not rise to circularity.

Assumptions & free parameters 8 free parameters · 7 assumptions · 0 invented entities

The central construction rests on standard VAE machinery, conditional independence of tasks given z, and ignorable missingness. Eight free parameters or selection choices enter the reported numbers, most tuned or chosen on the validation split; the shared latent itself is a modeling device, not a new physical entity, and no independent observational handle is proposed for it.

free parameters (8)
  • beta (KL divergence weight) = 0.05, annealed from 0
    Chosen by sweep {0.01, 0.05, 0.1, 0.2} on validation (Fig. 10); it directly controls the expression gain.
  • latent dimension = 96
    Chosen by hand; a narrower bottleneck drops expression to 0.391, wider not reported. Load-bearing for the latent's benefit.
  • temporal window L = 64 frames
    Chosen from a dose-response sweep; PMTL climbs monotonically with window length (Fig. 4).
  • per-AU decision thresholds = 12 calibrated thresholds
    Each AU decision threshold is calibrated on the validation split and contributes to the reported AU F1.
  • task source selection = VA/AU from 2-backbone specialists; EXPR from shared latent on both backbones
    Each task's source is chosen on validation; PMTL=1.679 is therefore an in-sample endpoint (nested protocol estimates 0.010 optimism).
  • homoscedastic uncertainty weights = 3 learned scalars (sigma_t)
    Per-task observation precisions learned via Kendall et al. weighting; part of Eq. (1)-(5).
  • class weights for focal/BCE = not specified
    Expression and AU losses are class-weighted; exact weights and focal parameters are not stated.
  • rare-expression oversampling rate = not specified
    Training oversamples rare expressions; the exact factor is not reported.
assumptions (7)
  • domain assumption The three tasks are conditionally independent given the latent z (Eq. 3).
    If false, the factorization p(y|z)=prod p(y_t|z) is misspecified and Eq. (5) is not an exact ELBO.
  • domain assumption Partial-label masks are ignorable (missing at random given the frame features).
    The marginalization in Eq. (4)-(5) integrates out unobserved labels; if the mask pattern encoded label values, estimates would be biased.
  • domain assumption Standard-normal prior and linear Gaussian/categorical/Bernoulli decoders define the model.
    Specific choice of variational family; a weak posterior could limit the claimed marginalization benefit, as their own bottleneck ablation shows.
  • domain assumption Both backbones' pretraining data exclude s-Aff-Wild2.
    Benchmark compliance rests on this provenance claim for FSFM and MAE-Face (Sec. 4).
  • standard math Deep generative semi-supervised objective (Kingma et al., 2014) is correct and applicable.
    Basis for the ELBO in Eq. (5); cited and used without independent formal proof.
  • standard math Official PMTL metric is additive and macro-averaged (Eq. 6).
    Justifies per-task specialist assembly and cross-backbone averaging; defined by the benchmark protocol.
  • domain assumption Cached backbone features are sufficient; the backbone is frozen during temporal/latent training.
    The reported pipeline probes a frozen-feature representation; if joint backbone fine-tuning were required, the mechanism claim would not follow.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Shared Latent for Partially-Labeled Multi-Task Facial Affect Recognition." pith.science (2026). https://pith.science/paper/V3NMYD2G

@misc{pith2026260716285,
  author       = {Pith},
  title        = {Pith review of: A Shared Latent for Partially-Labeled Multi-Task Facial Affect Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V3NMYD2G}},
  note         = {Machine review of arXiv:2607.16285}
}
read the original abstract

Facial affect in the wild is naturally multi-task: valence-arousal, discrete expressions, and facial action units describe the same face. Yet real corpora annotate these tasks only partially and unevenly, so most systems mask the missing labels or impute pseudo-labels and forgo the cross-task signal. We instead cast partially-labeled multi-task learning as marginalization over a shared affect latent: one variational bottleneck mediates all three task decoders, so a frame annotated for one task shapes the representation the others use, and the masked objective reappears as the reconstruction term of an evidence lower bound. On s-Aff-Wild2, where only 37% of frames carry all three labels, the classes are severely imbalanced, and pretraining on the source data is disallowed, we isolate where this coupling acts. On a single backbone it lifts expression macro-F1 from 0.403 for a dedicated specialist to 0.446, which the masked-loss model does not reach; a second, near-peer backbone with decorrelated errors then breaks an action-unit ceiling that external action-unit data could not, while valence-arousal stays within noise. Every gain is disciplined by a matched-control negative; together these controls indicate that the rare-class failure is representational, not a matter of loss shaping. As each task's source is chosen on the evaluation split, we report the assembled result, a combined multi-task score of 1.679 on validation, as an in-sample endpoint and rest our conclusions on the controlled comparisons; a small, regime-dependent transfer of the expression advantage to AffectNet and RAF-DB is presented as exploratory rather than conclusive.

Figures

Figures reproduced from arXiv: 2607.16285 by the authors.

Figure 1
Figure 1. System overview. Two AffectNet-initialized ViT-B/16 backbones (FSFM, a face self-supervised model, and MAE-Face) form two [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 3
Figure 3. Validation PMTL by backbone, grouped by pretraining family. Temporal context is the largest driver. The single largest step in Tab. 2 is the learned temporal model, which raises PMTL from 1.276 to 1.509±0.016. To attribute this gain we train the same architecture with window size one, which removes all temporal context while keeping the recurrent network and its freshly learned heads; it reaches only 1.007, below bo… view at source ↗
Figure 4
Figure 4. Validation PMTL vs temporal window length (single seed, fixed stride; absolute scores sit below the multi-seed 1.509). Dashed: per-frame model; dotted: window-one control. per-task assembly takes the same form as the strongest reported validation systems, which also combine the best model per sub-task (Liu et al., 2025). Diversity helps at the output. Combining models by av￾eraging their predictions is standard prac… view at source ↗
Figures from the paper (2 more)
Figure 8
Figure 8. Figure 8: Row-normalized expression confusion (validation): base 1 [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 10
Figure 10. Figure 10: Latent KL weight β (validation; static prior, z of size 96). Expression macro-F1 rises with β above the masked-loss baseline (dashed); PMTL peaks at β = 0.05; annealing (stars) is best. Filled markers six-seed, open three-seed. falls. The overall score therefore peaks…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Strength-Parity Ensembling with Parameter-Isolated Experts for Multi-Task Affect Recognition

    cs.CV 2026-07 conditional novelty 5.0 of 10

    Parameter-isolated LoRA experts on one face backbone stay decorrelated (0.91 vs 0.98 for full fine-tuning) and improve an ABAW affect ensemble from 1.6669 to 1.6949–1.7259 validation score.

Reference graph

Works this paper leans on

54 extracted references · 11 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Computer Vision -- ECCV 2022 Workshops , series=

    Affective Behavior Analysis using Action Unit Relation Graph and Multi-task Cross Attention , author=. Computer Vision -- ECCV 2022 Workshops , series=. 2023 , publisher=

    2022

  2. [2]

    Affective Behavior Analysis using Task-adaptive and

    Li, Xiaodong and Du, Wenchao and Yang, Hongyu , booktitle=. Affective Behavior Analysis using Task-adaptive and. 2025 , publisher=

    2025

  3. [3]

    Computer Vision -- ECCV 2024 Workshops , series=

    Affective Behaviour Analysis via Progressive Learning , author=. Computer Vision -- ECCV 2024 Workshops , series=. 2025 , publisher=

    2024

  4. [4]

    IEEE/CVF International Conference on Computer Vision (ICCV) , year=

    Asymmetric Loss for Multi-Label Classification , author=. IEEE/CVF International Conference on Computer Vision (ICCV) , year=

  5. [5]

    2023 , doi=

    Cai, Zhixi and Ghosh, Shreya and Stefanov, Kalin and Dhall, Abhinav and Cai, Jianfei and Rezatofighi, Hamid and Haffari, Reza and Hayat, Munawar , booktitle=. 2023 , doi=

    2023

  6. [6]

    NeurIPS , year=

    Learning Imbalanced Datasets with Label-Distribution-Aware Margin Loss , author=. NeurIPS , year=

  7. [7]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Multi-task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  8. [8]

    International Journal of Computer Vision , pages=

    Deep affect prediction in-the-wild: Aff-wild database and challenge, deep architectures, and beyond , author=. International Journal of Computer Vision , pages=. 2019 , publisher=

    2019

Show all 54 references
  1. [9]

    arXiv preprint arXiv:1910.04855 , year=

    Expression, Affect, Action Unit Recognition: Aff-Wild2, Multi-Task Learning and ArcFace , author=. arXiv preprint arXiv:1910.04855 , year=

    1910 arXiv

  2. [10]

    arXiv preprint arXiv:1910.11111 , year=

    Face Behavior a la carte: Expressions, Affect and Action Units in a Single Network , author=. arXiv preprint arXiv:1910.11111 , year=

    1910 arXiv

  3. [11]

    2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020) , pages=

    Analysing Affective Behavior in the First ABAW 2020 Competition , author=. 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020) , pages=. 2020 , doi=

    2020

  4. [12]

    arXiv preprint arXiv:2103.15792 , year=

    Affect Analysis in-the-wild: Valence-Arousal, Expressions, Action Units and a Unified Framework , author=. arXiv preprint arXiv:2103.15792 , year=

    arXiv

  5. [13]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Analysing affective behavior in the second abaw2 competition , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=. 2021 , doi=

    2021

  6. [14]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Abaw: Valence-arousal estimation, expression recognition, action unit detection & multi-task learning challenges , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=. 2022 , doi=

    2022

  7. [15]

    European Conference on Computer Vision , pages=

    Abaw: Learning from synthetic data & multi-task learning challenges , author=. European Conference on Computer Vision , pages=. 2023 , organization=

    2023

  8. [16]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Abaw: Valence-arousal estimation, expression recognition, action unit detection & emotional reaction intensity estimation challenges , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=. 2023 , doi=

    2023

  9. [17]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    The 6th affective behavior analysis in-the-wild (abaw) competition , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=. 2024 , doi=

    2024

  10. [18]

    European Conference on Computer Vision , pages=

    7th abaw competition: Multi-task learning and compound expression recognition , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  11. [19]

    arXiv preprint arXiv:2409.17717 , year=

    Behaviour4all: in-the-wild facial behaviour analysis toolkit , author=. arXiv preprint arXiv:2409.17717 , year=

    arXiv

  12. [20]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Distribution matching for multi-task learning of classification tasks: a large-scale study on faces & beyond , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2024 , doi=

    2024

  13. [21]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Advancements in Affective and Behavior Analysis: The 8th ABAW Workshop and Competition , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=. 2025 , doi=

    2025

  14. [22]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    From emotions to violence: Multimodal fine-grained behavior analysis at the 9th abaw , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=. 2025 , doi=

    2025

  15. [23]

    2026 , eprint=

    From Affect to Complex Behavior: Advancing Multimodal Human-Centered AI at the 10th ABAW Workshop & Competition , author=. 2026 , eprint=

    2026

  16. [24]

    ICLR , year=

    Long-tail learning via logit adjustment , author=. ICLR , year=

  17. [25]

    , journal=

    Savchenko, Andrey V. , journal=. 2024 , url=

    2024

  18. [26]

    arXiv preprint arXiv:2508.10104 , year=

    Sim. arXiv preprint arXiv:2508.10104 , year=

    arXiv

  19. [27]

    arXiv preprint arXiv:2502.14786 , year=

    Tschannen, Michael and Gritsenko, Alexey and Wang, Xiao and Naeem, Muhammad Ferjad and Alabdulmohsin, Ibrahim and Parthasarathy, Nikhil and Evans, Talfan and Beyer, Lucas and Xia, Ye and Mustafa, Basil and H. arXiv preprint arXiv:2502.14786 , year=

    arXiv

  20. [28]

    2025 , doi=

    Wang, Gaojian and Lin, Feng and Wu, Tong and Liu, Zhenguang and Ba, Zhongjie and Ren, Kui , booktitle=. 2025 , doi=

    2025

  21. [29]

    NeurIPS , year=

    Gradient Surgery for Multi-Task Learning , author=. NeurIPS , year=

  22. [30]

    Computer Vision and Pattern Recognition Workshops (CVPRW), 2017 IEEE Conference on , pages=

    Aff-wild: Valence and arousal `in-the-wild' challenge , author=. Computer Vision and Pattern Recognition Workshops (CVPRW), 2017 IEEE Conference on , pages=. 2017 , organization=

    2017

  23. [31]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    General Facial Representation Learning in a Visual-Linguistic Manner , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  24. [32]

    2026 , eprint=

    The Circumplex Degeneracy Behind the Rare-Class Limit in Affect Recognition , author=. 2026 , eprint=

    2026

  25. [33]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Semi-Supervised Learning with Deep Generative Models , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  26. [34]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , year=

    Multi-modal Facial Affective Analysis based on Masked Autoencoder , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , year=

  27. [35]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    A multi-task mean teacher for semi-supervised facial affective behavior analysis , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  28. [36]

    Gera, Darshan and Kumar, Badveeti Naveen Siva and Kumar, Bobbili Veerendra Raj and Balasubramanian, S. , year=. 2207.09012 , journal=

    arXiv

  29. [37]

    Pattern Recognition Letters , volume=

    Facial expression recognition based on a multi-task global-local network , author=. Pattern Recognition Letters , volume=. 2020 , publisher=

    2020

  30. [38]

    and Canavan, Shaun , journal=

    Hinduja, Saurabh and Nourivandi, Tara and Cohn, Jeffrey F. and Canavan, Shaun , journal=. Time to retire. 2024 , publisher=

    2024

  31. [39]

    , journal=

    Mollahosseini, Ali and Hasani, Behzad and Mahoor, Mohammad H. , journal=. 2019 , publisher=

    2019

  32. [40]

    IEEE Transactions on Image Processing , volume=

    Reliable Crowdsourcing and Deep Locality-Preserving Learning for Unconstrained Facial Expression Recognition , author=. IEEE Transactions on Image Processing , volume=. 2019 , publisher=

    2019

  33. [41]

    IEEE Transactions on Affective Computing , volume=

    Deep Facial Expression Recognition: A Survey , author=. IEEE Transactions on Affective Computing , volume=. 2022 , publisher=

    2022

  34. [42]

    and Canavan, Shaun and Reale, Michael and Horowitz, Andy and Liu, Peng and Girard, Jeffrey M

    Zhang, Xing and Yin, Lijun and Cohn, Jeffrey F. and Canavan, Shaun and Reale, Michael and Horowitz, Andy and Liu, Peng and Girard, Jeffrey M. , journal=. 2014 , publisher=

    2014

  35. [43]

    Savchenko, Andrey V. , year=. 2207.09508 , journal=

    arXiv

  36. [44]

    Machine Learning , volume=

    Multitask Learning , author=. Machine Learning , volume=. 1997 , publisher=

    1997

  37. [45]

    arXiv preprint arXiv:1706.05098 , doi=

    An Overview of Multi-Task Learning in Deep Neural Networks , author=. arXiv preprint arXiv:1706.05098 , doi=. 2017 , eprint=

    2017 arXiv

  38. [46]

    IEEE Conf

    Cross-Stitch Networks for Multi-Task Learning , author=. IEEE Conf. Computer Vision and Pattern Recognition (CVPR) , year=

  39. [47]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Multi-Task Learning as Multi-Objective Optimization , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  40. [48]

    arXiv preprint arXiv:1312.6114 , year=

    Auto-encoding variational bayes , author=. arXiv preprint arXiv:1312.6114 , year=

    arXiv

  41. [49]

    International Conference on Learning Representations (ICLR) , year=

    Higgins, Irina and Matthey, Lo. International Conference on Learning Representations (ICLR) , year=

  42. [50]

    Learning a Deep

    Durand, Thibaut and Mehrasa, Nazanin and Mori, Greg , booktitle=. Learning a Deep. 2019 , doi=

    2019

  43. [51]

    2015 IEEE International Conference on Computer Vision (ICCV) , pages=

    Multi-conditional Latent Variable Model for Joint Facial Action Unit Detection , author=. 2015 IEEE International Conference on Computer Vision (ICCV) , pages=. 2015 , doi=

    2015

  44. [52]

    2015 11th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG) , pages=

    Variable-state Latent Conditional Random Fields for Facial Expression Recognition and Action Unit Detection , author=. 2015 11th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG) , pages=. 2015 , doi=

    2015

  45. [53]

    2017 IEEE International Conference on Computer Vision (ICCV) , pages=

    Tran, Dieu Linh and Walecki, Robert and Rudovic, Ognjen and Eleftheriadis, Stefanos and Schuller, Bj. 2017 IEEE International Conference on Computer Vision (ICCV) , pages=. 2017 , doi=

    2017

  46. [54]

    arXiv preprint arXiv:1811.07771 , year=

    A multi-task learning & generation framework: Valence-arousal, action units & primary expressions , author=. arXiv preprint arXiv:1811.07771 , year=

    arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.