Pith. sign in

REVIEW 4 major objections 6 minor 40 references

Multi-task Learning For Joint Action and Gesture Recognition

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Multi-task learning for action and gesture recognition outperforms single-task training on nearly every tested configuration.

desk verdict Useful but under-specified empirical study; the core MTL-for-joint-action-gesture finding probably holds in qualified form, but the manual task splits and overbroad claims need fixing before I'd trust it. read the letter →

arxiv 2505.17867 v1 pith:MEMRK7J6 submitted 2025-05-23 cs.CV

Konstantinos Spathis , Nikolaos Kardaris , Petros Maragos This is my paper · ORCID
classification cs.CV
keywords multi-tasklearningactionrecognitiongesturehardparametersharingcross-stitchnetworkslearnedweightResNet-3Dhuman-robotinteraction
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 sets out to show that action and gesture recognition, usually trained as two separate classifiers, are better learned by one network that trains both at once. The authors adapt a ResNet-3D backbone into multi-task variants using three weight-sharing schemes and three loss-combination rules, then compare them against single-task baselines on four dataset combinations built from UCF-101, NTU-RGB+D, IsoGD, and NVGesture. Nearly all joint configurations match or beat the single-task models, and the gain is clearest for soft and learned weight sharing; hard sharing fails only on the gesture-dominated set. If the result holds, a practical system such as a robot could run one compact model instead of two separate ones and get better accuracy on both actions and gestures.

What carries the argument

The machinery has three moving parts. First, a ResNet-3D backbone pretrained on Kinetics-400, with its last two residual layers left trainable, supplies the shared spatio-temporal features. Second, weight sharing is varied across three regimes: hard parameter sharing (one shared trunk with duplicated task heads), soft parameter sharing through cross-stitch units that learn linear combinations of the two tasks' activation maps (parameterized by a sharing hyperparameter s), and learned weight sharing, which searches over layer-to-task assignments by alternating a natural evolution strategy for the assignment probabilities with SGD for the weights. Third, the multi-task loss is computed either as a plain average, as a dynamic weight average based on each task's recent loss rate, or as an uncertainty/automatic weighting with learned per-task noise. These components let the model exchange information at chosen depths and strengths, which is the mechanism the paper credits for the improved representations.

What would settle it

Re-run the joint-training comparison with the NTU action/gesture split made by multiple independent annotators, or without merging the overlapping gesture classes from NTU-RGB+D with IsoGD and NVGesture; if the SPS and LWS advantage over single-task baselines disappears or reverses, the claimed synergy is an artifact of the chosen class boundaries.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that actions and gestures are close enough as spatio-temporal tasks that a single architecture trained on both outperforms the same architecture trained on either alone. The evidence comes from four multi-task sets: UCF-101 with IsoGD, UCF-101 with NVGesture, NTU-RGB+D actions with IsoGD gestures, and NTU-RGB+D actions with NVGesture gestures. Across these, almost every SPS and LWS configuration beats both the ResNet-3D-18 and ResNet-3D-34 single-task baselines, and SPS with cross-stitch sharing at s=0.8 gives the best cumulative accuracies. The paper also reports that the best loss weighting depends on the data: automatic uncertainty weighting pairs well with SPS, while average and dynamic weight average are competitive with SPS and LWS. A notable efficiency result is that multi-task networks with a shallow ResNet-3D-18 backbone can outperform a deeper ResNet-3D-34 single-task network with similar trainable parameters.

Load-bearing premise

The load-bearing premise is the manual separation of NTU-RGB+D classes into action and gesture sets and the merging of overlapping gesture classes across datasets; if that split is imperfect, the reported multi-task advantage could be an artifact of that particular class partition rather than a genuine synergy between action and gesture recognition.

Editorial extensions

If this is right

  • A single joint model can replace separate action and gesture classifiers in an HRI system, cutting parameter count and inference cost while matching or exceeding accuracy.
  • Choosing the sharing mechanism matters more than choosing the loss rule: complete sharing (HPS) can hurt when gestures dominate, while adaptive sharing (SPS and LWS) is safer across all four sets.
  • The multi-task advantage is not tied to a bigger model: ResNet-3D-18 MTL models beat ResNet-3D-34 single-task models with comparable parameters.
  • Data balance shifts the conclusion: when actions dominate the joint set, even simple HPS helps; when gestures dominate, only flexible sharing methods retain the advantage.

Reading between the lines

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

  • Editorial extension: the authors flag in the conclusion that their models need to know the task of each input; a task-agnostic network would remove that requirement, and the same shared-feature argument suggests it should be trainable from the same data.
  • Editorial extension: because the NTU action/gesture split is manual and overlapping gesture classes are merged across datasets, the measured gain might depend on that partition; re-running with several independent annotators would show how stable the effect is.
  • Editorial extension: if the shared representations are genuinely general, joint training on UCF-101 and IsoGD should transfer to a small unseen gesture or action dataset with little fine-tuning; this is testable and goes beyond the paper's fixed four sets.
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

4 major / 6 minor

Summary. The paper proposes using multi-task learning (MTL) for joint action and gesture recognition, comparing three weight-sharing methods (hard parameter sharing, soft parameter sharing via Cross-Stitch units, and learned weight sharing) and three multi-task loss calculation methods (average, dynamic weight average, and automatic uncertainty weighting). The experiments combine UCF-101 and NTU-RGB+D for actions with IsoGD and NVGesture for gestures into four multi-task sets, including a manual split of NTU-RGB+D into action and gesture subsets. The authors report per-task and cumulative accuracies for ResNet-3D backbones and claim that MTL generally outperforms single-task learning (STL), yielding more efficient and robust representations.

Significance. If the central claim held, the paper would provide a useful practical demonstration that a single joint network can replace separate action and gesture classifiers, with potential efficiency and generalization benefits for human-robot interaction systems. The paper is also valuable as a systematic comparison of existing MTL components on a new task pair, and it includes a parameter-matched comparison with a deeper STL baseline. However, the claim is purely empirical, and the current evidence is weakened by an unreleased and unvalidated manual dataset split, missing statistical uncertainty, and several results that contradict the stated conclusions. These issues are fixable, but they currently prevent the paper from supporting its abstract-level claim.

major comments (4)
  1. [Section 4.1] The manual split of NTU-RGB+D into NTU AR and NTU GR, and the merging of 'common' gesture classes across NTU GR, IsoGD, and NVGesture, are not specified, validated, or released. The text states 'The separation in action and gesture sets is done manually' and that overlapping classes are handled 'by merging the samples of the different datasets as the same class,' but it does not list which of the 120 NTU classes go into each set, how overlap was determined, or how merged labels were created. Because every Set-3 and Set-4 result depends on this partition, the reported MTL advantage could be an artifact of a particular label assignment rather than a genuine action-gesture synergy. The authors should provide a complete class-level mapping, justify the criteria, release it, and ideally test robustness to alternative splits.
  2. [Section 4.3, Tables 1 and 2] The claim that 'All weight sharing methods outperform their STL variants, except for the case of the HPS models trained on Set-1' is contradicted by the tables. In Table 1, HPS average on Set-1 gives 54.16% versus 58.26% STL for actions and 20.49% versus 25.67% for gestures; HPS automatic on Set-1 is also below STL for gestures. In Table 2, HPS automatic on Set-3 gives 51.55% cumulative accuracy, below both STL r3d18 (55.41%) and STL r3d34 (59.62%). Many other rows improve one task while degrading the other. The abstract's statement that joint training achieves 'better performance for both tasks' is therefore too strong. The conclusions should be reworded to reflect that MTL helps in some configurations and on some tasks, and the paper should identify the conditions under which it helps and hurts.
  3. [Section 4.2] The single-task training protocol is underspecified, and this is load-bearing for the MTL-versus-STL comparison. The paper says both single-task and multi-task networks are trained for 20,000 iterations with batch size 8 and Adam, but it does not state whether each STL baseline is trained separately on each action and gesture dataset with the same number of iterations, how the 'same number of samples' is enforced when MTL batches contain samples from both tasks, or which data augmentation and evaluation protocol are used. It also does not report results over multiple runs or with different random seeds, which is important because many of the reported differences between MTL and STL are small (e.g., a few percentage points). The authors should specify the exact STL protocol and provide mean and standard deviation over at least three runs.
  4. [Section 4.3, Tables 1 and 2] The cumulative accuracy metric, defined as 'the percentage of all correct predictions to the number of total predictions across all tasks,' can hide per-task degradation. For example, a model that improves on the majority task while losing on the minority task can still show a higher cumulative accuracy, which does not support the abstract's claim of better performance for both tasks. The paper should either report the per-task accuracy as the primary metric or supplement the cumulative accuracy with a clear per-task breakdown and a discussion of task-wise trade-offs. Additionally, Table 2's numbers do not appear to be directly derivable from the per-task accuracies in Table 1 without the exact test-set sizes, so the weighting should be stated explicitly.
minor comments (6)
  1. [Section 3, around Eq. (6)] The parameter update for the weight optimization step is written as 'θ−ηθ∇π J(θ,π)' but should use the gradient with respect to θ, not π.
  2. [Section 3, loss formulas] The uncertainty loss and automatic loss formulas have unbalanced parentheses: the first writes 'Σ 1/(2σ^2)L_i + log(σ)' and the second ends with 'log(1 + σ^2))'. These should be corrected for clarity.
  3. [Table 1] The table header 'Set-1 (%)', 'Set-2 (%)', etc., is ambiguous because each set has two numbers per row. The columns should be explicitly labeled as the action accuracy and the gesture accuracy for each set.
  4. [Section 4.2] The sentence 'We left its last 2 Residual Layers (last 8 convolutional layers) trainable' leaves unclear whether batch-normalization statistics in the frozen layers are updated or kept fixed; this detail matters for reproducibility.
  5. [Section 4.3] The choice of the soft-sharing coefficient s is only tested for s=0.2 and s=0.8, and the DWA temperature T is never given. The dependence of the conclusions on these hyperparameters should be discussed.
  6. [Section 2.1] The reference to 'Hu and Sign' for UniT should be corrected to 'Hu and Singh'.

Circularity Check

0 steps flagged · score 0.0 of 10

Empirical MTL-vs-STL comparison with no circular derivation.

full rationale

The paper's central claim is an empirical observation: Tables 1 and 2 report test accuracies of single-task versus multi-task models on four constructed action/gesture sets. No equation is fitted to the reported outcomes, and no predicted quantity is defined in terms of the quantity it is supposed to explain. The MTL methods used (Cross-Stitch, LWS) are taken from external prior work, and the STL baselines are trained and evaluated under the same protocols rather than being algebraically derived from the reported results. The manual NTU split into action and gesture sets is a data-labeling choice that delimits the claim, but it does not make the MTL advantage true by construction; the advantage is contingent on measured performance. The author self-citations that appear in the related-work survey (e.g., [10], [21]) are illustrative examples and are not load-bearing for the central argument. No uniqueness theorem, ansatz, or fitted parameter is imported from the authors' own prior work to force the conclusion. The paper is therefore self-contained as an experimental study, and no circular step can be exhibited.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The empirical claim rests on task definitions and training hyperparameters rather than on a derivation. The key assumptions are the manual NTU split and class merging, which cannot be verified from the paper. No new entities are introduced.

free parameters (4)
  • SPS sharing coefficient s = 0.2, 0.8
    Controls the degree of information sharing in cross-stitch units in the SPS method. Two fixed values are evaluated; the optimal choice may depend on dataset composition.
  • DWA temperature T = not reported
    Temperature scaling in the dynamic weight average loss. The paper does not give the value used.
  • Training iterations = 20000
    Empirically fixed as the point where all evaluated models have reached convergence, as stated in Section 4.2.
  • LWS population sizes = not reported
    Population sizes for the NES assignment step and SGD weight step in LWS are not specified, despite being central to the method.
assumptions (3)
  • domain assumption The manual split of NTU-RGB+D into action and gesture classes is a valid representation of the two tasks.
    Section 4.1 states 'The separation in action and gesture sets is done manually.' The entire comparison depends on this definition.
  • domain assumption Merging common gesture classes from different datasets preserves label semantics.
    Section 4.1: 'When combining the sets... some of the classes are common between the datasets. We handle this issue by merging the samples of the different datasets as the same class.' No evidence is given that the gestures are visually equivalent.
  • domain assumption Pre-trained Kinetics-400 weights transfer equally to action and gesture video tasks.
    Section 4.2 states all models are initialized with Kinetics-400 pretrained weights; this is standard but assumes equal relevance to both tasks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-task Learning For Joint Action and Gesture Recognition." pith.science (2026). https://pith.science/paper/MEMRK7J6

@misc{pith2026250517867,
  author       = {Pith},
  title        = {Pith review of: Multi-task Learning For Joint Action and Gesture Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MEMRK7J6}},
  note         = {Machine review of arXiv:2505.17867}
}
read the original abstract

In practical applications, computer vision tasks often need to be addressed simultaneously. Multitask learning typically achieves this by jointly training a single deep neural network to learn shared representations, providing efficiency and improving generalization. Although action and gesture recognition are closely related tasks, since they focus on body and hand movements, current state-of-the-art methods handle them separately. In this paper, we show that employing a multi-task learning paradigm for action and gesture recognition results in more efficient, robust and generalizable visual representations, by leveraging the synergies between these tasks. Extensive experiments on multiple action and gesture datasets demonstrate that handling actions and gestures in a single architecture can achieve better performance for both tasks in comparison to their single-task learning variants.

Figures

Figures reproduced from arXiv: 2505.17867 by the authors.

Figure 1
Figure 1. An instance of a hard parameter sharing model for three tasks. The first layers of the model (gray [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Cross Stitch Units applied on two task specific CNNs. The task specific models, illustrated with [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. (a) Representation of an LWS architecture for two tasks. Layers between task-specific networks [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 2
Figure 2. Figure 2: HPS and SPS methods both share information across all MTL layers of the [PITH_FULL_IMAGE:figures/full_fig_p009_2.png]
Figure 4
Figure 4. Figure 4: Samples of action and gesture benchmark datasets. From top to bottom row: UCF-101, NTU [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Classes and samples distribution across di [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 38 canonical work pages

  1. [1]

    Caruana, Multitask learning, Machine Learning 28 (1997) 41–75

    R. Caruana, Multitask learning, Machine Learning 28 (1997) 41–75

    work page 1997

  2. [2]

    Standley, A

    T. Standley, A. Zamir, D. Chen, L. Guibas, J. Malik, S. Savarese, Which tasks should be learned together in multi-task learning?, in: Proc. of the International Conference on Machine Learning (ICML), 2020

    work page 2020

  3. [3]

    X. Liu, J. Gao, X. He, L. Deng, K. Duh, Y .-y. Wang, Representation learning us- ing multi-task deep neural networks for semantic classification and information retrieval, in: Proc. of the Conference of North American Chapter of the Associa- tion for Computational Linguistics (NAACL), 2015

    work page 2015

  4. [4]

    Taghavi, R

    P. Taghavi, R. Langari, G. Pandey, Swinmtl: A shared architecture for simultane- ous depth estimation and semantic segmentation from monocular camera images, in: Proc. of IEEE International Conference on Intelligent Robots and Systems (IROS), 2024

    work page 2024

  5. [5]

    Sener, V

    O. Sener, V . Koltun, Multi-task learning as multi-objective optimization, Ad- vances in Neural Information Processing Systems 31 (2018)

    work page 2018

  6. [6]

    Kokkinos, Ubernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory, in: Proc

    I. Kokkinos, Ubernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory, in: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017

    work page 2017

  7. [7]

    Bhattacharjee, T

    D. Bhattacharjee, T. Zhang, S. S ¨usstrunk, M. Salzmann, Mult: An end-to-end multitask learning transformer, in: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    work page 2022

  8. [8]

    R. Hu, A. Singh, Unit: Multimodal multitask learning with a unified transformer, in: Proc. of the IEEE International Conference on Computer Vision (ICCV), 2021

    work page 2021

Show all 40 references
  1. [9]

    X. Shi, Z. Chen, H. Wang, D.-Y . Yeung, W.-K. Wong, W.-c. Woo, Convolutional lstm network: A machine learning approach for precipitation nowcasting, Ad- vances in Neural Information Processing Systems 28 (2015) 802–810. 22

    2015

  2. [10]

    Rodomagoulakis, N

    I. Rodomagoulakis, N. Kardaris, V . Pitsikalis, E. Mavroudi, A. Katsamanis, A. Tsiami, P. Maragos, Multimodal human action recognition in assistive human- robot interaction, in: Proc. of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2016

    2016

  3. [11]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al., An image is worth 16x16 words: Transformers for image recognition at scale, in: Proc. of the International Conference on Learning Represent...

    2020

  4. [12]

    H. Lu, H. Jian, R. Poppe, A. A. Salah, Enhancing video transformers for action understanding with vlm-aided training, arXiv preprint arXiv:2403.16128 (2024). doi:10.48550/arXiv.2403.16128

    2024 doi arXiv

  5. [13]

    J. Li, D. Li, S. Savarese, S. Hoi, Blip-2: Bootstrapping language-image pre- training with frozen image encoders and large language models, in: Proc. of the International Conference on Machine Learning (ICML), 2023

    2023

  6. [14]

    L. Wang, B. Huang, Z. Zhao, Z. Tong, Y . He, Y . Wang, Y . Wang, Y . Qiao, Video- mae v2: Scaling video masked autoencoders with dual masking, in: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  7. [15]

    Goyal, S

    R. Goyal, S. Ebrahimi Kahou, V . Michalski, J. Materzynska, S. Westphal, H. Kim, V . Haenel, I. Fruend, P. Yianilos, M. Mueller-Freitag, et al., The” something something” video database for learning and evaluating visual common sense, in: Proc. of the IEEE International Confer...

    2017

  8. [16]

    Y . Ma, B. Zhou, R. Wang, P. Wang, Multi-stage factorized spatio-temporal rep- resentation for rgb-d action and gesture recognition, in: Proc. of the 31st ACM International Conference on Multimedia (ACM MM), 2023

    2023

  9. [17]

    Kopuklu, N

    O. Kopuklu, N. Kose, G. Rigoll, Motion fused frames: Data level fusion strategy for hand gesture recognition, in: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2018. 23

    2018

  10. [18]

    B. Zhou, P. Wang, J. Wan, Y . Liang, F. Wang, D. Zhang, Z. Lei, H. Li, R. Jin, Decoupling and recoupling spatiotemporal representation for rgb-d-based motion recognition, in: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  11. [19]

    M. Garg, D. Ghosh, P. M. Pradhan, Gestformer: Multiscale wavelet pooling trans- former network for dynamic hand gesture recognition, in: Proc. of the IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  12. [20]

    D. C. Luvizon, D. Picard, H. Tabia, Multi-task deep learning for real-time 3d hu- man pose estimation and action recognition, IEEE Transactions on Pattern Anal- ysis and Machine Intelligence 43 (8) (2020) 2752–2764

    2020

  13. [21]

    Koutras, P

    P. Koutras, P. Maragos, Susinet: See, understand and summarize it, in: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2019

    2019

  14. [22]

    Simonyan, A

    K. Simonyan, A. Zisserman, Two-stream convolutional networks for action recognition in videos, Advances in Neural Information Processing Systems 27 (2014) 568–576

    2014

  15. [23]

    D. Fan, H. Lu, S. Xu, S. Cao, Multi-task and multi-modal learning for rgb dy- namic gesture recognition, IEEE Sensors Journal 21 (23) (2021) 27026–27036

    2021

  16. [24]

    K. Hara, H. Kataoka, Y . Satoh, Learning spatio-temporal features with 3d residual networks for action recognition, in: Proc. of the IEEE International Conference on Computer Vision Workshops (ICCVW), 2017

    2017

  17. [25]

    Y . Yang, T. Hospedales, Trace norm regularised deep multi-task learning, in: Proc. of the International Conference on Learning Representations (ICLR), 2017

    2017

  18. [26]

    Misra, A

    I. Misra, A. Shrivastava, A. Gupta, M. Hebert, Cross-stitch networks for multi- task learning, in: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 24

    2016

  19. [27]

    Prellberg, O

    J. Prellberg, O. Kramer, Learned weight sharing for deep multi-task learning by natural evolution strategy and stochastic gradient descent, in: Proc. of the Inter- national Joint Conference on Neural Networks (IJCNN), 2020

    2020

  20. [28]

    Wierstra, T

    D. Wierstra, T. Schaul, T. Glasmachers, Y . Sun, J. Peters, J. Schmidhuber, Natural evolution strategies, The Journal of Machine Learning Research 15 (1) (2014) 949–980

    2014

  21. [29]

    Kendall, Y

    A. Kendall, Y . Gal, R. Cipolla, Multi-task learning using uncertainty to weigh losses for scene geometry and semantics, in: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  22. [30]

    Liebel, M

    L. Liebel, M. K ¨orner, Auxiliary tasks in multi-task learning, arXiv preprint arXiv:1805.06334 (2018). doi:10.48550/arXiv.1805.06334

    2018 doi arXiv

  23. [31]

    S. Liu, E. Johns, A. J. Davison, End-to-end multi-task learning with attention, in: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  24. [32]

    Soomro, A

    K. Soomro, A. R. Zamir, M. Shah, UCF101: A Dataset of 101 Human Action Classes From Videos in The Wild, Tech. Rep. CRCV-TR-12-01, Center for Re- search in Computer Vision (CRCV), November, 2012

    2012

  25. [33]

    J. Liu, A. Shahroudy, M. Perez, G. Wang, L.-Y . Duan, A. C. Kot, Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding, IEEE transactions on pattern analysis and machine intelligence 42 (10) (2019) 2684–2701

    2019

  26. [34]

    J. Wan, Y . Zhao, S. Zhou, I. Guyon, S. Escalera, S. Z. Li, Chalearn looking at people rgb-d isolated and continuous datasets for gesture recognition, in: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2016

    2016

  27. [35]

    Molchanov, X

    P. Molchanov, X. Yang, S. Gupta, K. Kim, S. Tyree, J. Kautz, Online detection and classification of dynamic hand gestures with recurrent 3d convolutional neu- ral network, in: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 25

    2016

  28. [36]

    Carreira, A

    J. Carreira, A. Zisserman, Quo vadis, action recognition? a new model and the ki- netics dataset, in: Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017

    2017

  29. [37]

    Mehmood, A

    T. Mehmood, A. E. Gerevini, A. Lavelli, I. Serina, Combining multi-task learning with transfer learning for biomedical named entity recognition, Procedia Com- puter Science 176 (2020) 848–857

    2020

  30. [38]

    B. Liu, X. Liu, X. Jin, P. Stone, Q. Liu, Conflict-averse gradient descent for multi- task learning, Advances in Neural Information Processing Systems 34 (2021) 18878–18890

    2021

  31. [39]

    L. Liu, Y . Li, Z. Kuang, J. Xue, Y . Chen, W. Yang, Q. Liao, W. Zhang, Towards impartial multi-task learning, in: Proc. of the International Conference on Learn- ing Representations (ICLR), 2021

    2021

  32. [40]

    Elich, L

    C. Elich, L. Kirchdorfer, J. M. K ¨ohler, L. Schott, Examining common paradigms in multi-task learning, in: Proc. of the DAGM German Conference on Pattern Recognition (GCPR), 2024. 26

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.