Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

DeepChest: Dynamic Gradient-Free Task Weighting for Effective Multi-Task Learning in Chest X-ray Classification

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read DeepChest claims a gradient-free rule that raises the loss weight of tasks below average accuracy and lowers the rest reaches 94.96% on ChestX-ray14, beating prior models while tripling training speed.

desk verdict The weighting heuristic is a reasonable cheap idea, but the headline 94.96% result is a metric-mismatch artifact: Table 1 compares published AUC baselines against DeepChest's thresholded accuracy. read the letter →

arxiv 2505.23595 v1 pith:5WZFQZMG submitted 2025-05-29 cs.CV cs.AI

classification cs.CVcs.AI
keywords multi-tasklearningdynamictaskweightingchestX-rayclassificationX-ray14gradient-freeoptimizationlossbalancingclassimbalancemedicalimaging
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

DeepChest argues that multi-task training for chest X-ray classification can be rebalanced by a rule that needs no gradient mathematics: each epoch, tasks whose training accuracy is below the average have their loss weight multiplied upward (bounded by a cap), and tasks above the average have their weight divided down. The authors claim this simple feedback loop, initialized from single-task accuracies, reaches 94.96% average accuracy on the 14-way ChestX-ray14 benchmark, above AG-CNN D-121's 87.54%, and lowers every task's loss relative to training the tasks separately. It also reports a per-epoch training time of 34 minutes versus 102 for PCGrad, making gradient-free weighting a practical alternative to gradient-based conflict resolution in medical imaging. If the claims hold, it suggests a fast, interpretable balancing rule can substitute for costly gradient manipulations in multi-label medical classification.

What carries the argument

The load-bearing object is the multiplicative dynamic-weighting rule defined by three scalars and a per-epoch comparison: initial weights from single-task accuracy, then for each task $w_t \leftarrow \min(w_t \cdot \alpha, w_{\max})$ if the task's training accuracy is below the average, and $w_t \leftarrow w_t / \beta$ if it is at or above the average, with $\alpha = 1.1$, $\beta = 1.05$. This rule converts an ordinary accuracy signal into a balanced objective $\sum_t w_t \mathcal{L}_t$ without ever computing gradients of the weighting scheme, which is what removes the memory and compute overhead associated with methods like PCGrad.

What would settle it

Train the same backbone on the same ChestX-ray14 split under identical preprocessing with three weight schemes—fixed equal weights, DeepChest's dynamic weights, and PCGrad—and measure both average accuracy and per-class AUC. The claim of a 7-point gain from weighting stands only if DeepChest beats fixed equal weights by a comparable margin and if the gain survives under AUC.

Watch

Extended reading notes

Core claim

The central discovery the paper asserts is that a performance-driven, gradient-free task-weighting rule can outperform both fixed-weight multi-task baselines and gradient-based methods on a large chest X-ray dataset. DeepChest initializes the weight of each task from its single-task accuracy, $w_t^{(0)} = 1.0 + (1.0 - A^{(STL,t)}) \cdot 0.5$, then at every epoch compares each task's training accuracy to the average; underperforming tasks get their weight multiplied by $\alpha = 1.1$ (capped at $w_{\max}$), and outperforming tasks get divided by $\beta = 1.05$. Applied to a shared-backbone network with 14 disease heads, this rule yields 94.96% average classification accuracy, a total $\Delta_m$ of $-0.44$ (all 14 tasks show lower loss under MTL than STL), and a threefold speedup relative to PCGrad. The authors frame this as evidence that balancing via relative task difficulty, rather than gradient surgery, is sufficient to mitigate negative transfer in multi-label chest X-ray classification.

Load-bearing premise

The headline 94.96% accuracy is reported as a single number without specifying the ChestX-ray14 split, number of runs, or standard deviation; the comparison to 87.54% assumes the same evaluation protocol as AG-CNN D-121, so part of the gain could come from differences in data split, preprocessing, or backbone rather than from the weighting rule itself.

Editorial extensions

If this is right

  • If the reported 94.96% average accuracy is reproducible, multi-label chest X-ray classifiers can be trained more cheaply than with PCGrad or other gradient-based MTL methods while still beating previously published accuracies.
  • The weighting rule gives researchers a simple, portable component: any multi-task architecture with per-task heads can adopt it by replacing fixed or learned weights with the accuracy-based update, requiring only the STL accuracies for initialization.
  • Because MTL loss is lower than STL loss for all 14 pathologies, the method implies that dynamic weighting actively counters negative transfer in imbalanced medical label sets.
  • The threefold training-time reduction suggests that gradient-free balancing can make MTL feasible on datasets and hardware where gradient-based balancing is too slow.

Reading between the lines

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

  • Editorial extension: replacing the raw-accuracy signal with per-class AUC or balanced accuracy would test whether the rule's benefit survives in a class-imbalanced metric; the paper reports accuracy only, so AUC behavior is unknown.
  • Editorial extension: because the update rewards any task that lies below the average, weights will tend to oscillate as tasks cross the mean during training; an ablation that logs weight trajectories or sweeps $\alpha$, $\beta$, and $w_{\max}$ would reveal whether this oscillation is neutral, helpful, or harmful.
  • Editorial extension: the rule should transfer to other multi-label, imbalanced domains—such as multi-organ radiology or multi-label document classification—since it only needs per-task training accuracy; the paper does not test these settings.
  • Editorial extension: a same-backbone, same-preprocessing ablation versus fixed equal weights and PCGrad would separate the weighting rule's contribution from the architecture and data-augmentation choices, since Table 1 compares models trained under different protocols.
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

5 major / 5 minor

Summary. The paper introduces DeepChest, a gradient-free dynamic task-weighting rule for multi-task learning, applied to multi-label chest X-ray classification on the ChestX-ray14 dataset. Task weights are initialized from single-task accuracies and updated each epoch by multiplying underperforming tasks by a factor alpha (capped at w_max) and dividing well-performing tasks by beta, based on whether each task's current training accuracy is below or above the average accuracy. The authors report an average accuracy of 94.96%, claim an improvement of about 7% over prior methods, report lower multi-task losses than single-task losses for all 14 diseases with a total Delta_m of -0.44, and report a threefold reduction in per-epoch training time relative to PCGrad. Code is publicly available.

Significance. The proposed weighting rule is simple, self-contained, and model-agnostic, and the public release of code is a strength. If validated under a common evaluation protocol, such a gradient-free approach would be a practically useful baseline for multi-task medical image classification. However, the central quantitative claim is not supported by the evidence: the baseline entries in Table 1 are recognizable as published AUC values (e.g., CheXNet's 0.8414 is the average AUC in Rajpurkar et al., 2017), while the DeepChest column is labeled as accuracy without any definition. The paper also omits the data split, run count, and implementation details, includes an uncontrolled placeholder-image substitution step, and provides no sensitivity analysis for its hyperparameters. The significance of the contribution is therefore not established as written.

major comments (5)
  1. [Section 4.4, Table 1]
  2. [Section 4.1 and Section 4.4]
  3. [Section 4.2, Handling Missing and Invalid Data]
  4. [Section 3.1.2, Algorithm 1]
  5. [Section 4.4, Tables 2 and 3]
minor comments (5)
  1. [Section 4.4]
  2. [Section 4.2, Equation (1)]
  3. [Section 4.2, Label Encoding]
  4. [Figures 4 and 5]
  5. [Limitations section]

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the task-weighting rule is self-contained and the reported gains are not definitionally entailed by the update equations.

full rationale

DeepChest's central derivation is not circular. The weighting rule is fully specified as explicit functions of observable training accuracies: w(0)_t = 1.0 + (1.0 - A_STL,t) * 0.5, and during training w_t = min(w_t * alpha, w_max) when A_t < A_avg, else w_t / beta. These equations determine how task weights evolve from STL accuracies; they do not encode the reported test-set numbers, and 94.96% is not a rearrangement of the weight-update formula. The constants alpha = 1.1, beta = 1.05, and the unspecified w_max are hand-picked rather than fitted to the target metric, so the 'fitted input called prediction' pattern does not apply, although the missing sensitivity analysis weakens robustness. No load-bearing self-citation or imported uniqueness theorem is present; Delta_m is taken from external Maninis et al. and the comparison value -0.22 from external Achituve et al. The genuine problem with the headline comparison is that Table 1 appears to list published AUC values (e.g., CheXNet 0.8414, AG-CNN D-121 0.8754) while DeepChest's 0.9496 is thresholded accuracy, which is a metric misalignment; that is a serious correctness and validity concern, not a circular derivation. The appended Limitations paragraph acknowledges that broader validation is still needed, reinforcing that the claims are empirical. Therefore no circular step is established, and the score is 0.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

DeepChest does not introduce new physical entities or a latent mechanism. It introduces a weighting rule, not an entity. The main ledger entries are the hyperparameters (alpha, beta, w_max) that are fixed without ablations, and the implicit assumption that training-accuracy comparisons produce good weights. The absence of a stated data split is a significant unstated experimental choice that influences every number in the paper.

free parameters (5)
  • alpha (weight increase factor) = 1.1 (stated as example)
    Controls how strongly underperforming tasks are boosted. No sensitivity analysis is provided, and the final results depend on this choice.
  • beta (weight decay factor) = 1.05 (stated as example)
    Controls how strongly well-performing tasks are reduced. No ablation or sensitivity analysis is provided.
  • w_max (maximum allowable weight) = not reported
    Caps the weight of underperforming tasks. The exact value is never stated in the paper, yet it modulates the loss and therefore the reported accuracies.
  • STL training budget = not reported
    The initialization relies on STL accuracies, but the number of epochs and setup used to obtain those accuracies are not reported, so the resulting weights are not reproducible.
  • MTL epoch schedule = not reported
    The dynamic update runs per epoch, but the total number of epochs and learning rate schedule are not stated, which affects the final accuracy.
assumptions (3)
  • domain assumption Training accuracy is a reliable proxy for task difficulty and for final test performance.
    The weighting rule increases weights for tasks with low training accuracy relative to the average. This assumes that low training accuracy signals a task that needs more emphasis, and that this emphasis will improve final generalization. The paper provides no evidence for this monotonic relationship in the chest X-ray setting.
  • domain assumption The ChestX-ray14 evaluation protocol used in prior reported numbers is comparable to the protocol used for DeepChest.
    Table 1 compares DeepChest to literature values without describing the split or evaluation protocol for the new model. The near-perfect scores on several diseases suggest the protocols may differ, which would invalidate the headline accuracy comparison.
  • domain assumption Downsampling 1024x1024 X-rays to 128x128 preserves the clinically relevant signal for all 14 diseases.
    The preprocessing section resizes images to 128x128. The paper does not assess whether fine-grained findings such as small nodules or pneumothorax remain detectable at this low resolution, yet the reported accuracies are treated as diagnostic performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeepChest: Dynamic Gradient-Free Task Weighting for Effective Multi-Task Learning in Chest X-ray Classification." pith.science (2026). https://pith.science/paper/5WZFQZMG

@misc{pith2026250523595,
  author       = {Pith},
  title        = {Pith review of: DeepChest: Dynamic Gradient-Free Task Weighting for Effective Multi-Task Learning in Chest X-ray Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5WZFQZMG}},
  note         = {Machine review of arXiv:2505.23595}
}
read the original abstract

While Multi-Task Learning (MTL) offers inherent advantages in complex domains such as medical imaging by enabling shared representation learning, effectively balancing task contributions remains a significant challenge. This paper addresses this critical issue by introducing DeepChest, a novel, computationally efficient and effective dynamic task-weighting framework specifically designed for multi-label chest X-ray (CXR) classification. Unlike existing heuristic or gradient-based methods that often incur substantial overhead, DeepChest leverages a performance-driven weighting mechanism based on effective analysis of task-specific loss trends. Given a network architecture (e.g., ResNet18), our model-agnostic approach adaptively adjusts task importance without requiring gradient access, thereby significantly reducing memory usage and achieving a threefold increase in training speed. It can be easily applied to improve various state-of-the-art methods. Extensive experiments on a large-scale CXR dataset demonstrate that DeepChest not only outperforms state-of-the-art MTL methods by 7% in overall accuracy but also yields substantial reductions in individual task losses, indicating improved generalization and effective mitigation of negative transfer. The efficiency and performance gains of DeepChest pave the way for more practical and robust deployment of deep learning in critical medical diagnostic applications. The code is publicly available at https://github.com/youssefkhalil320/DeepChest-MTL

Figures

Figures reproduced from arXiv: 2505.23595 by the authors.

Figure 1
Figure 1. General process of proposed model agnostic dynamic gradient-free task weighting MTL classification [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. General process of multi-task learning (MTL) training. The shared feature extraction backbone (e.g., VGG Net, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Distributions of the ChestX-ray14 dataset [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The accuracy of the proposed method for different [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The accuracy of DeepChest for different diseases [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

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. CURE: Curriculum-guided Multi-task Training for Reliable Anatomy Grounded Report Generation

    cs.CV 2026-01 reject novelty 5.0 of 10

    CURE's curriculum-guided multi-task training improves bounding-box grounding for chest X-ray report generation, but its claimed hallucination reduction is not confirmed by the paper's full evaluation.

Reference graph

Works this paper leans on

49 extracted references · 34 canonical work pages · cited by 1 Pith paper

  1. [1]

    Idan Achituve, Idit Diamant, Arnon Netzer, Gal Chechik, and Ethan Fetaya. 2024. Bayesian Uncertainty for Gradient Aggregation in Multi-Task Learning. arXiv preprint arXiv:2402.04005 (2024)

  2. [2]

    Saleh Albahli, Hafiz Tayyab Rauf, Abdulelah Algosaibi, and Valentina Emilia Balas. 2021. AI-driven deep CNN approach for multi-label pathology classification using chest X-Rays. PeerJ Computer Science 7 (2021), e495

  3. [3]

    BJ Bakker and TM Heskes. 2003. Task clustering and gating for bayesian multitask learning. (2003)

  4. [4]

    Ivo M Baltruschat, Hannes Nickisch, Michael Grass, Tobias Knopp, and Axel Saalbach. 2019. Comparison of deep learning approaches for multi-label chest X-ray classification. Scientific reports 9, 1 (2019), 6381

  5. [5]

    Hakan Bilen and Andrea Vedaldi. 2016. Integrated perception with recurrent multi-task neural networks. Advances in neural information processing systems 29 (2016)

  6. [6]

    Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. 2019. Learning imbalanced datasets with label-distribution-aware margin loss. Ad- vances in neural information processing systems 32 (2019)

  7. [7]

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

  8. [8]

    Kai Chen, Xuqi Wang, and Shanwen Zhang. 2022. Thorax disease classification based on pyramidal convolution shuffle attention neural network. IEEE Access 10 (2022), 85571–85581

Show all 49 references
  1. [9]

    Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. 2018. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning . PMLR, 794–803

  2. [10]

    Jean-Antoine Désidéri. 2012. Multiple-gradient descent algorithm (MGDA) for multiobjective optimization. Comptes Rendus Mathematique 350, 5-6 (2012), 313–318

  3. [11]

    Qingji Guan, Yaping Huang, Zhun Zhong, Zhedong Zheng, Liang Zheng, and Yi Yang. 2018. Diagnose like a radiologist: Attention guided convolutional neural network for thorax disease classification. arXiv preprint arXiv:1801.09927 (2018)

  4. [12]

    Sebastian Guendel, Sasa Grbic, Bogdan Georgescu, Siqi Liu, Andreas Maier, and Dorin Comaniciu. 2019. Learning to recognize abnormalities in chest x-rays with location-aware dense networks. In Progress in Pattern Recognition, Image Analysis, Computer Vision, and Applications: 2...

  5. [13]

    Mengjie Hu, Hezheng Lin, Zimeng Fan, Wenjie Gao, Lu Yang, Chun Liu, and Qing Song. 2020. Learning to recognize chest-Xray images faster and more efficiently based on multi-kernel depthwise convolution. IEEE Access 8 (2020), 37265–37274

  6. [14]

    Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al

  7. [15]

    Adrián Javaloy and Isabel Valera. 2021. Rotograd: Gradient homogenization in multitask learning. ICLR (2021)

  8. [16]

    Alex Kendall, Yarin Gal, and Roberto Cipolla. 2018. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition . 7482–7491

  9. [17]

    Eunji Kim, Siwon Kim, Minji Seo, and Sungroh Yoon. 2021. XProtoNet: diagnosis in chest radiography with global and local explanations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 15719–15728

  10. [18]

    I Kokkinos. 2017. Ubernet: Training universal cnn for low mid and high level vision with diverse datasets and limited memory. CVPR

  11. [19]

    Baijiong Lin, Feiyang Ye, and Yu Zhang. 2021. A closer look at loss weighting in multi-task learning. (2021)

  12. [20]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024)

  13. [21]

    Liyang Liu, Yi Li, Zhanghui Kuang, J Xue, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. 2021. Towards impartial multi-task learning. iclr

  14. [22]

    Shikun Liu, Edward Johns, and Andrew J Davison. 2019. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 1871–1880

  15. [23]

    Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Philip S Yu. 2017. Learn- ing multiple tasks with multilinear relationship networks. Advances in neural information processing systems 30 (2017)

  16. [24]

    Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. 2019. Attentive single-tasking of multiple tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 1851–1860

  17. [25]

    Yuren Mao, Zekai Wang, Weiwei Liu, Xuemin Lin, and Pengtao Xie. 2022. Metaweighting: learning to weight tasks in multi-task learning. In Findings of the Association for Computational Linguistics: ACL 2022 . 3436–3448

  18. [26]

    Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. 2016. Cross-stitch networks for multi-task learning. InProceedings of the IEEE conference on computer vision and pattern recognition . 3994–4003

  19. [27]

    Daniel A Moses. 2021. Deep learning applied to automatic disease detection using chest x-rays. Journal of Medical Imaging and Radiation Oncology 65, 5 (2021), 498–517

  20. [28]

    P Rajpurkar. 2017. CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning. ArXiv abs/1711 5225 (2017)

  21. [29]

    Sebastian Ruder. 2017. An overview of multi-task learning in deep neural net- works. arXiv preprint arXiv:1706.05098 (2017)

  22. [30]

    Ozan Sener and Vladlen Koltun. 2018. Multi-task learning as multi-objective optimization. Advances in neural information processing systems 31 (2018)

  23. [31]

    Laleh Seyyed-Kalantari, Guanxiong Liu, Matthew McDermott, Irene Y Chen, and Marzyeh Ghassemi. 2020. CheXclusion: Fairness gaps in deep chest X-ray classifiers. In BIOCOMPUTING 2021: proceedings of the Pacific symposium . World Scientific, 232–243

  24. [32]

    Trevor Standley, Amir Zamir, Dawn Chen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese. 2020. Which tasks should be learned together in multi-task learning?. In International conference on machine learning . PMLR, 9120–9132

  25. [33]

    Sina Taslimi, Soroush Taslimi, Nima Fathi, Mohammadreza Salehi, and Moham- mad Hossein Rohban. 2022. Swinchex: Multi-label classification on chest x-ray images with transformers. arXiv preprint arXiv:2206.04246 (2022)

  26. [34]

    Tom Van Sonsbeek, Xiantong Zhen, Dwarikanath Mahapatra, and Marcel Wor- ring. 2023. Probabilistic integration of object level annotations in chest x-ray classification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 3630–3640

  27. [35]

    H Wang and Y Xia. [n. d.]. Chestnet: A deep neural network for classifica- tion of thoracic diseases on chest radiography. arXiv 2018. arXiv preprint arXiv:1807.03058 ([n. d.])

  28. [36]

    Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. 2017. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference...

  29. [37]

    Zirui Wang, Yulia Tsvetkov, Orhan Firat, and Yuan Cao. 2020. Gradient vaccine: Investigating and improving multi-task optimization in massively multilingual models. arXiv preprint arXiv:2010.05874 (2020)

  30. [38]

    Derrick Xin, Behrooz Ghorbani, Justin Gilmer, Ankush Garg, and Orhan Firat

  31. [39]

    H Mehta Yang, T Duan, D Ding, A Bagul, C Langlotz, K Shpanskaya, et al

  32. [40]

    Yongxin Yang and Timothy M Hospedales. 2016. Trace norm regularised deep multi-task learning. arXiv preprint arXiv:1606.04038 (2016)

  33. [41]

    Li Yao, Eric Poblenz, Dmitry Dagunts, Ben Covington, Devon Bernard, and Kevin Lyman. 2017. Learning to diagnose from scratch by exploiting dependencies among labels. arXiv preprint arXiv:1710.10501 (2017)

  34. [42]

    Li Yao, Jordan Prosky, Eric Poblenz, Ben Covington, and Kevin Lyman. 2018. Weakly supervised medical diagnosis and localization from multiple resolutions. arXiv preprint arXiv:1803.07703 (2018)

  35. [43]

    Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. 2020. Gradient surgery for multi-task learning. Advances in neural information processing systems 33 (2020), 5824–5836

  36. [44]

    Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. 2018. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition . 3712–3722

  37. [45]

    Yu Zhang and Qiang Yang. 2021. A survey on multi-task learning. IEEE transac- tions on knowledge and data engineering 34, 12 (2021), 5586–5609

  38. [46]

    Zhanpeng Zhang, Ping Luo, Chen Change Loy, and Xiaoou Tang. 2014. Facial landmark detection by deep multi-task learning. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13. Springer, 94–108. 10

  39. [2017]

    arXiv preprint arXiv:1711.05225 (2017)

    CheXNet: radiologist-level pneumonia detection on chest x-rays with deep learning. arXiv preprint arXiv:1711.05225 (2017)

  40. [2019]

    In Proceedings of the AAAI conference on artificial intelligence , Vol

    Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence , Vol. 33. 590–597

  41. [2022]

    Do current multi-task optimization methods in deep learning even help? Advances in neural information processing systems 35 (2022), 13597–13609

Pith tools

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