Pith. sign in

REVIEW 4 major objections 6 minor 52 references

AutoScale: Linear Scalarization Guided by Multi-Task Optimization Metrics

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Good multi-task scalarization weights keep task gradients balanced and stable, and these metric trends can guide weight selection inside one training run.

desk verdict A useful empirical observation and a practical two-phase heuristic, but the central correlation is under-evidenced and the exploration-phase transfer is unvalidated. read the letter →

arxiv 2508.13979 v1 pith:NWCIDJ73 submitted 2025-08-19 cs.LG

Yi Yang , Kei Ikemura , Qingwen Zhang , Xiaomeng Zhu , Ci Li , Nazre Batool , Sina Sharif Mansouri , John Folkesson This is my paper
classification cs.LG
keywords multi-tasklearninglinearscalarizationtaskweightselectiongradientmagnitudesimilarityconditionnumberlossbalancingoptimizationautonomousdrivingbenchmark
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

Multi-task learning usually relies on either complex gradient-balancing algorithms or a brute-force search over scalarization weights. This paper claims there is a middle path: the quality of a scalarization weight is visible in standard multi-task optimization metrics, with good weights showing high gradient magnitude similarity, a low condition number of the gradient matrix, and well-balanced losses. It backs this with 19 fixed-weight runs showing a clear correlation, then turns the correlation into AutoScale, a two-phase procedure that optimizes those metrics inside a single training run and applies the resulting weight for the rest of training. If the claim holds, expensive weight search becomes unnecessary, and plain linear scalarization becomes competitive with, and often faster than, state-of-the-art multi-task optimization methods.

What carries the argument

The load-bearing objects are the three cost functions F(w) that turn MTO heuristics into weight selectors: Equal |g|, which penalizes pairwise differences in rescaled task-gradient magnitudes through a difference matrix A^t; Equal |l|, the same construction on task-loss scales; and Low Cond., which minimizes the condition number κ(G^t_w) = σmax/σmin of the weight-scaled gradient matrix. Each cost is averaged over a 50-iteration window, minimized by SLSQP to pick the next window's weight, and the final fixed weight is the mean of the last 10 window weights (exploration ratio α=0.2). The pairwise magnitude-similarity identity (2|g_i||g_j|/(|g_i|²+|g_j|²)) and the condition-number definition co

What would settle it

Run AutoScale on NYUv2, extract the final weight ŵ, then retrain from a fresh initialization with ŵ fixed for the whole run. If this fixed-weight rerun lands below the top quartile of the 20 grid-search weights, or clearly below AutoScale's own reported result, then the exploration phase is not finding a good weight, only a weight that works within its own changing-weight history. A second check: replace the exploration phase's metric-based weight updates with random weights each 50-iteration window, keep the same final-weight averaging, and compare; if performance is unchanged, the MTO-metric

Watch

Extended reading notes

Core claim

This paper claims that the effectiveness of linear scalarization weights is governed by multi-task optimization (MTO) metrics and can be predicted from them. Across 19 fixed-weight runs, better-performing weights consistently show higher gradient magnitude similarity, a lower condition number of the scaled gradient matrix, and lower standard deviation of relative loss, while other candidate metrics (cosine similarity, loss descending rate) stay flat or inconsistent. On this basis the paper poses weight selection as an optimization problem: minimize a cost function built from these metrics subject to the weights summing to K. AutoScale solves that problem inside one training run, exploring we

Load-bearing premise

The correlation between MTO metrics and performance was measured on runs trained from scratch with a fixed weight (Section 4.2, Figure 3), but AutoScale selects weights from gradients and losses collected while the weight itself changes every 50 iterations; the paper assumes a weight that looks good by these metrics under changing-weight dynamics will also be a good fixed weight when applied from the exploration point onward, without ever isolating that transfer.

Editorial extensions

If this is right

  • Weight search for linear scalarization can be replaced by a single training run: AutoScale's metric-guided exploration matches the best of 20 grid-search trials on CityScapes and NYUv2, and beats it on NuScenes by mean rank and degradation metrics.
  • Linear scalarization with the low-condition-number cost outperforms gradient-manipulating MTO methods (MGDA, PCGrad, GradNorm, IMTL-G, Aligned-MTL) while running 2–3× faster, because only the second phase updates the network with the chosen fixed weights.
  • Any new MTO metric can be plugged into AutoScale's cost function F(w), so the framework's usefulness grows with the MTO literature rather than competing with it.
  • The paper's NuScenes benchmark (3D detection + BEV segmentation, about 156 hours per run) gives the community the first large-scale, high-task-complexity comparison point for MTO methods, where a 20-trial search baseline is essentially unaffordable.
  • Because AutoScale's exploration phase only manipulates scalar weights, it is architecture-agnostic and can be bolted onto any shared-parameter network with the same per-iteration cost as plain scalarization plus a small SLSQP solve per window.

Reading between the lines

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

  • If the metric-performance link is causal rather than coincidental, the same gradient and loss statistics could select weights online for non-stationary task sets (arriving tasks, changing data distributions), where grid search is impossible.
  • The paper treats its three cost functions separately but observes they interact (loss-scale optimization destabilizes the condition number early in training); a composite cost that penalizes low gradient similarity, high condition number, and loss imbalance simultaneously is the natural next experiment, and the correlation view predicts it should beat any single metric.
  • Because F(w) is just a function of gradients and losses, AutoScale extends to any architecture with shared parameters and to per-domain weighting in multi-domain learning, not just the three benchmarks tested.
  • The fixed hyperparameters α=0.2 and τ=50 were used across datasets of very different scale (CityScapes vs NuScenes, a roughly 30× difference in training time); the metric-correlation picture predicts deeper exploration helps where tasks conflict more, a check the paper leaves open.
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 investigates why certain fixed task weights in linear scalarization perform well in multi-task learning (MTL). It analyzes 19 fixed-weight runs on CityScapes, measuring ten 'MTO metrics' (gradient magnitude similarity, gradient cosine similarity, condition number, inverse learning rate, loss descending rate, relative loss scale, etc.), and claims that well-performing weights exhibit high gradient magnitude similarity, low condition number, and low standard deviation of relative loss (Section 4.2, Figures 2–3). Based on this claimed correlation, the paper proposes AutoScale (Section 5, Algorithm 1), a two-phase method: in the exploration phase, weights are updated every τ=50 iterations by minimizing a cost function built from one of three selected metrics; in the linear scalarization phase, the average of the last η weights is fixed for the remaining training. AutoScale is evaluated on CityScapes, NYUv2, and a new large-scale NuScenes benchmark, and the authors report competitive performance with lower training cost than gradient-manipulating MTO methods.

Significance. If the central claim holds, the paper would make a useful contribution: it offers a principled, low-cost alternative to grid search for scalarization weights, and it introduces a new large-scale benchmark (NuScenes with UniTR) that could benefit the MTL community. The cost functions are not fitted to final performance; they are computed from gradient and loss statistics, so the method itself is not circular in the usual sense. The authors also report time efficiency, which is often missing in MTO comparisons. However, the evidence for the claimed 'direct connection' is currently only a visual correlation on a single dataset, without quantitative correlation coefficients, multiple seeds, or held-out validation of the metric-selection procedure. The transfer from fixed-weight full runs to the changing-weight exploration phase is also unexamined. These gaps are load-bearing for the paper's main claims, so the paper is not yet ready for acceptance.

major comments (4)
  1. [Section 4.2, Figure 3] The central claim of a 'strong correlation' between scalarization performance and the three key MTO metrics is supported only by visual inspection of 19 runs on CityScapes. No quantitative correlation coefficient, confidence interval, or standard error is reported, and there is no indication of multiple seeds. Because these correlations are the motivation for the cost functions in Section 5, the claim is load-bearing. Please report, for each metric, a rank correlation (e.g., Spearman) with the performance ranking across the 19 runs, ideally with confidence intervals, and repeat the analysis at least on a held-out dataset (NYUv2 or NuScenes) to show the pattern is not idiosyncratic to CityScapes.
  2. [Section 5, Algorithm 1] The correlation in Section 4.2 is measured for fixed-weight runs trained from scratch. AutoScale, however, computes its weight updates from gradient/loss statistics collected over 50-iteration windows while the weight is itself changing (Algorithm 1), and the final averaged weight is applied only for the remaining (1−α)T iterations, not from scratch. The paper does not demonstrate that the metric–performance relationship holds at the window scale, at mid-training parameter states, or under a drifting weight schedule. Figure 6 merely shows that AutoScale's own trajectories have favorable metric values; it does not compare with a fixed-weight baseline at the same training state, nor does it show that optimizing these metrics causes good final performance. This transfer assumption is load-bearing. Please add a controlled experiment: e.g., compare AutoScale's selected weight applied from the
  3. [Section 6, Tables 1 and 2] The selection of the 'key metrics' was made by examining performance outcomes on CityScapes (Figure 3), and the method is then evaluated on the same dataset. This selection-on-test-data weakens the claim that these metrics are generally predictive. The results on NYUv2 are notably mixed: in Table 2, the Equal|l| variant has MR 7.67 and Δmdeg 9.81, substantially worse than several baselines, and even the Low Cond. variant, while competitive, does not clearly outperform FairGrad. The text's statement that AutoScale 'consistently shows superior performance' is therefore an overstatement. Please present the correlation analysis for NYUv2 and NuScenes (the supplementary is referenced but not included in the main text), and soften the claims in light of the mixed NYUv2 results.
  4. [Section 6, Default Settings] The exploration ratio α=0.2, window size τ=50, and aggregation size η=10 are introduced without any sensitivity analysis in the main text; the ablation is deferred to the supplementary. These hyperparameters control the exploration–exploitation tradeoff and are not derived from theory. A brief ablation in the main text (at least for α and η) is needed to show that the method's performance does not hinge on a specific choice of these values. Without this, the 'without expensive weight search' claim is only partially supported, since AutoScale itself introduces new hyperparameters.
minor comments (6)
  1. [Tables 1 and 2] Typos: 'Linear Scalization' should be 'Linear Scalarization'; 'bolt' should be 'bold'; 'trails' should be 'trials'.
  2. [Section 6, Baseline Methods] The baseline list says 'MDGA' but the referenced method is MGDA (Sener and Koltun, 2017). Also, the text says 'ITML' where IMTL is meant.
  3. [Section 4.1] Notation conflict: in the metric definitions, M^t_GMS(w_i) uses subscript i for the weight index in W, but the same subscript i is then reused as a task index in the definitions involving g_i and g_j. This makes equations like (2) ambiguous. Please use a different index for the weight (e.g., w_n) and task indices (e.g., a, b).
  4. [Section 4.1] The paper says 'ten MTO metrics' are evaluated, but only six are formally defined in the main text; the rest are deferred to the supplementary. Since the choice of metrics is central to the analysis, please at least list all ten metrics with their names and a reference to their source.
  5. [Figure 3] The caption should specify that the metric values are averaged over training iterations and should state the number of runs (19) and the dataset (CityScapes) directly.
  6. [Equation (3)] The definition of A^t is unclear: the matrix rows correspond to task pairs but the index k appears in the row definition. Please clarify the indexing or provide an explicit example for K=3 with labeled rows/columns.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: AutoScale's weight selection is driven by gradient/loss statistics, not by the final performance it later reports.

full rationale

The claimed derivation—that certain MTO metrics (high gradient magnitude similarity, low condition number, low std of relative loss) are associated with good fixed scalarization weights, and that AutoScale's cost functions optimize those metrics—does not reduce to its inputs by construction. The cost functions in Sec. 5.1 are computed from per-task gradient norms and losses (Eq. 3, Low Cond. definition), not from the final Δm/MR scores used as performance in Sec. 4.2 and Tables 1–2. The correlation evidence in Fig. 3 is an empirical observation, not a definitional identity: linear scalarization does not fix gradient cosine similarity (the paper itself notes this in Sec. 4.2), and high gradient-magnitude similarity is not equivalent to low Δm. AutoScale's exploration phase minimizes F(w|G,L) over local windows (Alg. 1), so the selected weights are not fitted to held-out performance. Neither the metric definitions nor the procedure import a uniqueness theorem or a self-citation chain; the paper cites external MTO methods for the individual heuristics. The main legitimate concern is external validity—the metric–performance correlation was measured on fixed-weight full runs while AutoScale selects weights on τ=50 changing-weight windows—but that is a generalization gap, not circularity, because nothing in Eq. 2 or Alg. 1 equates the optimized metric with the reported outcome. Minor caveat: the choice of which three metrics count as 'key' was made on CityScapes and CityScapes is also used as a validation benchmark, which weakens the discovery claim slightly, but the NYUv2 and NuScenes results provide independent grounding. Score 1.

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

The central claim rests on the empirical choice of 'key' MTO metrics, which are taken from existing MTO literature and selected using the paper's own CityScapes experiments. The only hand-chosen numeric parameters are the three AutoScale hyperparameters (alpha, tau, eta), which are fixed defaults across all experiments. No new physical or mathematical entities are introduced.

free parameters (3)
  • Exploration ratio alpha = 0.2
    Hand-chosen default in Section 6; controls the fraction of training spent in the exploration phase. Not tuned per dataset, but affects which weights are selected and final performance.
  • Window size tau = 50
    Hand-chosen default; the number of iterations between weight updates in Phase 1. Influences how many local weights are produced and the statistics used for optimization.
  • Aggregation size eta = 10
    Hand-chosen default; the number of recent local weights averaged to form the final fixed weight.
assumptions (3)
  • domain assumption The chosen MTO metrics (gradient magnitude similarity, condition number, relative loss scale) are appropriate surrogates for final multi-task performance.
    The claim that high GMS, low CN, and low loss-scale deviation imply good scalarization performance is an empirical heuristic, imported from prior MTO literature, not derived in this paper.
  • ad hoc to paper The metric-performance correlation observed on CityScapes (Figure 3) generalizes to NYUv2 and NuScenes and to the changing-weight exploration phase.
    The key metrics were selected from CityScapes outcomes; AutoScale's success depends on this correlation holding elsewhere and under non-fixed-weight training.
  • standard math Minimizing the proposed cost functions in each window (e.g., Eq. 3 and the condition-number objective) with SLSQP reaches a usable local optimum.
    The paper invokes SLSQP (Section 6) without convergence guarantees for these nonconvex objectives.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AutoScale: Linear Scalarization Guided by Multi-Task Optimization Metrics." pith.science (2026). https://pith.science/paper/NWCIDJ73

@misc{pith2026250813979,
  author       = {Pith},
  title        = {Pith review of: AutoScale: Linear Scalarization Guided by Multi-Task Optimization Metrics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NWCIDJ73}},
  note         = {Machine review of arXiv:2508.13979}
}
read the original abstract

Recent multi-task learning studies suggest that linear scalarization, when using well-chosen fixed task weights, can achieve comparable to or even better performance than complex multi-task optimization (MTO) methods. It remains unclear why certain weights yield optimal performance and how to determine these weights without relying on exhaustive hyperparameter search. This paper establishes a direct connection between linear scalarization and MTO methods, revealing through extensive experiments that well-performing scalarization weights exhibit specific trends in key MTO metrics, such as high gradient magnitude similarity. Building on this insight, we introduce AutoScale, a simple yet effective two-phase framework that uses these MTO metrics to guide weight selection for linear scalarization, without expensive weight search. AutoScale consistently shows superior performance with high efficiency across diverse datasets including a new large-scale benchmark.

Figures

Figures reproduced from arXiv: 2508.13979 by the authors.

Figure 1
Figure 1. Illustration showing different multi-task training issues, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Evaluation on different MTO metrics and how they evolve during the training process of seven linear scalarization weight sets [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Performance (∆m ↓: average of performance drop compared to single-task learning, lower value indicate higher performance, as detailed in Section 6) vs. metrics values (average over training iterations) of 19 weight sets of linear scalarization. The plots illustrate a clear correlation between high performance and high gradient magnitude similarity, low condition number, and low standard deviation of relative loss am… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Weight evolution in AutoScale. AutoScale divides the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Performance (∆m%) vs. training time. ⋆ The searched weights are from 20 trials, but only one trial’s time is plotted; actual time is 20 times longer [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: How metrics evolve during AutoScale training? In ad [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 44 canonical work pages

  1. [1]

    Bayesian uncertainty for gradient aggre- gation in multi-task learning, 2024

    Idan Achituve, Idit Diamant, Arnon Netzer, Gal Chechik, and Ethan Fetaya. Bayesian uncertainty for gradient aggre- gation in multi-task learning, 2024. 2

    work page 2024

  2. [2]

    Fair resource allocation in multi-task learning

    Hao Ban and Kaiyi Ji. Fair resource allocation in multi-task learning. arXiv preprint arXiv:2402.15638, 2024. 1, 6, 7

    arXiv 2024

  3. [3]

    nuscenes: A multi- modal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multi- modal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631, 2020. 6

    2020

  4. [4]

    Three-way trade-off in multi-objective learning: Op- timization, generalization and conflict-avoidance

    Lisha Chen, Heshan Fernando, Yiming Ying, and Tianyi Chen. Three-way trade-off in multi-objective learning: Op- timization, generalization and conflict-avoidance. Advances in Neural Information Processing Systems, 36:70045–70093,

    work page

  5. [5]

    Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks

    Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and An- drew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In In- ternational conference on machine learning, pages 794–803. PMLR, 2018. 1, 2, 4, 6, 7

    work page 2018

  6. [6]

    Just pick a sign: Optimizing deep multitask models with gra- dient sign dropout

    Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gra- dient sign dropout. Advances in Neural Information Process- ing Systems, 33:2039–2050, 2020. 2

    work page

  7. [7]

    Multinet++: Multi-stream feature ag- gregation and geometric loss strategy for multi-task learning

    Sumanth Chennupati, Ganesh Sistu, Senthil Yogamani, and Samir A Rawashdeh. Multinet++: Multi-stream feature ag- gregation and geometric loss strategy for multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 2

    work page 2019

  8. [8]

    The cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 3213–3223, 2016. 1, 6

    work page 2016

Show all 52 references
  1. [9]

    Instance-aware se- mantic segmentation via multi-task network cascades

    Jifeng Dai, Kaiming He, and Jian Sun. Instance-aware se- mantic segmentation via multi-task network cascades. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3150–3158, 2016. 2

    2016

  2. [10]

    K ¨ohler, and Lukas Schott

    Cathrin Elich, Lukas Kirchdorfer, Jan M. K ¨ohler, and Lukas Schott. Challenging common paradigms in multi-task learn- ing, 2024. 3

    2024

  3. [11]

    Mitigating gradi- ent bias in multi-objective learning: A provably convergent approach

    Heshan Fernando, Han Shen, Miao Liu, Subhajit Chaudhury, Keerthiram Murugesan, and Tianyi Chen. Mitigating gradi- ent bias in multi-objective learning: A provably convergent approach. In ICLR. International Conference on Learning Representations, 2023. 2

    2023

  4. [12]

    Dynamic task prioritization for multitask learning

    Michelle Guo, Albert Haque, De-An Huang, Serena Yeung, and Li Fei-Fei. Dynamic task prioritization for multitask learning. In Proceedings of the European conference on com- puter vision (ECCV), pages 270–287, 2018. 2

    2018

  5. [13]

    Ro- bust multi-task learning with excess risks

    Yifei He, Shiji Zhou, Guojun Zhang, Hyokun Yun, Yi Xu, Belinda Zeng, Trishul Chilimbi, and Han Zhao. Ro- bust multi-task learning with excess risks. arXiv preprint arXiv:2402.02009, 2024. 2

    2024 arXiv

  6. [14]

    Revisiting scalarization in multi-task learning: A theoretical perspective

    Yuzheng Hu, Ruicheng Xian, Qilong Wu, Qiuling Fan, Lang Yin, and Han Zhao. Revisiting scalarization in multi-task learning: A theoretical perspective. Advances in Neural In- formation Processing Systems, 36, 2024. 1, 3

    2024

  7. [15]

    Fuller: Unified multi-modality multi-task 3d perception via multi-level gradient calibration

    Zhijian Huang, Sihao Lin, Guiyu Liu, Mukun Luo, Chao- qiang Ye, Hang Xu, Xiaojun Chang, and Xiaodan Liang. Fuller: Unified multi-modality multi-task 3d perception via multi-level gradient calibration. In Proceedings of the IEEE/CVF International Conference on Computer Vision ,...

    2023

  8. [16]

    Online knowledge distillation for multi-task learning

    Geethu Miriam Jacob, Vishal Agarwal, and Bj ¨orn Stenger. Online knowledge distillation for multi-task learning. InPro- ceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, pages 2359–2368, 2023. 2

    2023

  9. [17]

    Selective task group updates for multi-task optimization, 2025

    Wooseong Jeong and Kuk-Jin Yoon. Selective task group updates for multi-task optimization, 2025. 2

    2025

  10. [18]

    Multi-task learning using uncertainty to weigh losses for scene geome- try and semantics

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

  11. [19]

    A software package for sequential quadratic programming

    Dieter Kraft. A software package for sequential quadratic programming. Forschungsbericht- Deutsche Forschungs- und Versuchsanstalt fur Luft- und Raumfahrt, 1988. 8

    1988

  12. [20]

    In defense of the uni- tary scalarization for deep multi-task learning

    Vitaly Kurin, Alessandro De Palma, Ilya Kostrikov, Shimon Whiteson, and Pawan K Mudigonda. In defense of the uni- tary scalarization for deep multi-task learning. Advances in Neural Information Processing Systems , 35:12169–12183,

  13. [21]

    Deep asymmetric multi-task feature learning

    Hae Beom Lee, Eunho Yang, and Sung Ju Hwang. Deep asymmetric multi-task feature learning. In International Conference on Machine Learning, pages 2956–2964. PMLR,

  14. [22]

    Reasonable effectiveness of random weighting: A litmus test for multi-task learning

    Baijiong Lin, Feiyang Ye, Yu Zhang, and Ivor W Tsang. Reasonable effectiveness of random weighting: A litmus test for multi-task learning. arXiv preprint arXiv:2111.10603 ,

    arXiv

  15. [23]

    Baijiong Lin, Weisen Jiang, Feiyang Ye, Yu Zhang, Peng- guang Chen, Ying-Cong Chen, Shu Liu, and James T. Kwok. Dual-balancing for multi-task learning, 2023. 2

    2023

  16. [24]

    Smooth tchebycheff scalar- ization for multi-objective optimization

    Xi Lin, Xiaoyuan Zhang, Zhiyuan Yang, Fei Liu, Zhenkun Wang, and Qingfu Zhang. Smooth tchebycheff scalar- ization for multi-objective optimization. arXiv preprint arXiv:2402.19078, 2024. 1, 6, 7

    2024 arXiv

  17. [25]

    Conflict-averse gradient descent for multi-task learn- ing

    Bo Liu, Xingchao Liu, Xiaojie Jin, Peter Stone, and Qiang Liu. Conflict-averse gradient descent for multi-task learn- ing. Advances in Neural Information Processing Systems , 34:18878–18890, 2021. 2 9

    2021

  18. [26]

    Famo: Fast adaptive multitask optimization

    Bo Liu, Yihao Feng, Peter Stone, and Qiang Liu. Famo: Fast adaptive multitask optimization. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 4, 6, 7

    2024

  19. [27]

    Towards impartial multi-task learning

    Liyang Liu, Yi Li, Zhanghui Kuang, J Xue, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Towards impartial multi-task learning. In iclr, 2021. 1, 2, 6, 7

    2021

  20. [28]

    Online mirror descent for tchebycheff scalarization in multi-objective optimization

    Meitong Liu, Xiaoyuan Zhang, Chulin Xie, Kate Donahue, and Han Zhao. Online mirror descent for tchebycheff scalarization in multi-objective optimization. arXiv preprint arXiv:2410.21764, 2024. 2

    2024

  21. [29]

    End- to-end multi-task learning with attention

    Shikun Liu, Edward Johns, and Andrew J Davison. End- to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871–1880, 2019. 2, 4

    2019

  22. [30]

    End- to-end multi-task learning with attention

    Shikun Liu, Edward Johns, and Andrew J Davison. End- to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871–1880, 2019. 7, 8

    2019

  23. [31]

    Auto-lambda: Disentangling dynamic task relation- ships

    Shikun Liu, Stephen James, Andrew J Davison, and Edward Johns. Auto-lambda: Disentangling dynamic task relation- ships. arXiv preprint arXiv:2202.03091, 2022. 1

    2022 arXiv

  24. [32]

    Learning multiple tasks with multilinear relationship net- works

    Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Philip S Yu. Learning multiple tasks with multilinear relationship net- works. Advances in neural information processing systems , 30, 2017. 2

    2017

  25. [33]

    Cross-stitch networks for multi-task learning

    Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Mar- tial Hebert. Cross-stitch networks for multi-task learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2

    2016

  26. [34]

    Multi- task learning as a bargaining game

    Aviv Navon, Aviv Shamsian, Idan Achituve, Haggai Maron, Kenji Kawaguchi, Gal Chechik, and Ethan Fetaya. Multi- task learning as a bargaining game. arXiv preprint arXiv:2202.01017, 2022. 2

    2022 arXiv

  27. [35]

    Jacobian descent for multi- objective optimization

    Pierre Quinton and Val ´erian Rey. Jacobian descent for multi- objective optimization. arXiv preprint arXiv:2406.16232 ,

    arXiv

  28. [36]

    Scalarization for multi-task and multi- domain learning at scale

    Amelie Royer, Tijmen Blankevoort, and Babak Ehte- shami Bejnordi. Scalarization for multi-task and multi- domain learning at scale. Advances in Neural Information Processing Systems, 36, 2024. 1, 3

    2024

  29. [37]

    Multi-task learning as multi-objective optimization

    Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural informa- tion processing systems, 31, 2018. 1, 6, 7

    2018

  30. [38]

    Independent component alignment for multi-task learning

    Dmitry Senushkin, Nikolay Patakin, Arseny Kuznetsov, and Anton Konushin. Independent component alignment for multi-task learning. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 20083–20093, 2023. 1, 2, 4, 6, 7

    2023

  31. [39]

    Go4align: Group optimization for multi-task alignment, 2024

    Jiayi Shen, Cheems Wang, Zehao Xiao, Nanne Van Noord, and Marcel Worring. Go4align: Group optimization for multi-task alignment, 2024. 2

    2024

  32. [40]

    Indoor segmentation and support inference from rgbd images

    Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision–ECCV 2012: 12th Eu- ropean Conference on Computer Vision, Florence, Italy, Oc- tober 7-13, 2012, Proceedings, Part V 12 , pages 746–760...

    2012

  33. [41]

    Which tasks should be learned together in multi-task learning? In Proceedings of the 37th International Conference on Machine Learning , pages 9120–9132

    Trevor Standley, Amir Zamir, Dawn Chen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese. Which tasks should be learned together in multi-task learning? In Proceedings of the 37th International Conference on Machine Learning , pages 9120–9132. PMLR, 2020. 2

    2020

  34. [42]

    Regularizing deep multi- task networks using orthogonal gradients

    Mihai Suteu and Yike Guo. Regularizing deep multi- task networks using orthogonal gradients. arXiv preprint arXiv:1912.06844, 2019. 2, 4

    1912 arXiv

  35. [43]

    Discovering struc- ture in multiple learning tasks: The tc algorithm

    Sebastian Thrun and Joseph O’Sullivan. Discovering struc- ture in multiple learning tasks: The tc algorithm. In ICML, pages 489–497. Citeseer, 1996. 2

    1996

  36. [44]

    Unitr: A unified and efficient multi-modal transformer for bird’s-eye-view repre- sentation

    Haiyang Wang, Hao Tang, Shaoshuai Shi, Aoxue Li, Zhen- guo Li, Bernt Schiele, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird’s-eye-view repre- sentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6792–680...

    2023

  37. [45]

    Direction-oriented multi-objective learning: Simple and provable stochastic al- gorithms

    Peiyao Xiao, Hao Ban, and Kaiyi Ji. Direction-oriented multi-objective learning: Simple and provable stochastic al- gorithms. Advances in Neural Information Processing Sys- tems, 36:4509–4533, 2023. 1

    2023

  38. [46]

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

    Derrick Xin, Behrooz Ghorbani, Justin Gilmer, Ankush Garg, and Orhan Firat. Do current multi-task optimization methods in deep learning even help? Advances in neural information processing systems, 35:13597–13609, 2022. 1, 3

    2022

  39. [47]

    Multi-objective meta learning

    Feiyang YE, Baijiong Lin, Zhixiong Yue, Pengxin Guo, Qiao Xiao, and Yu Zhang. Multi-objective meta learning. In Advances in Neural Information Processing Systems , pages 21338–21351. Curran Associates, Inc., 2021. 1

    2021

  40. [48]

    Taskprompter: Spatial-channel multi-task prompting for dense scene understanding

    Hanrong Ye and Dan Xu. Taskprompter: Spatial-channel multi-task prompting for dense scene understanding. In ICLR, 2023. 2

    2023

  41. [49]

    Gradient surgery for multi-task learning

    Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Pro- cessing Systems, 33:5824–5836, 2020. 1, 2, 4, 6, 7

    2020

  42. [50]

    Achievement-based training progress balancing for multi-task learning

    Hayoung Yun and Hanjoo Cho. Achievement-based training progress balancing for multi-task learning. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 16935–16944, 2023. 2

    2023

  43. [51]

    Taskonomy: Disentangling task transfer learning

    Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 3712–3722, 2018. 2

    2018

  44. [52]

    Pyramid scene parsing network

    Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881–2890, 2017. 6, 7 10

    2017

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.