Pith. sign in

REVIEW 4 major objections 5 minor 54 references

Towards Unified Modeling in Federated Multi-Task Learning via Subspace Decoupling

T0 review · 4 major / 5 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 FedDEA keeps only the top-$\rho$ fraction of each client's update by absolute magnitude, zeros the rest, and rescales by $1/\rho$ before aggregation, claiming this suppresses cross-task interference and lets one unified global model serve…

desk verdict FedDEA is a simple top-magnitude mask plus 1/rho rescale aggregation heuristic for federated multi-task learning; it shows consistent empirical gains across many baselines, but the paper's subspace-decoupling story is not actually validated. read the letter →

arxiv 2505.24185 v1 pith:YJY5JXGY submitted 2025-05-30 cs.LG cs.AIcs.CV

Yipan Wei , Yuchen Zou , Yapeng Li , Bo Du This is my paper
classification cs.LGcs.AIcs.CV
keywords federatedmulti-tasklearningheterogeneoustaskaggregationparameterupdateinterferencesubspacedecouplingmagnitude-basedmaskingrescalingunifiedglobalmodelplug-inmodule
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 argues that in federated multi-task learning, each client's local update can be split into task-relevant signal concentrated in a sparse subspace and task-irrelevant noise spread across shared dimensions, and that standard aggregation mixes both, causing directional drift and degraded multi-task performance. FedDEA is a server-side aggregation step that retains only the top-$\rho$ fraction of each client's update by absolute magnitude, zeroes out the rest, and multiplies the survivors by $1/\rho$ before the usual weighted aggregation. This makes a single global model serve several heterogeneous tasks simultaneously, without requiring task labels or architectural changes. The paper reports that wrapping eight existing federated optimizers with FedDEA yields consistently positive overall gains on NYUD-V2 and PASCAL-Context, whereas the two prior enhancement modules tested often degrade performance.

What carries the argument

The central object is a per-client binary mask $m_k$ built by thresholding the local update $\Delta_k^{(t)}$ at its $\lfloor \rho d \rfloor$-th largest absolute magnitude, so that exactly the top-$\rho$ fraction of dimensions is kept and all others are set to zero (Equations 2-4), followed by a uniform rescaling of the retained dimensions by $1/\rho$ (Equation 5) before the standard weighted aggregation step (Equation 6). The selection ratio $\rho \in (0,1]$ is the only new hyperparameter and controls the trade-off between suppressing interference and preserving optimization energy; the paper reports an inverted-U pattern in $\Delta$ as $\rho$ varies, with the best values in the 0.1-0.2 range for these benchmarks. This two-stage decoupling-and-recalibration mechanism is what the paper claims converts a conventional federated optimizer into a structure-aware multi-task aggregator.

What would settle it

Construct a two-client experiment with synthetic updates in which one task's useful signal is deliberately placed in low-magnitude dimensions while the other task's noise dominates the high-magnitude dimensions; if the FedDEA-masked model performs worse on the first task than plain FedAvg, the magnitude-as-relevance assumption is violated. Alternatively, probe which dimensions each task truly needs by ablating individual parameter dimensions and comparing that ground truth with the FedDEA mask: if frequently needed dimensions are routinely masked out and rescaling does not compensate, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that task-level decoupled aggregation can be achieved purely from update structure: at each communication round, the dimensions where a client's update has the largest absolute magnitudes form that task's support subspace, and different tasks' support subspaces are mostly disjoint. Keeping only those dimensions and rescaling by $1/\rho$ preserves the update's direction and roughly its energy while suppressing shared-dimension interference, so the aggregated global model retains each task's learning signal. The paper positions this as a new capability, unified heterogeneous-task aggregation, that personalized federated multi-task methods cannot provide because they return separate models per task. The evidence is a set of comparisons where FedDEA-integrated versions of FedAvg, FedProx, FedNova, MOON, FedDyn, FedBN, FedGA, and FedACG all show positive $\Delta$ gains, with representative gains of +10.60% to +12.65% on NYUD-V2 and +7.16% to +12.89% on PASCAL-Context for the Swin-T backbone, and similar improvement patterns with a ResNet-18 backbone.

Load-bearing premise

The load-bearing premise is that the dimensions of a client's update with the largest absolute magnitudes are exactly the dimensions that carry task-relevant signal, and that different tasks' important dimensions are mostly disjoint; the paper's only direct evidence is a two-round visualization.

Editorial extensions

If this is right

  • FedDEA can be wrapped around an existing federated optimizer without changing the model architecture or adding communication overhead, so multi-task decoupling is achievable as a drop-in server-side change.
  • Because FedDEA produces one unified global model that handles all client tasks, it changes the deployment footprint of federated multi-task systems compared with personalization-based methods that require a separate model per task.
  • The paper's comparisons suggest that aggregation modules which ignore update structure, such as PCGrad and FedHEAL in these tests, can harm heterogeneous multi-task performance, while magnitude-based structural filtering consistently improves it.
  • The selection ratio $\rho$ is a meaningful tuning dial: too large a ratio admits interference, too small a ratio strips optimization energy, and the reported peak around 0.1-0.2 indicates that aggressive but not extreme filtering is the effective operating point.

Reading between the lines

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

  • This suggests a close link to task-vector arithmetic: FedDEA can be read as a federated analogue that builds the global model as a weighted sum of cleaned task vectors, so it may inherit both the strengths and the failure modes of task-vector methods, including sensitivity to shared initialization and model width.
  • Because the method uses update magnitude as a proxy for task relevance, its success likely depends on clients having roughly comparable local optimization schedules; a testable extension is to vary local epochs or learning rates across clients and measure whether FedDEA's gains shrink when update magnitudes become less comparable.
  • The $1/\rho$ rescaling only approximately preserves the original update energy, since the true energy remaining after masking is generally less than the full update energy at small $\rho$; whether this approximation matters for convergence is an open question the paper does not resolve.
  • A natural non-federated test is to apply the same magnitude-masking and rescaling inside a single multi-task learner that trains all tasks jointly, which would separate whether the disjoint-subspace structure is intrinsic to the tasks or an artifact of independently optimized client paths.
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 / 5 minor

Summary. The paper proposes FedDEA, a server-side aggregation wrapper for federated multi-task learning with heterogeneous tasks. At each communication round, FedDEA computes each client's parameter update, masks it to the top-ρ fraction of dimensions by absolute magnitude, rescales the retained coordinates by 1/ρ, and then applies the conventional weighted aggregation rule. The authors argue that different tasks activate approximately disjoint parameter subspaces and that magnitude-based filtering decouples task-relevant updates from cross-task interference. They evaluate FedDEA on NYUD-V2 and PASCAL-Context with each task assigned to one client, integrating it into eight federated optimization algorithms and comparing against PCGrad and FedHEAL, supplemented by ablations and a ResNet-18 backbone extension.

Significance. If the subspace-decoupling premise holds, FedDEA is an appealingly simple and deployment-friendly module: it requires no task labels, no architecture changes, and no additional communication, and it produces a single unified model supporting all tasks. The paper's strengths are the breadth of base optimizers tested, the two-backbone evaluation, and the inclusion of ablations that separate the mask criterion from the rescaling step, with random-mask and no-rescale controls. The main weaknesses are the lack of direct quantitative evidence for the load-bearing task-support-subspace assumption, the absence of any statistical characterization of the results, and an internal inconsistency in the reported FedProx numbers. The contribution is potentially useful, but the evidence as presented does not yet establish that the mechanism works through task-level subspace decoupling rather than through a generic high-magnitude upweighting effect.

major comments (4)
  1. [Section 3.2, Eqs. (3)-(5)] The central premise that top-magnitude update dimensions are the task-relevant dimensions, and that such dimensions are disjoint or approximately orthogonal across tasks, is asserted from Figure 2 but never quantitatively validated. Figure 2 is a two-round visualization with no reported overlap statistics, no longitudinal trend, and no link between mask overlap and final task performance. The method's entire mechanism is defined by the magnitude ranking in Eqs. (3)-(4), so the paper should provide, at minimum, the average intersection-over-union of masks across task-clients, the fraction of update energy captured by the top-ρ dimensions, and the stability of these masks across rounds. Without such evidence, the consistent positive Δ in Tables 3 and 5 could be explained by a generic effect of upweighting informative high-magnitude coordinates rather than by task-level decoupling; the Random Mask ablation is helpful but only rules out arbitrary masking and does not establish disjointness or task relevance.
  2. [Tables 1 and 3] There is an internal inconsistency in the reported FedProx+DEA results on NYUD-V2. Table 1 lists Semseg mIoU 31.46, Depth RMSE 0.6958, Normals mErr 24.42, and Edge odsF 74.83, while Table 3 lists Semseg mIoU 32.36, Depth RMSE 0.7229, Normals mErr 24.61, and Edge odsF 75.31 for the same configuration. The FedAvg+DEA rows match between the two tables, so this is not a general formatting issue. The authors should clarify which numbers are correct, whether different values of ρ or different random seeds were used, and should ensure that all tables report the same configuration consistently. This discrepancy undermines confidence in the reliability of the other reported numbers.
  3. [Tables 3, 5, and Figure 4] All experimental results appear to be single runs with no error bars, no standard deviations, and no significance tests. Several of the claimed improvements are small (e.g., FedGA +1.50% and FedACG +2.20% on Swin-T; FedGA +1.49% on ResNet-18), so the abstract's claim of 'consistently delivers significant overall performance improvements' is not statistically supported. In addition, the selection ratio ρ is tuned on NYUD-V2 in Figure 4 (the best Δ of 10.60% occurs at ρ=0.2 for FedAvg), and the same or a selected value is then used for PASCAL-Context without a reported sensitivity analysis on that dataset. The authors should state how ρ was chosen, whether PASCAL-Context results are a held-out check or also tuned, and should provide at least three seeds with variance for the main tables.
  4. [Tables 1-2 and Section 4.2] The comparison with personalized and multi-task federated methods is not apples-to-apples. FedAMP, FedMTL, and FedHCA2 produce separate models per task, while FedDEA produces a single unified model, so the comparison conflates the modeling paradigm with the aggregation method. On several metrics FedDEA is worse than the personalization baselines (e.g., NYUD Semseg 30.78 vs. 33.86 for FedAMP), and the paper's argument is that a unified model is more valuable. That is a legitimate trade-off, but it needs to be quantified: the authors should report the total number of parameters and communication cost for the unified model versus the sum of the per-task models, and ideally include a centralized multi-task training upper bound trained on all client data to calibrate how much performance is lost by the federated constraint.
minor comments (5)
  1. [Eq. (6)] Equation (6) appears to have a missing learning-rate factor: it reads θ(t+1) = θ(t) + · Σ_k w_k Δ̃_k, where the '+' is followed by a dot. This should be written with an explicit η, and the text should define whether η is the same as the local optimizer's learning rate or a separate server-side learning rate.
  2. [Section 4.4 and Table 4] The text says 'Experimental results in Table 6 show that update structure-unaware masking significantly degrades performance,' but the ablation table in the main paper is numbered Table 4. The cross-reference appears to point to the supplementary ResNet-18 ablation instead. Please correct the numbering and ensure all in-text references match the table numbers.
  3. [Figure 2] The rendering of Figure 2 in the manuscript is not interpretable: the axis labels and panel headings appear as glyph-like placeholder characters. The figure should be regenerated with clear English labels, and it would be more informative to include quantitative heatmaps of mask overlap and update-energy concentration rather than binary activation maps.
  4. [Throughout] There are numerous typos and grammatical slips, including 'clien k erforms' in Section 3.3, 'a update structure aware aggregation strategy' in Section 1, 'plug-in-play' for 'plug-and-play' in several places, and an inconsistent use of 'FedHEAL' versus 'HEAL.' These should be corrected.
  5. [Table 3] In the Moon+DEA row on NYUD-V2, the Δ value is printed as '21.03' without a leading '+' sign, unlike all other positive Δ entries. Please make the formatting consistent.

Circularity Check

1 steps flagged · score 2.0 of 10

No equation-level circularity: the central benchmark claims are measured externally, but the 'task-relevant subspace' framing is self-referential because it defines task-relevance by the same top-magnitude statistic that the method masks.

  1. self definitional [Section 3.2 (Motivation) and Section 3.3, Eqs. (3)-(4)]
    "To extract task-relevant update signals, we introduce a hyperparameter ρ ∈ (0, 1](selection ratio), referred to as the selection ratio, which controls the proportion of update dimensions retained by each client. Specifically, among all elements in ∆(t) k , we identify the top ⌊ρd⌋ dimensions with the largest absolute magnitudes and retain them."

    The motivating observation in Section 3.2 is that 'different tasks activate distinct regions of the model parameter space,' and Figure 2 visualizes 'the top 10% activated update dimensions per task.' This activation statistic is the same top-magnitude criterion later codified in Eq. (3) as the mask. Thus the observation does not independently validate that high-magnitude dimensions are task-relevant; 'task-relevant' is defined by the paper as high response strength, i.e., the selection rule itself. The disjointness of the top-magnitude sets is an empirical fact, but labeling these sets as task-relevant subspaces is a definitional framing rather than a tested hypothesis.

full rationale

FedDEA's aggregation rule (Eqs. 2-6) is a fixed, explicit function of the client updates and the hyperparameter ρ; no parameter is fitted to the benchmark tasks and then reported as a prediction. The performance gains in Tables 3 and 5 are evaluated against external task metrics on NYUD-V2 and PASCAL-Context, so the central claim of consistent improvement is measured independently of the method's construction. The ablation study includes Random Mask and Small Mask variants, which test whether magnitude-based selection matters rather than assuming it. The only self-referential element is the conceptual framing of the top-magnitude mask as 'task-relevant': Section 3.2's motivating observation uses the same top-10% magnitude statistic that Eq. (3) selects, so it does not separately establish that high-magnitude dimensions are task-specific or that tasks' relevant subspaces are disjoint. This weakens the strength of the conceptual 'subspace decoupling' narrative, but it does not make the benchmark results circular. No load-bearing self-citations, imported uniqueness theorems, or ansatz-smuggling citations were found; the reference to task arithmetic [22] is external and used only as motivation. Overall, the reported predictions do not reduce by construction to the method's inputs.

Assumptions & free parameters 1 free parameters · 3 assumptions · 1 invented entities

The method depends on one tuned hyperparameter, the selection ratio ρ, and on three stated or implicit assumptions: update subspace disjointness, magnitude as task-relevance, and energy preservation by the 1/ρ rescale. No new physical or architectural entities are introduced beyond the conceptual 'task-support subspace', which is defined by the mask itself.

free parameters (1)
  • selection ratio ρ = 0.2 (peak of Fig. 4 sweep on NYUD-V2)
    Controls the fraction of top-magnitude update dimensions retained; swept in {0.1, ..., 1.0} and the value giving the largest Δ on NYUD is used in the main tables; no held-out selection protocol is described.
assumptions (3)
  • domain assumption Updates generated by different tasks are concentrated in disjoint or approximately orthogonal subspaces of parameter space.
    Section 3.2 states this based on a toy experiment (Fig. 2); the entire decoupling mechanism depends on this structural separation being true across tasks and rounds.
  • domain assumption The top-ρ fraction by absolute magnitude of a client update approximates the task-relevant subspace; the remaining dimensions are noise.
    Section 3.3 defines the mask purely by magnitude; if task-relevant signal lies in low-magnitude dimensions, masking will remove it. The paper provides no quantitative validation of this proxy.
  • ad hoc to paper Rescaling retained updates by 1/ρ approximately preserves optimization energy.
    Equation (5) and the text claim the rescaling preserves the original energy level; this holds only under specific distributional assumptions about update magnitudes, which are not stated or verified.
invented entities (1)
  • task-support subspace
    purpose: Conceptual subspace said to contain task-relevant update directions; motivates filtering and rescaling.
    Defined in Section 3.2 and operationalized as the top-ρ magnitude mask in Section 3.3; no independent measurement outside the masking criterion is offered.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Unified Modeling in Federated Multi-Task Learning via Subspace Decoupling." pith.science (2026). https://pith.science/paper/YJY5JXGY

@misc{pith2026250524185,
  author       = {Pith},
  title        = {Pith review of: Towards Unified Modeling in Federated Multi-Task Learning via Subspace Decoupling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YJY5JXGY}},
  note         = {Machine review of arXiv:2505.24185}
}
read the original abstract

Federated Multi-Task Learning (FMTL) enables multiple clients performing heterogeneous tasks without exchanging their local data, offering broad potential for privacy preserving multi-task collaboration. However, most existing methods focus on building personalized models for each client and unable to support the aggregation of multiple heterogeneous tasks into a unified model. As a result, in real-world scenarios where task objectives, label spaces, and optimization paths vary significantly, conventional FMTL methods struggle to achieve effective joint training. To address this challenge, we propose FedDEA (Federated Decoupled Aggregation), an update-structure-aware aggregation method specifically designed for multi-task model integration. Our method dynamically identifies task-relevant dimensions based on the response strength of local updates and enhances their optimization effectiveness through rescaling. This mechanism effectively suppresses cross-task interference and enables task-level decoupled aggregation within a unified global model. FedDEA does not rely on task labels or architectural modifications, making it broadly applicable and deployment-friendly. Experimental results demonstrate that it can be easily integrated into various mainstream federated optimization algorithms and consistently delivers significant overall performance improvements on widely used NYUD-V2 and PASCAL-Context. These results validate the robustness and generalization capabilities of FedDEA under highly heterogeneous task settings.

Figures

Figures reproduced from arXiv: 2505.24185 by the authors.

Figure 1
Figure 1. Conceptual comparison between traditional [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Visualization of the top 10% activated update dimensions per task in early and late [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed method. Our approach addresses task heterogeneity in federated [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Hyperparameter analysis of selection rate ρ. 0 20 40 60 80 100 Epoch 0 5 10 15 20 25 30 mIoU FedAVG FedAVG+PCGrad FedAVG+HEAL FedAVG+DEA [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Convergence comparison of Fe￾dAvg and its variants with PCGrad [51], Fed￾HEAL [10], and FedDEA. 0 20 40 60 80 100 Epoch 0 5 10 15 20 25 30 mIoU FedAVG+DEA FedAVG FedDyn+DEA FedDyn FedProx+DEA FedProx [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 7
Figure 7. Figure 7: Convergence comparison of Fe￾dAvg [35] and its variants with PCGrad [51], FedHEAL [10], and FedDEA. 0 20 40 60 80 100 Epoch 0 2 4 6 8 10 12 mIoU FedAVG+DEA FedAVG FedDyn+DEA FedDyn FedProx+DEA FedProx [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 49 canonical work pages

  1. [1]

    Neural multi-task learning in drug design

    Stephan Allenspach, Jan A Hiss, and Gisbert Schneider. Neural multi-task learning in drug design. Nature Machine Intelligence, 6(2):124–137, 2024

    work page 2024

  2. [2]

    Advancing covid-19 diagnosis with privacy-preserving collaboration in artificial intelligence

    Xiang Bai, Hanchen Wang, Liya Ma, Yongchao Xu, Jiefeng Gan, Ziwei Fan, Fan Yang, Ke Ma, Jiehua Yang, Song Bai, et al. Advancing covid-19 diagnosis with privacy-preserving collaboration in artificial intelligence. Nature Machine Intelligence, 3(12):1081–1089, 2021

    work page 2021

  3. [3]

    Federated disentangled representation learning for unsupervised brain anomaly detection

    Cosmin I Bercea, Benedikt Wiestler, Daniel Rueckert, and Shadi Albarqouni. Federated disentangled representation learning for unsupervised brain anomaly detection. Nature Machine Intelligence, 4(8):685– 695, 2022

    work page 2022

  4. [4]

    Many-task federated learning: A new problem setting and a simple baseline

    Ruisi Cai, Xiaohan Chen, Shiwei Liu, Jayanth Srinivasa, Myungjin Lee, Ramana Kompella, and Zhangyang Wang. Many-task federated learning: A new problem setting and a simple baseline. In CVPR, pages 5037–5045, 2023

    work page 2023

  5. [5]

    Multitask learning

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

    1997

  6. [6]

    Fraug: Tackling federated learning with non-iid features via representation augmentation

    Haokun Chen, Ahmed Frikha, Denis Krompass, Jindong Gu, and V olker Tresp. Fraug: Tackling federated learning with non-iid features via representation augmentation. In ICCV, pages 4849–4859, 2023

    work page 2023

  7. [7]

    Multi-task learning in natural language processing: An overview

    Shijie Chen, Yu Zhang, and Qiang Yang. Multi-task learning in natural language processing: An overview. ACM Computing Surveys, 56(12):1–32, 2024

    work page 2024

  8. [8]

    Adamv-moe: Adaptive multi-task vision mixture-of-experts

    Tianlong Chen, Xuxi Chen, Xianzhi Du, Abdullah Rashwan, Fan Yang, Huizhong Chen, Zhangyang Wang, and Yeqing Li. Adamv-moe: Adaptive multi-task vision mixture-of-experts. In ICCV, pages 17346–17357, 2023

    work page 2023

Show all 54 references
  1. [9]

    Fed- bone: Towards large-scale federated multi-task learning

    Yi-Qiang Chen, Teng Zhang, Xin-Long Jiang, Qian Chen, Chen-Long Gao, and Wu-Liang Huang. Fed- bone: Towards large-scale federated multi-task learning. Journal of Computer Science and Technology, 39(5):1040–1057, 2024

    2024

  2. [10]

    Fair federated learning under domain skew with local consistency and domain diversity

    Yuhang Chen, Wenke Huang, and Mang Ye. Fair federated learning under domain skew with local consistency and domain diversity. In CVPR, 2024

    2024

  3. [11]

    Federated learning for predicting clinical outcomes in patients with covid-19

    Ittai Dayan, Holger R Roth, Aoxiao Zhong, Ahmed Harouni, Amilcare Gentili, Anas Z Abidin, Andrew Liu, Anthony Beardsworth Costa, Bradford J Wood, Chien-Sung Tsai, et al. Federated learning for predicting clinical outcomes in patients with covid-19. Nature medicine, 27(10):1735...

    2021

  4. [12]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. Ieee, 2009

    2009

  5. [13]

    Multi-task self-supervised visual learning

    Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In ICCV, pages 2051– 2060, 2017

    2017

  6. [14]

    Federated learning based on dynamic regularization

    Alp Emre Durmus, Zhao Yue, Matas Ramon, Mattina Matthew, Whatmough Paul, and Saligrama Venkatesh. Federated learning based on dynamic regularization. In ICLR, 2021

    2021

  7. [15]

    Spreadgnn: Decentralized multi-task federated learning for graph neural networks on molecular data

    Chaoyang He, Emir Ceyani, Keshav Balasubramanian, Murali Annavaram, and Salman Avestimehr. Spreadgnn: Decentralized multi-task federated learning for graph neural networks on molecular data. In AAAI, volume 36, pages 6865–6873, 2022

    2022

  8. [16]

    Delving deep into rectifiers: Surpassing human-level performance on imagenet classification

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, pages 1026–1034, 2015

    2015

  9. [17]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770–778, 2016

    2016

  10. [18]

    Rethinking federated learning with domain shift: A prototype view

    Wenke Huang, Mang Ye, Zekun Shi, He Li, and Bo Du. Rethinking federated learning with domain shift: A prototype view. In CVPR, pages 16312–16322. IEEE, 2023

    2023

  11. [19]

    Federated learning for generalization, robustness, fairness: A survey and benchmark

    Wenke Huang, Mang Ye, Zekun Shi, Guancheng Wan, He Li, Bo Du, and Qiang Yang. Federated learning for generalization, robustness, fairness: A survey and benchmark. PAMI, 2024

    2024

  12. [20]

    Personalized cross-silo federated learning on non-iid data

    Yutao Huang, Lingyang Chu, Zirui Zhou, Lanjun Wang, Jiangchuan Liu, Jian Pei, and Yong Zhang. Personalized cross-silo federated learning on non-iid data. In AAAI, volume 35, pages 7865–7873, 2021

    2021

  13. [21]

    Personalized cross-silo federated learning on non-iid data

    Yutao Huang, Lingyang Chu, Zirui Zhou, Lanjun Wang, Jiangchuan Liu, Jian Pei, and Yong Zhang. Personalized cross-silo federated learning on non-iid data. In AAAI, volume 35, pages 7865–7873, 2021. 10

    2021

  14. [22]

    Editing models with task arithmetic

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. In ICLR, 2022

    2022

  15. [23]

    Scaffold: Stochastic controlled averaging for federated learning

    Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In ICML, pages 5132–5143. PMLR, 2020

    2020

  16. [24]

    Communication-efficient federated learning with accelerated client gradient

    Geeho Kim, Jinkyu Kim, and Bohyung Han. Communication-efficient federated learning with accelerated client gradient. In CVPR, pages 12385–12394, 2024

    2024

  17. [25]

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

    Iasonas Kokkinos. Ubernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In CVPR, pages 6129–6138, 2017

    2017

  18. [26]

    Model-contrastive federated learning

    Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In CVPR, pages 10713–10722, 2021

    2021

  19. [27]

    Ditto: Fair and robust federated learning through personalization

    Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. Ditto: Fair and robust federated learning through personalization. In ICML, pages 6357–6368. PMLR, 2021

    2021

  20. [28]

    Federated optimization in heterogeneous networks

    Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. In Proceedings of Machine learning and systems , volume 2, pages 429–450, 2020

    2020

  21. [29]

    Fedbn: Federated learning on non-iid features via local batch normalization

    Xiaoxiao Li, Meirui JIANG, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fedbn: Federated learning on non-iid features via local batch normalization. In ICLR, 2021

    2021

  22. [30]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012–10022, 2021

    2021

  23. [31]

    Fully convolutional networks for semantic segmenta- tion

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmenta- tion. In CVPR, pages 3431–3440, 2015

    2015

  24. [32]

    Fedhca2: Towards hetero-client federated multi-task learning

    Yuxiang Lu, Suizhi Huang, Yuwen Yang, Shalayiding Sirejiding, Yue Ding, and Hongtao Lu. Fedhca2: Towards hetero-client federated multi-task learning. In CVPR, pages 5599–5609, 2024

    2024

  25. [33]

    Attentive single-tasking of multiple tasks

    Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In CVPR, pages 1851–1860, 2019

    2019

  26. [34]

    Federated multi-task learning under a mixture of distributions

    Othmane Marfoq, Giovanni Neglia, Aurélien Bellet, Laetitia Kameni, and Richard Vidal. Federated multi-task learning under a mixture of distributions. In NeurIPS, volume 34, pages 15434–15447, 2021

    2021

  27. [35]

    Communication-efficient learning of deep networks from decentralized data

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. InAISTATS, pages 1273–1282. PMLR, 2017

    2017

  28. [36]

    Fedseg: Class-heterogeneous federated learning for semantic segmentation

    Jiaxu Miao, Zongxin Yang, Leilei Fan, and Yi Yang. Fedseg: Class-heterogeneous federated learning for semantic segmentation. In CVPR, pages 8042–8052, 2023

    2023

  29. [38]

    Multi-task federated learning for personalised deep neural networks in edge computing

    Jed Mills, Jia Hu, and Geyong Min. Multi-task federated learning for personalised deep neural networks in edge computing. TPDS, 33(3):630–641, 2021

    2021

  30. [39]

    The role of context for object detection and semantic segmentation in the wild

    Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In CVPR, pages 891–898, 2014

    2014

  31. [40]

    Federated split task- agnostic vision transformer for covid-19 cxr diagnosis

    Sangjoon Park, Gwanghyun Kim, Jeongsol Kim, Boah Kim, and Jong Chul Ye. Federated split task- agnostic vision transformer for covid-19 cxr diagnosis. volume 34, pages 24617–24630, 2021

    2021

  32. [41]

    Rethinking architecture design for tackling data heterogeneity in federated learning

    Liangqiong Qu, Yuyin Zhou, Paul Pu Liang, Yingda Xia, Feifei Wang, Ehsan Adeli, Li Fei-Fei, and Daniel Rubin. Rethinking architecture design for tackling data heterogeneity in federated learning. In CVPR, pages 10061–10071, 2022

    2022

  33. [42]

    Adaptive federated optimization

    Sashank J Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Koneˇcn`y, Sanjiv Kumar, and Hugh Brendan McMahan. Adaptive federated optimization. In ICLR, 2021

    2021

  34. [43]

    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 ECCV, pages 746–760. Springer, 2012. 11

    2012

  35. [44]

    Federated multi-task learning

    Virginia Smith, Chao-Kai Chiang, Maziar Sanjabi, and Ameet S Talwalkar. Federated multi-task learning. In NeurIPS, volume 30, 2017

    2017

  36. [45]

    Mti-net: Multi-scale task interaction networks for multi-task learning

    Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In ECCV, pages 527–543. Springer, 2020

    2020

  37. [46]

    Tackling the objective incon- sistency problem in heterogeneous federated optimization

    Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective incon- sistency problem in heterogeneous federated optimization. In NeurIPS, volume 33, pages 7611–7623, 2020

    2020

  38. [47]

    Pad-net: Multi-tasks guided prediction-and- distillation network for simultaneous depth estimation and scene parsing

    Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and- distillation network for simultaneous depth estimation and scene parsing. In CVPR, pages 675–684, 2018

    2018

  39. [48]

    Federated machine learning: Concept and applications

    Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1–19, 2019

    2019

  40. [49]

    Heterogeneous federated learning: State-of-the-art and research challenges

    Mang Ye, Xiuwen Fang, Bo Du, Pong C Yuen, and Dacheng Tao. Heterogeneous federated learning: State-of-the-art and research challenges. ACM Computing Surveys, 56(3):1–44, 2023

    2023

  41. [50]

    Unleashing the power of multi-task learning: A comprehensive survey spanning traditional, deep, and pretrained foundation model eras

    Jun Yu, Yutong Dai, Xiaokang Liu, Jin Huang, Yishan Shen, Ke Zhang, Rong Zhou, Eashan Adhikarla, Wenxuan Ye, Yixin Liu, et al. Unleashing the power of multi-task learning: A comprehensive survey spanning traditional, deep, and pretrained foundation model eras. CoRR, 2024

    2024

  42. [51]

    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. In NeurIPS, volume 33, pages 5824–5836, 2020

    2020

  43. [52]

    Achievement-based training progress balancing for multi-task learning

    Hayoung Yun and Hanjoo Cho. Achievement-based training progress balancing for multi-task learning. In ICCV, pages 16935–16944, 2023

    2023

  44. [53]

    Federated learning for non-iid data via unified feature learning and optimization objective alignment

    Lin Zhang, Yong Luo, Yan Bai, Bo Du, and Ling-Yu Duan. Federated learning for non-iid data via unified feature learning and optimization objective alignment. In ICCV, pages 4420–4428, 2021

    2021

  45. [54]

    Federated domain generalization with generalization adjustment

    Ruipeng Zhang, Qinwei Xu, Jiangchao Yao, Ya Zhang, Qi Tian, and Yanfeng Wang. Federated domain generalization with generalization adjustment. In CVPR, pages 3954–3963, 2023

    2023

  46. [55]

    A survey on multi-task learning

    Yu Zhang and Qiang Yang. A survey on multi-task learning. TKDE, 34(12):5586–5609, 2021. 12 A Datasets and Preprocessing Details We conduct experiments on two widely-used multi-task learning benchmarks: NYUD-V2 [43] and PASCAL-Context [39]. Both datasets are processed following...

    2021

Pith tools

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