Pith. sign in

REVIEW 4 major objections 5 minor 86 references

FOCoOp: Enhancing Out-of-Distribution Robustness in Federated Prompt Learning for Vision-Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims FOCoOp, a federated prompt-learning framework for vision-language models, resolves the accuracy-versus-out-of-distribution-robustness trade-off by training three prompt sets and aligning them with semi-unbalanced optimal…

desk verdict Impressive empirical breadth and a plausible new architecture, but the headline OOD gains are confounded by metric alignment with the training loss, and the theory is not yet rigorous. read the letter →

arxiv 2506.16218 v3 pith:63ZIIKF3 submitted 2025-06-19 cs.CV

Xinting Liao , Weiming Liu , Jiaming Qian , Pengyang Zhou , Jiahe Xu , Wenjie Wang , Chaochao Chen , Xiaolin Zheng
show 1 more author
Tat-Seng Chua
This is my paper
classification cs.CV
keywords federatedpromptlearningout-of-distributiondetectionvision-languagemodelsdistributionallyrobustoptimizationoptimaltransportdataheterogeneityCLIP
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

Federated prompt learning lets many clients adapt a pretrained vision-language model by sharing only text prompts, but existing methods sacrifice out-of-distribution (OOD) robustness when they raise accuracy. FOCoOp claims to break that trade-off by training three prompt sets: shared in-distribution (ID) global prompts, per-client local prompts, and OOD prompts designed to mismatch ID images. A bi-level distributionally robust optimization widens class-level and distribution-level separation locally, while a server-side semi-unbalanced optimal transport step calibrates OOD prompts against global prompts so clients agree on what counts as novel. On CIFAR-100 with pathological non-overlap, FOCoOp reports the best accuracy (93.85) and the best OOD detection (FPR95 19.50, AUROC 95.42) among seven baselines. If these gains hold, federated deployments of vision-language models can detect novel inputs without giving up predictive performance.

What carries the argument

The load-bearing object is a three-set prompt architecture: ID global prompts $T^g$, client-local prompts $T^l_k$, and OOD prompts $T^o$, all of which are text-context vectors fed to a frozen vision-language model. Class-level separation comes from the prediction loss $\ell_P$ (Eq. 3), which uses a fused prompt $t_c = (1-\rho)t^l_c + \rho t^g_c$ and a softmax denominator that includes all OOD prompt scores; distribution-level separation comes from $\ell_D$ (Eq. 4), which pushes the summed ID-prompt score above the summed OOD-prompt score. Bi-level distributionally robust optimization (Eq. 6) turns both objectives into minimax problems, perturbing global prompts under an optimal-transport divergence and OOD prompts under an unbalanced optimal-transport divergence. The server then solves a semi-unbalanced optimal transport problem (Eq. 8) via Frank-Wolfe to produce a coupling between global and OOD prompts, calibrating global prompts by Eq. (10) and reselecting OOD prompts by Eq. (11). These pieces carry the argument because each corresponds to a failure mode the paper identifies: local overfitting, missing distribution-level separation, and inconsistent OOD judgments across clients.

What would settle it

Re-run the CIFAR-100 comparison with a fixed OOD score that does not include the trained OOD prompts in the denominator, such as energy or maximum-ID-logit scoring, and check whether FOCoOp's FPR95 lead over FedGalLoP and PromptFolio persists.

Watch

Extended reading notes

Core claim

The central claim is that OOD robustness in federated prompt learning should be engineered as two explicit separations: class-level separation between ID classes and distribution-level separation between ID and OOD data, maintained consistently across clients. FOCoOp trains ID global prompts and local prompts for generalization and personalization, plus OOD prompts sampled from WordNet negative labels for detection. Locally, the bi-level objective Eq. (5) maximizes the probability that an image matches its class while suppressing OOD prompt matches; the distributionally robust version (Eq. 6) perturbs global and OOD prompts to worst-case positions under optimal-transport constraints so limited local data still yields wide separation. On the server, semi-unbalanced optimal transport aligns all clients' OOD prompts with the aggregated global prompts, uses the closest 'seemly OOD' prompts to calibrate global prompts, and sends the most distant ones back as fresh OOD prompts. The paper reports that this resolves the trade-off, with CIFAR-100 pathological non-overlap accuracy at 93.85, covariate-shift accuracy at 91.47, FPR95 at 19.50, and AUROC at 95.42, and that both modules are needed since removing either degrades results.

Load-bearing premise

The reported OOD detection advantage assumes that the MCM score computed with the model's own 100 OOD prompts is a fair yardstick; if that scoring convention favors any model that adds extra OOD prompts to the denominator, the comparison overstates the gain.

Editorial extensions

If this is right

  • On CIFAR-100 pathological non-overlap, FOCoOp reports ACC 93.85 and FPR95 19.50, roughly halving the false-positive rate of the best baseline while raising accuracy.
  • Both modules matter: removing the bi-level OOD separations or the global-view OOD consistency degrades detection, and removing the server module causes the larger drop.
  • The gains are claimed to persist across heterogeneity settings (Dirichlet and pathological, 10 and 100 clients), covariate shifts, domain generalization, and five additional label-shift datasets.
  • Only prompts are exchanged between clients and server, so the OOD robustness improvement does not require communicating model weights.
  • OOD prompts initialized from WordNet negative labels and then recalibrated on the server keep clients from mislabeling each other's unseen data as outliers.

Reading between the lines

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

  • Editorial inference: the reported OOD gains may be partly attributable to the 100 OOD prompts sitting in the softmax denominator (Eq. 3) of the scoring rule; baselines that do not add such prompts to their denominator could look weaker under the MCM metric regardless of true separation.
  • Editorial inference: the same three-prompt mechanism with bi-level distributionally robust optimization and semi-unbalanced optimal transport could transfer to other parameter-efficient federated fine-tuning methods, such as LoRA or adapter tuning, beyond textual prompts.
  • Editorial inference: the server calibration might be achievable with simpler matching algorithms, for example nearest-neighbor alignment instead of optimal transport; the paper does not test whether the transport geometry is essential to the consistency gain.
  • Editorial inference: evaluating detection with an independent OOD score that does not use the trained OOD prompts in the denominator would clarify whether the method achieves genuine geometric separation or a scoring convention advantage.
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 FOCoOp, a federated prompt learning method for CLIP that trains three sets of prompts — global ID prompts, local ID prompts, and OOD prompts — using a bi-level distributionally robust optimization (BDRO) loss at the client and a semi-unbalanced optimal transport (SemiUOT) calibration at the server. The central empirical claim is that FOCoOp simultaneously improves ID accuracy, covariate-shift accuracy, and OOD detection (AUROC/FPR95) relative to seven federated baselines across 15 datasets. The paper also ablates the BOS and GOC modules and provides sensitivity analyses over shot counts and hyperparameters.

Significance. If the empirical claims hold, the paper would be a meaningful contribution: it targets OOD robustness specifically in federated prompt learning for VLMs, a setting where the performance/robustness trade-off is real and understudied. The evaluation is broad (15 datasets, two heterogeneity types, cross-device and cross-silo settings), and the consistent gains across many tables, including the ablation variants, suggest the method is not trivially overfit to one benchmark. The paper also includes a full theoretical appendix and a comparison set that mixes FPL methods with adapted centralized OOD methods, which is appropriate. The main weaknesses are that the theoretical derivation of the BDRO updates contains a mathematically invalid step, the stated objective in Eq. (1) does not match the implemented loss, and the OOD evaluation protocol may be aligned with the training objective in a way that favors FOCoOp over baselines. These issues are load-bearing for the paper's central claims and need to be resolved before the contribution can be fully credited.

major comments (4)
  1. [Appendix C.3, Eq. (19), and Theorem 3.1] The proof of Theorem 3.1/Theorem C.3 is not valid. The third line of Eq. (19) replaces the OT discrepancy D_OT(\hat P, P0) — which by Definition C.1 is an infimum over couplings — with the product expectation E_{\hat t~\hat P} E_{t~P0} c(\hat t, t). These two quantities are not equal; the product expectation corresponds to the independent coupling, not the optimal transport plan. The subsequent line also drops the inf over τ1 and the term τ1η1 without justification. The claimed solution forms in Eq. (7) and the update rules in Algorithm 2 therefore do not follow from the stated optimization problem. In addition, the main text states Theorem 3.1 under 'continuous and differentiable' while Appendix Theorem C.3 assumes 'concave and differentiable', and no concavity of L in the prompt arguments is established. Please provide a correct derivation, or weaken the theorem to a heuristic motivation for the BDRO updates.
  2. [Section 3.1, Eq. (1)] The stated global objective includes an explicit out-of-distribution detection loss ℓO(D_OOD, T^l_k, T^g, T^o), but no OOD training data D_OOD is ever defined or used: Eq. (5), Algorithm 2, and the local update only consume ID samples x ∈ D and WordNet-initialized OOD prompts. Either specify how ℓO and D_OOD are constructed, or remove ℓO from Eq. (1) so that the stated objective matches the implemented loss.
  3. [Section 3.2, Eqs. (3)-(4), and Section 4.1] The test-time OOD score is called MCM, but its exact formula is never given. The training probabilities in Eqs. (3)-(4) use a denominator that includes the 100 OOD prompts; if the MCM score at evaluation is computed from the same denominator, the reported FPR95/AUROC comparisons against baselines that do not add OOD prompts to the denominator are not made on the same yardstick. Please state the test-time MCM formula explicitly. If the OOD-prompt denominator is used, include a control that augments the MCM denominator with fixed WordNet initializations of the same 100 OOD prompts but without the ℓD/BDRO/GOC training, to separate the effect of the score change from the effect of learned OOD separation.
  4. [Section 4.1 and Tables 1-5] The paper reports averages over three random seeds but no standard deviations or confidence intervals. Because the FPR95 differences on the headline CIFAR-100 row are large (e.g., 19.50 vs. 38.22 in Table 1), the absence of variance is not fatal to the qualitative ordering, but it leaves the sensitivity of the smaller gaps (e.g., TinyImageNet ACC 88.35 vs. 88.03 in Table 1) unquantified. Please report error bars or per-seed results.
minor comments (5)
  1. [Abstract, Section 3.3, Fig. 2 caption, Section 4.1] The phrase 'seemly OOD prompts' should be 'seemingly OOD prompts', and 'Pathlogical' should be 'Pathological'.
  2. [Section 4.2 and Fig. 7(a-b)] The text says 'optimal threshold ρ = 2', but the sensitivity grid in Fig. 7(a-b) is ρ ∈ {0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.8}; this should be ρ = 0.2 or the grid should be corrected.
  3. [Section 2.1 and Appendix D] The related work refers to 'FedFolio' while the experiments use 'PromptFolio'; please unify the naming. Similarly, Appendix D lists pFedprompt among comparison methods, but it does not appear in the main tables.
  4. [Appendix C, Eq. (8)] The KL term in Eq. (8) is written with a sum over c outside the bracket, while the argument π^T 1 is indexed by j; the notation should be aligned with the definition of KL(π^T 1 ‖ a).
  5. [Algorithm 2, lines 5 and 12] The gradients are taken of an expression containing 'arg min' with no indication of which variables are held fixed; the pseudocode should be rewritten to make the differentiation variables explicit.

Circularity Check

1 steps flagged · score 3.0 of 10

OOD-detection gains are partly a re-statement of the ℓD training objective because MCM shares the OOD-prompt softmax denominator.

  1. fitted input called prediction [Section 3.2, Eqs. (3)-(4); Section 4.1, Evaluation Metrics (MCM)]
    "Similarly, we encourage distribution-level separation, by optimizing the loss for prediction probability of all ID prompts that are larger than OOD prompts, i.e., ℓD = Ex∈D [− log p(yID = 1|x)] , (4) where p(yID = 1|x) = PC c=1 S(x,tc)PC c=1 S(x,tc)+PU u=1 S(x,tou) ... We compute maximum concept matching (MCM) (Ming et al., 2022) as OOD detection score, which is based on similarity between textual features and image features."

    MCM is the max-over-classes of the prediction probability defined in Eq. (3), whose denominator explicitly includes the trained OOD-prompt scores. Eq. (4)'s ℓD directly maximizes p(yID=1|x) = Σ_c S(x,t_c) / (Σ_c S(x,t_c) + Σ_u S(x,t_o_u)) on ID data. The reported OOD-detection metrics (FPR95, AUROC) therefore measure, by construction, the very ID-vs-OOD score ratio FOCoOp was trained to maximize; baselines that do not place OOD prompts in the MCM denominator, or that never train those scores down with ℓD, are not an equally matched yardstick. Part of the AUROC/FPR95 gain is a direct effect of the trained OOD-prompt set used in the score, not evidence of better OOD representation. The paper provides no control keeping the MCM denominator fixed while ablating the ℓD-trained OOD prompts.

full rationale

The paper's main empirical claim is benchmarked against seven external baselines on ACC/CACC/FPR95/AUROC and on domain-transfer tables, so the central comparison is not globally circular, and no load-bearing self-citation chain appears. Self-citations to FOOGD (Liao et al., 2024b) and to the authors' UOT works are motivational and background, not reductions. The proof of Theorem 3.1 has a support gap rather than a circularity: the main text assumes continuity and differentiability, while Supplement Theorem C.3 assumes concavity and never proves it; this needs fixing but does not make the DRO solution equivalent to its input. The one substantive circularity concern is evaluation-score alignment: Eq. (4) trains FOCoOp to maximize p(yID=1|x) = Σ_ID S / (Σ_ID S + Σ_OOD S), and MCM is computed from the same softmax-denominator family defined in Eq. (3), which contains the trained OOD prompts. Consequently, part of the reported FPR95/AUROC margin over baselines without OOD-prompt denominators re-states the training objective rather than providing independent evidence, and the missing matched-denominator control means the headline trade-off claim is only partially supported.

Assumptions & free parameters 7 free parameters · 4 assumptions · 1 invented entities

The central claim rests on several tuned hyperparameters (ρ, α, τ1, τ2, U, M, γ) and on domain assumptions about the OOD candidate pool and the scoring function. None of these are derived from first principles.

free parameters (7)
  • ρ = swept over [0, 0.8]; text mentions optimal 2 (likely typo)
    Fusion weight between local and global ID prompts, tuned per dataset via sensitivity analysis (Fig. 7).
  • α = swept over [0, 1.0]
    EMA coefficient for global prompt calibration in Eq. (10); tuned per dataset.
  • τ1 = swept over [0.1, 10]
    DRO penalty for global prompt perturbation; trade-off point selected via sensitivity (Fig. 7e).
  • τ2 = swept over [0.1, 10]
    DRO penalty for OOD prompt perturbation; higher values favor detection (Fig. 7f).
  • U = 100
    Number of OOD prompts; fixed across experiments.
  • M = not specified
    Number of seemly OOD prompts selected for calibration in Section 3.3; the value is never reported.
  • γ = not specified
    L1 penalty on perturbation noise in Algorithm 2; no value is given.
assumptions (4)
  • domain assumption WordNet supplies a sufficient lexical pool of OOD candidate classes for distribution-level separation
    OOD prompts are initialized by selecting top U negative labels from WordNet (Eq. 2 and Section 3.2); if the pool lacks words relevant to the OOD test distributions, detection will fail.
  • domain assumption MCM with the modified softmax denominator is a valid OOD detection score
    The evaluation (Section 4.1) uses MCM based on similarities with the same OOD prompts that appear in the training loss denominator, aligning metric and objective.
  • ad hoc to paper The product-distribution expectation equals the OT distance in the dual derivation
    Appendix Eq. (19) substitutes E_{hat P × P0} c for D_OT(P, P0), which is not the OT cost; the derivation of Theorem 3.1 depends on this identity.
  • domain assumption All clients share the identical class label space
    Global and local prompts are defined per class and aggregated across clients; the method does not address label-space heterogeneity or open-set class shifts.
invented entities (1)
  • OOD prompts T^o
    purpose: A set of 100 learnable context vectors tuned to mismatch local ID data, providing distribution-level separation for OOD detection.
    These are new model parameters introduced by the method; their effectiveness is only demonstrated within the paper's own evaluation, with no falsifiable handle outside the reported datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FOCoOp: Enhancing Out-of-Distribution Robustness in Federated Prompt Learning for Vision-Language Models." pith.science (2026). https://pith.science/paper/63ZIIKF3

@misc{pith2026250616218,
  author       = {Pith},
  title        = {Pith review of: FOCoOp: Enhancing Out-of-Distribution Robustness in Federated Prompt Learning for Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/63ZIIKF3}},
  note         = {Machine review of arXiv:2506.16218}
}
read the original abstract

Federated prompt learning (FPL) for vision-language models is a powerful approach to collaboratively adapt models across distributed clients while preserving data privacy. However, existing FPL approaches suffer from a trade-off between performance and robustness, particularly in out-of-distribution (OOD) shifts, limiting their reliability in real-world scenarios. The inherent in-distribution (ID) data heterogeneity among different clients makes it more challenging to maintain this trade-off. To fill this gap, we introduce a Federated OOD-aware Context Optimization (FOCoOp) framework, which captures diverse distributions among clients using ID global prompts, local prompts, and OOD prompts. Specifically, FOCoOp leverages three sets of prompts to create both class-level and distribution-level separations, which adapt to OOD shifts through bi-level distributionally robust optimization. Additionally, FOCoOp improves the discrimination consistency among clients, i.e., calibrating global prompts, seemingly OOD prompts, and OOD prompts by semi-unbalanced optimal transport. The extensive experiments on real-world datasets demonstrate that FOCoOp effectively captures decentralized heterogeneous distributions and enhances robustness of different OOD shifts. The project is available at GitHub.

Figures

Figures reproduced from arXiv: 2506.16218 by the authors.

Figure 1
Figure 1. The performance and OOD robustness of FPL methods. The FPR95 reflects detection capability, while ACC and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Framework of FOCoOp. For each client, FOCoOp uses bi-level OOD separations module to fine-tune three sets of prompts adapting to pretrained VLM. While in the server, FOCoOp leverages the global-view OOD consistency module to enhance the discrimination among ID global prompts and OOD prompts. 3.2. Client: Bi-Level OOD Separations For the sake of limited access to training data, it mainly includes two aspects of OOD r… view at source ↗
Figure 3
Figure 3. The average generalization results on data with different covariate-shifts [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Detection comparison on CIFAR-100. 1 2 4 8 16 Number of shots 50.0 60.0 70.0 80.0 90.0 Performance (%) (a) CIFAR100 Comparison 1 2 4 8 16 Number of shots 40.0 50.0 60.0 70.0 80.0 90.0 Performance (%) (b) TinyImageNet Comparison PromptFL FedOTP FedPGP PromptFolio FedLoC…
Figure 7
Figure 7. Figure 7: Hyperparameter sensitivity studies. Method capability on other datasets. To comprehensively evaluate the OOD robustness of FPL methods, we present the average ID-C covariate shift generalization in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Detection Comparison on CIFAR-100. PromptFL FedOTP FedPGP PromptFolio FedLoCoOp FedGalLoP FedLAPT FOCoOp 0 20 40 60 80 FPR95 (%) (a) FPR95 Comparison PromptFL FedOTP FedPGP PromptFolio FedLoCoOp FedGalLoP FedLAPT FOCoOp 0 20 40 60 80 100 AUROC (%) (b) AUROC Comparison …
Figure 9
Figure 9. Figure 9: Detection comparison on TinyImageNet [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: Hyperparameter sensitivity studies. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

86 extracted references · 59 canonical work pages

  1. [1]

    Minimization of functions having lipschitz continuous first partial derivatives

    Armijo, L. Minimization of functions having lipschitz continuous first partial derivatives. Pacific Journal of mathematics, 16 0 (1): 0 1--3, 1966

    1966

  2. [2]

    D., and Li, Y

    Bai, H., Canal, G., Du, X., Kwon, J., Nowak, R. D., and Li, Y. Feed two birds with one scone: Exploiting wild data for both out-of-distribution generalization and detection. In International Conference on Machine Learning, pp.\ 1454--1471. PMLR, 2023

    2023

  3. [3]

    Diprompt: Disentangled prompt tuning for multiple latent domain generalization in federated learning

    Bai, S., Zhang, J., Guo, S., Li, S., Guo, J., Hou, J., Han, T., and Lu, X. Diprompt: Disentangled prompt tuning for multiple latent domain generalization in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 27284--27293, 2024 a

    2024

  4. [4]

    Id-like prompt learning for few-shot out-of-distribution detection

    Bai, Y., Han, Z., Cao, B., Jiang, X., Hu, Q., and Zhang, C. Id-like prompt learning for few-shot out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 17480--17489, 2024 b

    2024

  5. [5]

    Food-101--mining discriminative components with random forests

    Bossard, L., Guillaumin, M., and Van Gool, L. Food-101--mining discriminative components with random forests. In Computer vision--ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13, pp.\ 446--461. Springer, 2014

    2014

  6. [6]

    A Unified Wasserstein Distributional Robustness Framework for Adversarial Training

    Bui, T. A., Le, T., Tran, Q., Zhao, H., and Phung, D. A unified wasserstein distributional robustness framework for adversarial training. arXiv preprint arXiv:2202.13437, 2022

    work page Pith review arXiv 2022

  7. [7]

    and Chao, W.-L

    Chen, H.-Y. and Chao, W.-L. On bridging generic and personalized federated learning for image classification. In International Conference on Learning Representations, 2021

    2021

  8. [8]

    Describing textures in the wild

    Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., and Vedaldi, A. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 3606--3613, 2014 a

    work page 2014

Show all 86 references
  1. [9]

    Describing textures in the wild

    Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., and Vedaldi, A. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 3606--3613, 2014 b

    2014

  2. [10]

    Clarkson, K. L. Coresets, sparse greedy approximation, and the frank-wolfe algorithm. ACM Transactions on Algorithms (TALG), 6 0 (4): 0 1--30, 2010

    2010

  3. [11]

    Harmonizing generalization and personalization in federated prompt learning

    Cui, T., Li, H., Wang, J., and Shi, Y. Harmonizing generalization and personalization in federated prompt learning. In Salakhutdinov, R., Kolter, Z., Heller, K., Weller, A., Oliver, N., Scarlett, J., and Berkenkamp, F. (eds.), Proceedings of the 41st International Conference o...

    2024

  4. [12]

    Q., Li, A., and Kung, H

    Dong, X., Zhang, S. Q., Li, A., and Kung, H. Spherefed: Hyperspherical federated learning. In ECCV, pp.\ 165--184. Springer, 2022

    2022

  5. [13]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Dosovitskiy, A. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  6. [14]

    Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories

    Fei-Fei, L., Fergus, R., and Perona, P. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pp.\ 178--178. IEEE, 2004

    2004

  7. [15]

    Wordnet: An electronic lexical database

    Fellbaum, C. Wordnet: An electronic lexical database. MIT Press google schola, 2: 0 678--686, 1998

    1998

  8. [16]

    Learning federated visual prompt in null space for mri reconstruction

    Feng, C.-M., Li, B., Xu, X., Liu, Y., Fu, H., and Zuo, W. Learning federated visual prompt in null space for mri reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 8064--8073, 2023

    2023

  9. [17]

    Geodesic flow kernel for unsupervised domain adaptation

    Gong, B., Shi, Y., Sha, F., and Grauman, K. Geodesic flow kernel for unsupervised domain adaptation. In 2012 IEEE conference on computer vision and pattern recognition, pp.\ 2066--2073. IEEE, 2012

    2012

  10. [18]

    Pfedprompt: Learning personalized prompt for vision-language models in federated learning

    Guo, T., Guo, S., and Wang, J. Pfedprompt: Learning personalized prompt for vision-language models in federated learning. In Proceedings of the ACM Web Conference 2023, pp.\ 1364--1374, 2023 a

    2023

  11. [19]

    Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model

    Guo, T., Guo, S., Wang, J., Tang, X., and Xu, W. Promptfl: Let federated participants cooperatively learn prompts instead of models-federated learning in age of foundation model. IEEE Transactions on Mobile Computing, 2023 b

    2023

  12. [20]

    Out-of-distribution generalization of federated learning via implicit invariant relationships

    Guo, Y., Guo, K., Cao, X., Wu, T., and Chang, Y. Out-of-distribution generalization of federated learning via implicit invariant relationships. In International Conference on Machine Learning, pp.\ 11905--11933. PMLR, 2023 c

    2023

  13. [21]

    Deep residual learning for image recognition

    He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 770--778, 2016

    2016

  14. [22]

    and Dietterich, T

    Hendrycks, D. and Dietterich, T. G. Benchmarking neural network robustness to common corruptions and surface variations. arXiv preprint arXiv:1807.01697, 2018

    2018 arXiv

  15. [23]

    and Gimpel, K

    Hendrycks, D. and Gimpel, K. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2016

    2016

  16. [24]

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

    Huang, W., Ye, M., Shi, Z., Wan, G., Li, H., Du, B., and Yang, Q. Federated learning for generalization, robustness, fairness: A survey and benchmark. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  17. [25]

    Revisiting frank-wolfe: Projection-free sparse convex optimization

    Jaggi, M. Revisiting frank-wolfe: Projection-free sparse convex optimization. In International conference on machine learning, pp.\ 427--435. PMLR, 2013

    2013

  18. [26]

    Scaling up visual and vision-language representation learning with noisy text supervision

    Jia, C., Yang, Y., Xia, Y., Chen, Y.-T., Parekh, Z., Pham, H., Le, Q., Sung, Y.-H., Li, Z., and Duerig, T. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pp.\ 4904--4916. PMLR, 2021

    2021

  19. [27]

    and Lin, T

    Jiang, L. and Lin, T. Test-time robust personalization for federated learning. In The Eleventh International Conference on Learning Representations, 2022

    2022

  20. [28]

    Negative label guided ood detection with pretrained vision-language models

    Jiang, X., Liu, F., Fang, Z., Chen, H., Liu, T., Zheng, F., and Han, B. Negative label guided ood detection with pretrained vision-language models. In The Twelfth International Conference on Learning Representations, 2024

    2024

  21. [29]

    P., Kale, S., Mohri, M., Reddi, S., Stich, S., and Suresh, A

    Karimireddy, S. P., Kale, S., Mohri, M., Reddi, S., Stich, S., and Suresh, A. T. Scaffold: Stochastic controlled averaging for federated learning. In International conference on machine learning, pp.\ 5132--5143. PMLR, 2020

    2020

  22. [30]

    Learning multiple layers of features from tiny images

    Krizhevsky, A., Hinton, G., et al. Learning multiple layers of features from tiny images. 2009

    2009

  23. [31]

    Fine-tuning can distort pretrained features and underperform out-of-distribution

    Kumar, A., Raghunathan, A., Jones, R., Ma, T., and Liang, P. Fine-tuning can distort pretrained features and underperform out-of-distribution. arXiv preprint arXiv:2202.10054, 2022

    2022 arXiv

  24. [32]

    Gallop: Learning global and local prompts for vision-language models

    Lafon, M., Ramzi, E., Rambour, C., Audebert, N., and Thome, N. Gallop: Learning global and local prompts for vision-language models. In European Conference on Computer Vision, pp.\ 264--282. Springer, 2025

    2025

  25. [33]

    and Yang, X

    Le, Y. and Yang, X. Tiny imagenet visual recognition challenge. CS 231N, 7 0 (7): 0 3, 2015

    2015

  26. [34]

    Li, D., Yang, Y., Song, Y.-Z., and Hospedales, T. M. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pp.\ 5542--5550, 2017

    2017

  27. [35]

    Global and local prompts cooperation via optimal transport for federated learning

    Li, H., Huang, W., Wang, J., and Shi, Y. Global and local prompts cooperation via optimal transport for federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 12151--12161, 2024

    2024

  28. [36]

    Model-contrastive federated learning

    Li, Q., He, B., and Song, D. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10713--10722, 2021

    2021

  29. [37]

    K., Zaheer, M., Sanjabi, M., Talwalkar, A., and Smith, V

    Li, T., Sahu, A. K., Zaheer, M., Sanjabi, M., Talwalkar, A., and Smith, V. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2: 0 429--450, 2020

    2020

  30. [38]

    and Wang, J

    Li, X. and Wang, J. Position paper: Assessing robustness, privacy, and fairness in federated learning integrated with foundation models. arXiv preprint arXiv:2402.01857, 2024

    2024

  31. [39]

    Hyperfed: hyperbolic prototypes exploration with consistent aggregation for non-iid data in federated learning

    Liao, X., Liu, W., Chen, C., Zhou, P., Zhu, H., Tan, Y., Wang, J., and Qi, Y. Hyperfed: hyperbolic prototypes exploration with consistent aggregation for non-iid data in federated learning. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intell...

    2023

  32. [40]

    Rethinking the representation in federated unsupervised learning with non-iid data

    Liao, X., Liu, W., Chen, C., Zhou, P., Yu, F., Zhu, H., Yao, B., Wang, T., Zheng, X., and Tan, Y. Rethinking the representation in federated unsupervised learning with non-iid data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 2284...

    2024

  33. [41]

    Foogd: Federated collaboration for both out-of-distribution generalization and detection

    Liao, X., Liu, W., Zhou, P., Yu, F., Xu, J., Wang, J., Wang, W., Chen, C., and Zheng, X. Foogd: Federated collaboration for both out-of-distribution generalization and detection. arXiv preprint arXiv:2410.11397, 2024 b

    2024 arXiv

  34. [42]

    Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space

    Liu, Q., Chen, C., Qin, J., Dou, Q., and Heng, P.-A. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 1013--1023, 2021

    2021

  35. [43]

    Reducing item discrepancy via differentially private robust embedding alignment for privacy-preserving cross domain recommendation

    Liu, W., Zheng, X., Chen, C., Xu, J., Liao, X., Wang, F., Tan, Y., and Ong, Y.-S. Reducing item discrepancy via differentially private robust embedding alignment for privacy-preserving cross domain recommendation. In Proceedings of the 41st International Conference on Machine ...

    2024

  36. [44]

    Hyperbolic variational graph auto-encoder for next poi recommendation

    Liu, Y., Qi, L., Mao, X., Liu, W., Wang, F., Xu, X., Zhang, X., Dou, W., Zhou, X., and Beheshti, A. Hyperbolic variational graph auto-encoder for next poi recommendation. In Proceedings of the ACM on Web Conference 2025, pp.\ 3267--3275, 2025 a

    2025

  37. [45]

    Guardreasoner-vl: Safeguarding vlms via reinforced reasoning

    Liu, Y., Zhai, S., Du, M., Chen, Y., Cao, T., Gao, H., Wang, C., Li, X., Wang, K., Fang, J., et al. Guardreasoner-vl: Safeguarding vlms via reinforced reasoning. arXiv preprint arXiv:2505.11049, 2025 b

    2025 arXiv

  38. [46]

    Does clip's generalization performance mainly stem from high train-test similarity? arXiv preprint arXiv:2310.09562, 2023

    Mayilvahanan, P., Wiedemer, T., Rusak, E., Bethge, M., and Brendel, W. Does clip's generalization performance mainly stem from high train-test similarity? arXiv preprint arXiv:2310.09562, 2023

    2023 arXiv

  39. [47]

    McMahan, B., Moore, E., Ramage, D., Hampson, S., and y Arcas, B. A. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pp.\ 1273--1282. PMLR, 2017

    2017

  40. [48]

    Delving into out-of-distribution detection with vision-language representations

    Ming, Y., Cai, Z., Gu, J., Sun, Y., Li, W., and Li, Y. Delving into out-of-distribution detection with vision-language representations. Advances in neural information processing systems, 35: 0 35087--35102, 2022

    2022

  41. [49]

    Locoop: Few-shot out-of-distribution detection via prompt learning

    Miyai, A., Yu, Q., Irie, G., and Aizawa, K. Locoop: Few-shot out-of-distribution detection via prompt learning. Advances in Neural Information Processing Systems, 36, 2024

    2024

  42. [50]

    T., Torr, P., and Lim, S

    Nguyen, A. T., Torr, P., and Lim, S. N. Fedsr: A simple and effective domain generalization method for federated learning. Advances in Neural Information Processing Systems, 35: 0 38831--38843, 2022 a

    2022

  43. [51]

    T., Torr, P., and Lim, S

    Nguyen, A. T., Torr, P., and Lim, S. N. Fedsr: A simple and effective domain generalization method for federated learning. Advances in Neural Information Processing Systems, 35: 0 38831--38843, 2022 b

    2022

  44. [52]

    and Zisserman, A

    Nilsback, M.-E. and Zisserman, A. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pp.\ 722--729. IEEE, 2008

    2008

  45. [53]

    Federated learning from vision-language foundation models: Theoretical analysis and method

    Pan, B., Huang, W., and Shi, Y. Federated learning from vision-language foundation models: Theoretical analysis and method. arXiv preprint arXiv:2409.19610, 2024

    2024 arXiv

  46. [54]

    M., Vedaldi, A., Zisserman, A., and Jawahar, C

    Parkhi, O. M., Vedaldi, A., Zisserman, A., and Jawahar, C. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pp.\ 3498--3505. IEEE, 2012

    2012

  47. [55]

    Moment matching for multi-source domain adaptation

    Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., and Wang, B. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 1406--1415, 2019

    2019

  48. [56]

    Counterfactual user sequence synthesis augmented with continuous time dynamic preference modeling for sequential poi recommendation

    Qi, L., Liu, Y., Liu, W., Pei, S., Xu, X., Zhang, X., Wang, Y., and Dou, W. Counterfactual user sequence synthesis augmented with continuous time dynamic preference modeling for sequential poi recommendation. In Proceedings of the Thirty-Third International Joint Conference on...

    2024

  49. [57]

    K., Ganesh, M

    Qiu, C., Li, X., Mummadi, C. K., Ganesh, M. R., Li, Z., Peng, L., and Lin, W.-Y. Text-driven prompt generation for vision-language models in federated learning. In International Workshop on Federated Learning in the Age of Foundation Models in Conjunction with NeurIPS, 2023

    2023

  50. [58]

    Generalized federated learning via sharpness aware minimization

    Qu, Z., Li, X., Duan, R., Liu, Y., Tang, B., and Lu, Z. Generalized federated learning via sharpness aware minimization. In International Conference on Machine Learning, pp.\ 18250--18280. PMLR, 2022

    2022

  51. [59]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp.\ 8748--8763. PMLR, 2021

    2021

  52. [60]

    Clip-guided federated learning on heterogeneity and long-tailed data

    Shi, J., Zheng, S., Yin, X., Lu, Y., Xie, Y., and Qu, Y. Clip-guided federated learning on heterogeneity and long-tailed data. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 14955--14963, 2024

    2024

  53. [61]

    Clipood: Generalizing clip to out-of-distributions

    Shu, Y., Guo, X., Wu, J., Wang, X., Wang, J., and Long, M. Clipood: Generalizing clip to out-of-distributions. In International Conference on Machine Learning, pp.\ 31716--31731. PMLR, 2023

    2023

  54. [62]

    Certifying some distributional robustness with principled adversarial training

    Sinha, A., Namkoong, H., Volpi, R., and Duchi, J. Certifying some distributional robustness with principled adversarial training. arXiv preprint arXiv:1710.10571, 2017

    2017 arXiv

  55. [63]

    Is heterogeneity notorious? taming heterogeneity to handle test-time shift in federated learning

    Tan, Y., Chen, C., Zhuang, W., Dong, X., Lyu, L., and Long, G. Is heterogeneity notorious? taming heterogeneity to handle test-time shift in federated learning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  56. [64]

    Exploiting personalized invariance for better out-of-distribution generalization in federated learning

    Tang, X., Guo, S., and Zhang, J. Exploiting personalized invariance for better out-of-distribution generalization in federated learning. arXiv preprint arXiv:2211.11243, 2022

    2022 arXiv

  57. [65]

    The inaturalist species classification and detection dataset

    Van Horn, G., Mac Aodha, O., Song, Y., Cui, Y., Sun, C., Shepard, A., Adam, H., Perona, P., and Belongie, S. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 8769--8778, 2018

    2018

  58. [66]

    Ce-rcfr: Robust counterfactual regression for consensus-enabled treatment effect estimation

    Wang, F., Chen, C., Liu, W., Fan, T., Liao, X., Tan, Y., Qi, L., and Zheng, X. Ce-rcfr: Robust counterfactual regression for consensus-enabled treatment effect estimation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.\ 3013--3023, 2024 a

    2024

  59. [67]

    Inter- and intra- similarity preserved counterfactual incentive effect estimation for recommendation systems

    Wang, F., Qi, L., Liu, W., Yu, B., Chen, J., and Xu, Y. Inter- and intra- similarity preserved counterfactual incentive effect estimation for recommendation systems. ACM Trans. Inf. Syst., 2025 a . doi:10.1145/3722104

    2025 doi

  60. [68]

    Clipn for zero-shot ood detection: Teaching clip to say no

    Wang, H., Li, Y., Yao, H., and Li, X. Clipn for zero-shot ood detection: Teaching clip to say no. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 1802--1812, 2023

    2023

  61. [69]

    Outlier-robust distributionally robust optimization via unbalanced optimal transport

    Wang, Z. Outlier-robust distributionally robust optimization via unbalanced optimal transport. In Annual Conference on Neural Information Processing Systems, 2025

    2025

  62. [70]

    Flora: Federated fine-tuning large language models with heterogeneous low-rank adaptations

    Wang, Z., Shen, Z., He, Y., Sun, G., Wang, H., Lyu, L., and Li, A. Flora: Federated fine-tuning large language models with heterogeneous low-rank adaptations. In Advances in Neural Information Processing Systems, volume 37, pp.\ 22513--22533. Curran Associates, Inc., 2024 b

    2024

  63. [71]

    M., and Johansson, K

    Wang, Z., Shen, Y., Zavlanos, M. M., and Johansson, K. H. Outlier-robust distributionally robust optimization via unbalanced optimal transport. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2025 b

    2025

  64. [72]

    T., Weng, L., and Hoang, N

    Weng, P.-Y., Hoang, M., Nguyen, L., Thai, M. T., Weng, L., and Hoang, N. Probabilistic federated prompt-tuning with non-iid and imbalanced data. Advances in Neural Information Processing Systems, 37: 0 81933--81958, 2024

    2024

  65. [73]

    A., Oliva, A., and Torralba, A

    Xiao, J., Hays, J., Ehinger, K. A., Oliva, A., and Torralba, A. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pp.\ 3485--3492. IEEE, 2010

    2010

  66. [74]

    A., Zhang, Y., Finkelstein, A., Kulkarni, S

    Xu, P., Ehinger, K. A., Zhang, Y., Finkelstein, A., Kulkarni, S. R., and Xiao, J. Turkergaze: Crowdsourcing saliency with webcam based eye tracking. arXiv preprint arXiv:1504.06755, 2015

    2015 arXiv

  67. [75]

    Yang, F.-E., Wang, C.-Y., and Wang, Y.-C. F. Efficient model personalization in federated learning via client-specific prompt generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 19159--19168, 2023

    2023

  68. [76]

    Generalized out-of-distribution detection: A survey

    Yang, J., Zhou, K., Li, Y., and Liu, Z. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, 132 0 (12): 0 5635--5662, 2024

    2024

  69. [77]

    Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop

    Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., and Xiao, J. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015

    2015 arXiv

  70. [78]

    Turning the curse of heterogeneity in federated learning into a blessing for out-of-distribution detection

    Yu, S., Hong, J., Wang, H., Wang, Z., and Zhou, J. Turning the curse of heterogeneity in federated learning into a blessing for out-of-distribution detection. In 2023 International Conference on Learning Representations, 2023

    2023

  71. [79]

    R., Ning, L., and Singhal, K

    Yuan, H., Morningstar, W. R., Ning, L., and Singhal, K. What do we mean by generalization in federated learning? In International Conference on Learning Representations, 2021

    2021

  72. [80]

    Rethinking misalignment in vision-language model adaptation from a causal perspective

    Zhang, Y., Li, J., Liu, L., and Qiang, W. Rethinking misalignment in vision-language model adaptation from a causal perspective. In Advances in Neural Information Processing Systems, volume 37, pp.\ 39224--39248, 2024

    2024

  73. [81]

    Lapt: Label-driven automated prompt tuning for ood detection with vision-language models

    Zhang, Y., Zhu, W., He, C., and Zhang, L. Lapt: Label-driven automated prompt tuning for ood detection with vision-language models. In European Conference on Computer Vision, pp.\ 271--288. Springer, 2025

    2025

  74. [82]

    Places: A 10 million image database for scene recognition

    Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., and Torralba, A. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40 0 (6): 0 1452--1464, 2017

    2017

  75. [83]

    C., and Liu, Z

    Zhou, K., Yang, J., Loy, C. C., and Liu, Z. Learning to prompt for vision-language models. International Journal of Computer Vision, 130 0 (9): 0 2337--2348, 2022

    2022

  76. [84]

    Fedgog: Federated graph out-of-distribution generalization with diffusion data exploration and latent embedding decorrelation

    Zhou, P., Chen, C., Liu, W., Liao, X., Shen, W., Xu, J., Fu, Z., Wang, J., Wen, W., and Zheng, X. Fedgog: Federated graph out-of-distribution generalization with diffusion data exploration and latent embedding decorrelation. In Proceedings of the AAAI Conference on Artificial ...

    2025

  77. [85]

    Fedgf: Enhancing structural knowledge via graph factorization for federated graph learning

    Zhou, P., Chen, C., Liu, W., Liao, X., Yu, F., Fu, Z., Lou, X., Wen, W., Zheng, X., and Wang, J. Fedgf: Enhancing structural knowledge via graph factorization for federated graph learning. In Proceedings of the Eighteenth ACM International Conference on Web Search and Data Min...

    2025

  78. [86]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.