Pith. sign in

REVIEW 4 major objections 6 minor 57 references

Domain Borders Are There to Be Crossed With Federated Few-Shot Adaptation

T0 review · 4 major / 6 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 FedAcross+ claims a federated client can adapt a pretrained model to a shifted target domain by freezing the backbone and classifier and fine-tuning a single domain-adaptive layer on 3-10 labeled samples per class.

desk verdict A clean static few-shot FL framework, but the paper's only new increment—stream sampling—is never actually run, and every reported number comes from prior work. read the letter →

arxiv 2507.10160 v1 pith:G2VIW6OI submitted 2025-07-14 cs.LG cs.CRcs.DC

Manuel Röder , Christoph Raab , Frank-Michael Schleif This is my paper
classification cs.LGcs.CRcs.DC
keywords FederatedLearningDomainAdaptationFew-shotDeepTransferStreamingDataResourceConstraintsPrototype-basedinferenceSporadicmodelupdates
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

FedAcross+ asks whether a resource-constrained federated client can adapt a shared pretrained model to its own domain with only a handful of labeled examples per class. The paper's claim is that it can: freeze the deep feature extractor and the classifier, fine-tune only a small domain-adaptive linear layer on a $k$-shot support set ($k \in \{3,5,10\}$), and classify by distance to class prototypes. On Office-31 this reaches 89.0% average accuracy at $k = 10$, the best result among the methods compared, and on Office-Home it reaches 70.9% at $k = 10$, within one point of a method that uses all target data. The authors also add a streaming-sampling branch intended for non-stationary environments, but no experiment in the paper exercises that branch; the empirical case that carries the paper is the frozen-backbone, adaptation-layer-only recipe.

What carries the argument

The carrying object is the domain-adaptive linear layer, a task-specific adapter (Eq. 1) that combines a linear map $W,b$ with conditional batch-normalization statistics $\mu,\sigma$ and affine parameters $\gamma,\beta$; it is the only trainable module on the client. The companion mechanism is class-prototype generation (Eq. 4), which averages the $k$ labeled support embeddings per class, and nearest-prototype inference (Eq. 5), which turns prediction into an $L^2$ distance lookup. Together these make client-side backpropagation through the deep backbone unnecessary and compress what the client sends back to the server to a few vectors and adaptation parameters.

What would settle it

Run FedAcross+ on a non-stationary target stream with injected drift, comparing VeSSal-sampled $k$-shot sets against randomly sampled $k$-shot sets at the same labeling budget; if random sampling matches or beats the sampler in downstream target accuracy, the paper's streaming claim is refuted. A second check: find a target domain where updating only the adaptation module fails to improve over no adaptation while updating the backbone or classifier does.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that for cross-device domain shift the bottleneck sits in a single conditional layer, not in the deep network. After server-side pretraining on a source domain, the client receives the feature extractor and classifier frozen and updates only the domain-adaptive linear layer $A(\psi) = ((W\tilde{x} + b) - \mu)/\sigma \cdot \gamma + \beta$ on a balanced $k$-shot support set from the target domain. Class prototypes are then formed as mean adapted embeddings (Eq. 4), and inference is nearest-prototype in Euclidean distance (Eq. 5). With $k = 10$ on Office-31, FedAcross+ averages 89.0% over six domain shifts, the best result among the compared methods, while LCCS leads at $k=5$ with 88.9%; on Office-Home, FedAcross+ reaches 70.9% at $k=10$, within one point of SHOT, which uses all target data. The paper also asserts that a volume-sampling rule (Eq. 6) extends the pipeline to streaming, non-stationary environments, but its experiments keep data provisioning static and defer the sampling claim to prior work.

Load-bearing premise

The load-bearing assumption is that the VeSSal-based sampling rule (Eq. 6), which the paper never tests in its experiments, selects $k$ samples per class that are as informative for fine-tuning the adaptation layer as the static $k$-shot support sets used in the benchmarks.

Editorial extensions

If this is right

  • Clients with severely limited computation can still take part in federated domain adaptation, because no gradient flows through the frozen backbone and the only uploaded objects are adaptation-module parameters and class prototypes.
  • At $k=10$ labels per class, FedAcross+ reaches 89.0% average accuracy on Office-31, the top result in its comparison, and comes within about one point of full-data SHOT on Office-Home (70.9% versus 71.8%), indicating that a small labeled support set can substitute for large unlabeled target data in some domain-shift regimes.
  • Increasing $k$ from 3 to 10 gives operators a direct accuracy-versus-annotation-effort trade-off, with the waste-sorting subsets showing the same scaling (for example, Office-Home Waste rising from 84.42% at $k=3$ to 93.45% at $k=10$ for one direction).
  • Because client knowledge is compressed into prototypes, the server can combine or average them to onboard new clients with small parameter deltas, making sporadic model updates feasible in bandwidth-constrained settings.
  • If the streaming branch works as claimed, the same frozen-backbone recipe would support continuous adaptation on non-stationary sensor data without retraining the deep network on the client.

Reading between the lines

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

  • The paper evaluates only one client adapting at a time, so the multi-client benefit of upstreaming prototypes remains unmeasured; a natural next experiment would aggregate prototypes from several clients and check whether the global adaptation module improves as clients join.
  • The streaming claim is currently an assertion inherited from earlier work; running FedAcross+ on a stream with injected drift and comparing VeSSal-sampled $k$-shot sets against randomly drawn support sets at the same labeling budget is the missing experiment.
  • The reporting of only average accuracies leaves the variance across domain shifts hidden; a per-direction breakdown shows some endpoints (e.g., D→A and D→W on Office-31) underperform centralized baselines, so robustness across shift directions deserves separate study.
  • The privacy argument rests on prototypes being class means in embedding space; stress-testing uploaded prototypes against embedding-space inversion would directly probe a core deployment assumption.
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 presents FedAcross+, a federated few-shot domain adaptation method. The server pre-trains a source model consisting of a feature extractor, a domain-adaptive linear layer, and a classifier. On each client, the feature extractor and classifier are frozen and only the adaptation module is fine-tuned on a k-shot support set; inference is performed by nearest-prototype classification. The claimed new contribution over the prior FedAcross framework is an extension to streaming data: a VeSSal-inspired sampling rule (Eq. 6) is proposed to select items from a client data stream for labelling and inclusion in the support set. The experimental section reports results on Office-31, Office-Home, and waste-sorting subsets of Office-Home and DomainNet, and argues that FedAcross+ achieves competitive accuracy with low client-side cost.

Significance. If the static framework is taken as prior work, the manuscript's practical contribution is the integration of a stream-sampling module into an already published federated few-shot adaptation pipeline. The underlying design is clean, the equations in Sections 3.2-3.4 are internally consistent, and the static method is benchmarked on public datasets against several external baselines. The authors also provide a GitHub repository and use standard frameworks (PyTorch Lightning and Flower), which supports reproducibility. However, the paper's advertised novelty, the streaming extension for non-stationary environments, is never exercised in the experiments, and the reported accuracy numbers are explicitly inherited from the authors' prior paper [36]. The empirical validation is therefore self-referential, and the new claim of suitability for dynamic environments is currently unsupported.

major comments (4)
  1. [§4.1, §3.6, Algorithm 1] The headline extension to streaming data is never evaluated. Section 4.1 states that "data provisioning for both the FL server and the FL client is static" and delegates the efficacy of stream sampling to references [39] and [37]. Algorithm 1's SAMPLING_ENABLED branch (lines 4-7) is therefore not executed in any experiment. Consequently, the abstract's claim that FedAcross+ "is extended to encompass the processing of streaming data, thereby rendering the solution suitable for non-stationary environments" is not supported by the reported results. The authors should either add a streaming/non-stationary experiment that exercises Eq. (6), or substantially weaken the non-stationary claim.
  2. [§3.6, Eq. (6)] There is a missing mechanism between the volume-based sampling probability and the balanced k-shot support set. Eq. (6) defines pt proportional to a determinantal/leverage score, while Section 3.6 asserts that the support set "is constructed to ensure that each category is represented equally, with k samples per category." Section 3.1 additionally requires exactly k annotated samples per class. No rule is given for converting a potentially class-imbalanced stream into a balanced support set. If the sampler enforces class balance, the effective selection rule is no longer the cited VeSSal rule; if it does not, Eq. (3) and Eq. (4) may not operate on a balanced k-shot set. The paper needs to specify the exact selection procedure, including how balance is guaranteed.
  3. [§4.2, Tables 2 and 3] The experimental validation is largely inherited from prior work, which is a circularity concern. Tables 2 and 3 are marked "Adapted from [36]", and the footnote states that "FedAcross(+)" is used "to indicate interchangeability of FedAcross [36] and FedAcross+ in terms of experimental results." Since FedAcross+ differs from FedAcross only by the stream-sampling extension, and that extension is untested, the reported accuracies do not validate FedAcross+ as defined in this manuscript. The paper also delegates the federated stream-sampling behavior to reference [37], which is cited as "forthcoming" at the time of submission. New experiments, or a clear reframing that the static results are reproduced from prior work, are required.
  4. [§4.1, Table 4] The k=0 condition in Table 4 is not defined in the methodology. Section 3.1 sets k = {0,3,5,10}, but Eq. (4) requires a support set with |Dspt_Ti,n| > 0 to form prototypes, and Section 4.1 describes fine-tuning on a k-shot support set. If k=0 means no adaptation and a different inference rule, that should be stated explicitly; if it means using the server model without client fine-tuning, the experimental protocol for that row needs to be documented.
minor comments (6)
  1. [§3.2] There are typos in the text, including "disciminatory" (should be "discriminatory") and "emphazising" (should be "emphasizing").
  2. [§3.5] The sentence "We argue that access to raw client data points is restraint in three ways" should read "is restricted in three ways."
  3. [§3.6] "This is a modular compliment to the existing components" should read "a modular complement to the existing components."
  4. [§3.6, Eq. (6)] The notation in Eq. (6) needs clarification: the definition of the covariance matrix Σ_t and its online update are not specified, and the text says it is the "inverse of the system covariance matrix over samples already selected for labelling" without stating whether this is a class-conditional or global covariance and how it is computed incrementally.
  5. [§4.2, Table 2] The text refers to method "SDAA" while Table 2 lists "SDDA"; the discrepancy should be fixed.
  6. [§4.1] The statement that the original x-ray and multi-spectral data "could not be made available for copyright reasons" is vague; the authors should clarify which dataset or license prevents release.

Circularity Check

2 steps flagged · score 7.0 of 10

FedAcross+ is a relabeling of the authors' prior FedAcross results, and its only new streaming branch is validated solely by the authors' own forthcoming paper [37], never run in this paper.

  1. renaming known result [Section 4.2, footnote to Table 2; same note to Table 3; Table 4 caption 'Source: [36]']
    "Results referenced from [57]; we useFedAcross(+)in the table to indicate interchangeability of FedAcross [36] and FedAcross+ in terms of experimental results."

    The headline empirical claim — 'FedAcross+ produces the best overall adaptation results of all methods under inspection with k = 10 (89.0%)' — is backed by tables whose FedAcross+ rows are explicitly declared interchangeable with the authors' prior FedAcross [36], and Table 4 is captioned 'Source: [36]'. Combined with Section 4.1's static provisioning, no experiment in this paper runs the new SAMPLING_ENABLED path. The reported FedAcross+ accuracies are therefore the prior paper's numbers relabeled; the new component contributes nothing to them. This is a rename of a known result rather than a derivation of a new one.

  2. self citation load bearing [Section 4.1, first paragraph; see also Section 3.6 and Algorithm 1 (SAMPLING_ENABLED)]
    "For the sake of simplicity, it is assumed that the data provisioning for both the FL server and the FL client is static. The efficacy of stream sampling strategies has already been thoroughly evaluated in [39] and subsequently applied and tested in the context of FL in [37]."

    The only new contribution of FedAcross+ over FedAcross [36] is the VeSSal-based stream sampling extension of Section 3.6/Eq. (6), which the abstract claims 'render[s] the solution suitable for non-stationary environments'. The experiments never enable SAMPLING_ENABLED (Algorithm 1, Part 2, line 4) and instead assume static provisioning. The sole validation offered for this load-bearing component is a citation to [37], a 'forthcoming' ESANN paper by the same authors that 'modified the sampling strategy from [39]'. The new claim's support therefore reduces to a self-citation chain ([36] plus the authors' own unpublished [37]) rather than evidence contained in this paper.

full rationale

The paper separates into two claims. For the static pipeline, FedAcross+ is presented as interchangeable with FedAcross [36], and the accuracy tables are explicitly marked as such, with competitor numbers drawn from [57]. This is a re-presentation of a known result; it has independent external benchmark anchoring (Office-31, OfficeHome, DomainNet), so the static baseline carries some real content, but it is not new evidence for the newly named FedAcross+. For the streaming extension, the paper supplies no in-paper experiment: Section 4.1 fixes static data provisioning, Algorithm 1's SAMPLING_ENABLED branch is never executed, and the only support offered is a citation to the authors' own forthcoming paper [37]. Section 3.6 additionally asserts that the support set 'is constructed to ensure that each category is represented equally, with k samples per category' without deriving this balance from Eq. (6), whose determinantal sampling score has no explicit class-balancing mechanism. Overall, the central advertised claim — a streaming-capable FedAcross+ — reduces to self-citation (own [36] and [37]) plus an untested Eq. (6). Because the static baseline is benchmarked against external methods, the circularity is partial but substantial, warranting a score of 7 rather than 8 or 10.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

No invented entities in the ontological sense: the domain-adaptive linear layer is a parameterized module (W, b, gamma, beta in Eq. 1) rather than a new force, particle, or conserved quantity. The free parameters and axioms above are the real cost of admission. The heaviest burdens are the untested transfer of the VeSSal sampling rule into the frozen-backbone federated pipeline, and the assumption that operators can and do provide exactly k labeled samples per class in production.

free parameters (4)
  • k, labeled samples per class in the support set = 3, 5, 10 (plus k = 0 baseline)
    Hand-chosen as the few-shot regime; Section 3.1 assumes operators annotate exactly k samples per class, and all headline results scale with this setting.
  • Client fine-tuning hyperparameters = learning rate 0.1, 200 local epochs, batch size 32
    Hand-chosen in Section 4.1 with no sensitivity analysis; the reported accuracy depends on them.
  • Server pretraining hyperparameters = learning rate 0.01, 300 epochs, batch size 128, early stopping
    Hand-chosen in Section 4.1 following [9]; not swept or justified beyond citation.
  • Sampling probability components q_t and covariance estimator in Eq. 6 = not specified in this paper
    Borrowed from VeSSal [39] and the authors' companion work [37]; the values and their tuning are never stated, yet they drive the new streaming claim.
assumptions (6)
  • domain assumption Source and target domains share the same discrete label space Y with |Y| = L
    Stated in Section 3.1. This excludes open-set or partial-label deployments and is what makes prototype matching well defined.
  • domain assumption Operators annotate exactly k samples per class for each client
    Stated in the bullet assumptions of Section 3.1. The few-shot premise is true by construction rather than by measurement, and annotation cost is not quantified.
  • ad hoc to paper Freezing backbone and classifier preserves enough transferable features that only the domain-adaptive linear layer needs fine-tuning
    Section 3.3 design choice based on task-specific adapters [23]. It is the mechanism that produces the claimed resource savings, but it is not derived and is only evaluated indirectly through inherited experiments.
  • standard math Euclidean-distance nearest-centroid matching is a suitable classifier for the adapted embedding space
    Section 3.4 justifies via the Bregman divergence property of Euclidean distance, citing ProtoNet [41]. Standard background from the few-shot literature.
  • ad hoc to paper The VeSSal sampling rule transfers unchanged to the federated frozen-backbone setting
    Section 3.6 adopts Eq. 6 and cites [37] for the FL adaptation. This paper provides no experiment that exercises the SAMPLING_ENABLED branch of Algorithm 1.
  • domain assumption Samples arriving on the stream are preprocessed, single-item frames
    Section 3.6 scenario: items are 'identified and framed' before classification, which presumes an upstream detection system that is not part of the evaluated pipeline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Domain Borders Are There to Be Crossed With Federated Few-Shot Adaptation." pith.science (2026). https://pith.science/paper/G2VIW6OI

@misc{pith2026250710160,
  author       = {Pith},
  title        = {Pith review of: Domain Borders Are There to Be Crossed With Federated Few-Shot Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G2VIW6OI}},
  note         = {Machine review of arXiv:2507.10160}
}
read the original abstract

Federated Learning has emerged as a leading paradigm for decentralized, privacy-preserving learning, particularly relevant in the era of interconnected edge devices equipped with sensors. However, the practical implementation of Federated Learning faces three primary challenges: the need for human involvement in costly data labelling processes for target adaptation, covariate shift in client device data collection due to environmental factors affecting sensors, leading to discrepancies between source and target samples, and the impracticality of continuous or regular model updates in resource-constrained environments due to limited data transmission capabilities and technical constraints on channel availability and energy efficiency. To tackle these issues, we expand upon an efficient and scalable Federated Learning framework tailored for real-world client adaptation in industrial settings. This framework leverages a pre-trained source model comprising a deep backbone, an adaptation module, and a classifier running on a powerful server. By freezing the backbone and classifier during client adaptation on resource-constrained devices, we allow the domain adaptive linear layer to handle target domain adaptation, thus minimizing overall computational overhead. Furthermore, this setup, designated as FedAcross+, is extended to encompass the processing of streaming data, thereby rendering the solution suitable for non-stationary environments. Extensive experimental results demonstrate the effectiveness of FedAcross+ in achieving competitive adaptation on low-end client devices with limited target samples, successfully addressing the challenge of domain shift. Moreover, our framework accommodates sporadic model updates within resource-constrained environments, ensuring practical and seamless deployment.

Figures

Figures reproduced from arXiv: 2507.10160 by the authors.

Figure 1
Figure 1. Featured approach overview for federated client adaptation. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Challenges of real-world FL applications: Domain shift across clients and re [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Model Architecture for server pre-training on source data set [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Model Architecture for client fine-tuning on reduced [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Client Inference. Source: [36]. for computation of the pairwise L2 distance between the pre-computed target proto￾types and the query vector. The design choice to employ Euclidean distance as the distance metric in our proposed method is based on its integration with P…
Figure 6
Figure 6. Figure 6: Federated Stream Sampling Extension. The embedding pipeline is colored in [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: t-SNE plot of target class data (color-coded dots) and respective target class [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 39 canonical work pages

  1. [36]

    In: Proceedings of the 13th International Conference on Pattern Recog- nition Applications and Methods - V olume 1: ICPRAM

    Röder, M., Münch, M., Raab, C., Schleif, F.M.: Crossing domain borders with federated few-shot adaptation. In: Proceedings of the 13th International Conference on Pattern Recog- nition Applications and Methods - V olume 1: ICPRAM. pp. 511–521. INSTICC, SciTePress (2024). https://doi.org/10.5220/0012351900003654

    work page doi:10.5220/0012351900003654 2024

  2. [57]

    In: Raedt, L.D

    Zhang, W., Shen, L., Zhang, W., Foo, C.S.: Few-shot adaptation of pre-trained networks for domain shift. In: Raedt, L.D. (ed.) Proceedings of the thirty-first international joint conference on artificial intelligence, IJCAI-22. International Joint Conference on Artificial Intelligence, vol. abs/2205.15234, pp. 1665–1671. International Joint Conferences on...

    arXiv 2022

  3. [39]

    In: Proc

    Saran, A., Yousefi, S., Krishnamurthy, A., Langford, J., Ash, J.T.: Streaming active learning with deep neural networks. In: Proc. 40th ICML. ICML’23, JMLR.org (2023)

    work page 2023

  4. [37]

    In: Proceedings of the 32nd European Symposium on Artificial Neural Networks, Com- putational Intelligence and Machine Learning (ESANN 2024)

    Röder, M., Schleif, F.M.: Sparse uncertainty-informed sampling from federated streaming data. In: Proceedings of the 32nd European Symposium on Artificial Neural Networks, Com- putational Intelligence and Machine Learning (ESANN 2024). Bruges, Belgium (October 9-11 2024), forthcoming

    work page 2024

  5. [1]

    VisDA 2022 Challenge: Domain Adaptation for Industrial Waste Sorting

    Bashkirova, D., Mishra, S., Lteif, D., Teterwak, P., Kim, D., Alladkani, F.M., Akl, J., Çalli, B., Bargal, S.A., Saenko, K., Kim, D., Seo, M., Jeon, Y ., Choi, D.G., Ettedgui, S., Giryes, R., Hussein, S.A., Xie, B., Li, S.: VisDA 2022 challenge: Domain adaptation for indus- trial waste sorting. CoRR abs/2303.14828 (Mar 2023). https://doi.org/10.48550/arXi...

    work page Pith review arXiv doi:10.48550/arxiv.2303.14828 2022

  6. [2]

    Beutel, D.J., Topal, T., Mathur, A., Qiu, X., Fernandez-Marques, J., Gao, Y ., Sani, L., Kwing, H.L., Parcollet, T., Gusmão, P.P.d., Lane, N.D.: Flower: A friendly federated learning re- search framework (Jul 2020), https://arxiv.org/abs/2007.14390

    arXiv 2020

  7. [3]

    Neurocomputing 342, 125–136 (May 2019)

    Brinkrolf, J., Göpfert, C., Hammer, B.: Differential privacy for learning vector quantization. Neurocomputing 342, 125–136 (May 2019). https://doi.org/10.1016/j.neucom.2018.11.095, https://linkinghub.elsevier.com/retrieve/pii/S0925231219301511

    work page doi:10.1016/j.neucom.2018.11.095 2019

  8. [4]

    Domain-Specific Batch Normalization for Unsupervised Domain Adaptation

    Chang, W.G., You, T., Seo, S., Kwak, S., Han, B.: Domain-specific batch normalization for unsupervised domain adaptation. CoRR abs/1906.03950, 7346–7354 (Jun 2019). https: //doi.org/10.1109/cvpr.2019.00753, http://arxiv.org/abs/1906.03950

    work page Pith review arXiv 1906

Show all 57 references
  1. [5]

    Dai, Y ., Chen, Z., Li, J., Heinecke, S., Sun, L., Xu, R.: Tackling data heterogeneity in feder- ated learning with class prototypes (2023) FedAcross+ 23

    2023

  2. [6]

    In: International conference on learning representations

    Dhillon, G.S., Chaudhari, P., Ravichandran, A., Soatto, S.: A baseline for few-shot image classification. In: International conference on learning representations. International Con- ference on Learning Representations, vol. abs/1909.02729. OpenReview.net (Sep 2020), https://o...

    1909 arXiv

  3. [7]

    https://doi.org/10.5281/zenodo

    Falcon, W., team, T.P.L.: PyTorch lightning (Mar 2019). https://doi.org/10.5281/zenodo. 3828935, https://github.com/Lightning-AI/lightning

    2019 doi

  4. [8]

    In: Teh, Y .W., Titterington, M

    Glorot, X., Bengio, Y .: Understanding the difficulty of training deep feedforward neu- ral networks. In: Teh, Y .W., Titterington, M. (eds.) Proceedings of the thirteenth interna- tional conference on artificial intelligence and statistics. Proceedings of machine learning res...

    2010

  5. [9]

    In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M

    Guo, Y ., Codella, N.C., Karlinsky, L., Codella, J.V ., Smith, J.R., Saenko, K., Rosing, T., Feris, R.: A broader study of cross-domain few-shot learning. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer vision – ECCV 2020. European Conference on Computer Vi...

    2020 arXiv

  6. [10]

    ArXiv abs/1811.03604 (Nov 2018), https://www.semanticscholar.org/paper/ b72b6fae30561a7e29392e04e82ed1ad7bce8e78

    Hard, A., Rao, K., Mathews, R., Beaufays, F., Augenstein, S., Eichner, H., Kiddon, C., Ramage, D.: Federated learning for mobile keyboard predic- tion. ArXiv abs/1811.03604 (Nov 2018), https://www.semanticscholar.org/paper/ b72b6fae30561a7e29392e04e82ed1ad7bce8e78

    2018 arXiv

  7. [11]

    In: 2016 IEEE conference on computer vision and pattern recognition (CVPR)

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: 2016 IEEE conference on computer vision and pattern recognition (CVPR). pp. 770–778. Com- puter Vision and Pattern Recognition, IEEE (Jun 2016). https://doi.org/10.1109/CVPR.2016. 90

    2016 doi

  8. [12]

    https://doi.org/10.48550/ARXIV .2110.05076, https://arxiv.org/abs/2110.05076

    Hou, M., Sato, I.: A closer look at prototype classifier for few-shot image classification (Oct 2021). https://doi.org/10.48550/ARXIV .2110.05076, https://arxiv.org/abs/2110.05076

    2021 doi arXiv

  9. [14]

    In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Huang, W., Ye, M., Shi, Z., Li, H., Du, B.: Rethinking federated learning with do- main shift: A prototype view. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16312–16322. IEEE Computer Society, Los Alami- tos, CA, USA (jun 2023). https://...

    2023

  10. [15]

    Jothimurugesan, E., Hsieh, K., Wang, J., Joshi, G., Gibbons, P.B.: Federated learning under distributed concept drift (2023)

    2023

  11. [16]

    Kairouz, P., McMahan, H.B., Avent, B., Bellet, A., Bennis, M., Nitin Bhagoji, A., Bonawitz, K., Charles, Z., Cormode, G., Cummings, R., D’Oliveira, R.G.L., Eichner, H., El Rouayheb, S., Evans, D., Gardner, J., Garrett, Z., Gascón, A., Ghazi, B., Gibbons, P.B., Gruteser, M., Ha...

    2021

  12. [17]

    IEEE Transactions on Artificial Intelligence 2(6), 508–518 (Dec 2021)

    Kim, Y ., Cho, D., Han, K., Panda, P., Hong, S.: Domain adaptation without source data. IEEE Transactions on Artificial Intelligence 2(6), 508–518 (Dec 2021). https://doi.org/10. 1109/TAI.2021.3110179, http://arxiv.org/pdf/2007.01524

    2021

  13. [18]

    Kim, Y ., Oh, J., Kim, S., Yun, S.Y .: How to fine-tune models with few samples: Update, data augmentation, and test-time augmentation (2022)

    2022

  14. [19]

    In: Meila, M., 0001, T.Z

    Koh, P.W., Sagawa, S., Marklund, H., Xie, S.M., Zhang, M., Balsubramani, A., Hu, W., Ya- sunaga, M., Phillips, R.L., Gao, I., Lee, T., David, E., Stavness, I., Guo, W., Earnshaw, B.A., Haque, I.S., Beery, S., Leskovec, J., Kundaje, A., Pierson, E., Levine, S., Finn, C., Liang,...

    2021

  15. [20]

    CoRR abs/2102.09003, 615–625 (Jan 2021)

    Kurmi, V .K., Subramanian, V .K., Namboodiri, V .P.: Domain impression: A source data free domain adaptation method. CoRR abs/2102.09003, 615–625 (Jan 2021). https://doi.org/10. 1109/wacv48630.2021.00066, https://arxiv.org/abs/2102.09003

    2021 arXiv

  16. [21]

    https://www.wesort.ai/ (2023), accessed: 2023-10- 24

    Laier, N., Laier, J.: WeSort.AI homepage. https://www.wesort.ai/ (2023), accessed: 2023-10- 24

    2023

  17. [22]

    ACS Sustainable Chemistry & Engineering 9(47), 15722–15738 (Nov 2021)

    Lange, J.P.: Managing plastic waste-sorting, recycling, disposal, and product re- design. ACS Sustainable Chemistry & Engineering 9(47), 15722–15738 (Nov 2021). https://doi.org/10.1021/acssuschemeng.1c05013, https://ris.utwente.nl/ws/files/276811445/ acssuschemeng.1c05013.pdf

    2021

  18. [23]

    In: 2022 IEEE/CVF conference on computer vision and pattern recognition (CVPR)

    Li, W., Liu, X., Bilen, H.: Cross-domain few-shot learning with task-specific adapters. In: 2022 IEEE/CVF conference on computer vision and pattern recognition (CVPR). pp. 7151–7160. Computer Vision and Pattern Recognition, IEEE Computer Society, Los Alamitos, CA, USA (Jun 202...

    2022

  19. [24]

    In: 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Workshop Track Proceedings

    Li, Y ., Wang, N., Shi, J., Liu, J., Hou, X.: Revisiting batch normalization for practical do- main adaptation. In: 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Workshop Track Proceedings. OpenReview.net (2017), https:...

    2017

  20. [25]

    In: Proceedings of the 37th interna- tional conference on machine learning

    Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? Source hy- pothesis transfer for unsupervised domain adaptation. In: Proceedings of the 37th interna- tional conference on machine learning. pp. 6028–6039. ICML’20, JMLR.org (Feb 2020), http://proceeding...

    2020

  21. [26]

    Journal of Machine Learning Research 9(nov), 2579–2605 (2008)

    van der Maaten, L., Hinton, G.: Visualizing high-dimensional data using t-SNE. Journal of Machine Learning Research 9(nov), 2579–2605 (2008)

    2008

  22. [27]

    Marfoq, O., Neglia, G., Kameni, L., Vidal, R.: Federated learning for data streams (2023)

    2023

  23. [28]

    IEEE Trans- actions on Systems, Man, and Cybernetics: Systems (2023)

    Mawuli, C.B., Che, L., Kumar, J., Din, S.U., Qin, Z., Yang, Q., Shao, J.: Fedstream: Prototype-based federated learning on distributed concept-drifting data streams. IEEE Trans- actions on Systems, Man, and Cybernetics: Systems (2023)

    2023

  24. [29]

    In: Singh, A., Zhu, J

    McMahan, B., Moore, E., Ramage, D., Hampson, S., Arcas, B.A.y.: Communication- efficient learning of deep networks from decentralized data. In: Singh, A., Zhu, J. (eds.) Pro- ceedings of the 20th international conference on artificial intelligence and statistics. Proceed- ings...

    2017

  25. [30]

    (eds.) Neural Information Processing Systems, pp

    Müller, R., Kornblith, S., Hinton, G.: When does label smoothing help? In: Wal- lach, H.M., Larochelle, H., Beygelzimer, A., d’Alché Buc, F., Fox, E.A., Garnett, R. (eds.) Neural Information Processing Systems, pp. 4696–4705. Curran Associates Inc., Red Hook, NY , USA (Jun 201...

    2019

  26. [31]

    Nado, Z., Padhy, S., Sculley, D., D’Amour, A., Lakshminarayanan, B., Snoek, J.: Evaluating prediction-time batch normalization for robustness under covariate shift (Jun 2021), https: //arxiv.org/abs/2006.10963

    2021 arXiv

  27. [32]

    In: Proceedings of the IEEE international conference on com- puter vision

    Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi- source domain adaptation. In: Proceedings of the IEEE international conference on com- puter vision. pp. 1406–1415. IEEE International Conference on Computer Vision, IEEE (Oct 2019). https:...

    2019

  28. [33]

    Perez, L., Wang, J.: The effectiveness of data augmentation in image classification using deep learning (Dec 2017), http://arxiv.org/abs/1712.04621

    2017 arXiv

  29. [34]

    Neurocomputing 506, 418–429 (Sep 2022)

    Raab, C., Röder, M., Schleif, F.M.: Domain adversarial tangent subspace alignment for ex- plainable domain adaptation. Neurocomputing 506, 418–429 (Sep 2022). https://doi.org/10. 1016/j.neucom.2022.07.074, https://doi.org/10.1016/j.neucom.2022.07.074

    2022 doi

  30. [35]

    International Journal of Computer Vision (IJCV) 115(3), 211–252 (Dec 2015)

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpa- thy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet large scale visual recognition challenge. International Journal of Computer Vision (IJCV) 115(3), 211–252 (Dec 2015). ht...

    2015 doi

  31. [38]

    In: Daniilidis, K., Maragos, P., Paragios, N

    Saenko, K., Kulis, B., Fritz, M., Darrell, T.: Adapting visual category models to new do- mains. In: Daniilidis, K., Maragos, P., Paragios, N. (eds.) Computer vision – ECCV 2010. European Conference on Computer Vision, vol. 6314, pp. 213–226. Springer Berlin Heidel- berg, Berl...

    2010 doi

  32. [40]

    Acm Computing Surveys 54(9), 1–38 (Oct 2021)

    Siqueira, F., Davis, J.G.: Service computing for industry 4.0: State of the art, challenges, and research opportunities. Acm Computing Surveys 54(9), 1–38 (Oct 2021). https://doi.org/10. 1145/3478680, https://doi.org/10.1145/3478680

    2021 doi

  33. [41]

    In: Guyon, I., Luxburg, U.v., Bengio, S., Wallach, H.M., Fergus, R., Vishwanathan, S.V .N., Garnett, R

    Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In: Guyon, I., Luxburg, U.v., Bengio, S., Wallach, H.M., Fergus, R., Vishwanathan, S.V .N., Garnett, R. (eds.) Advances in neural information processing systems. pp. 4077–4087. NIPS (Mar 2017), htt...

    2017

  34. [42]

    Acm Computing Sur- veys abs/2205.06743 (Feb 2023)

    Song, Y ., Wang, T., Cai, P., Mondal, S.K., Sahoo, J.P.: A comprehensive survey of few- shot learning: Evolution, applications, challenges, and opportunities. Acm Computing Sur- veys abs/2205.06743 (Feb 2023). https://doi.org/10.1145/3582688, https://doi.org/10.1145/ 3582688

    2023 arXiv

  35. [43]

    IEEE Transactions on Neural Networks and Learning Systems 34, 9587–9603 (2021), https://api

    Tan, A.Z., Yu, H., zhen Cui, L., Yang, Q.: Towards personalized federated learning. IEEE Transactions on Neural Networks and Learning Systems 34, 9587–9603 (2021), https://api. semanticscholar.org/CorpusID:232076330

    2021

  36. [44]

    In: AAAI conference on artificial intelligence

    Tan, Y ., Long, G., Liu, L., Zhou, T., Lu, Q., Jiang, J., Zhang, C.: FedProto: Feder- ated prototype learning across heterogeneous clients. In: AAAI conference on artificial intelligence. Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, pp. 8432–8440. As...

    2022 doi

  37. [45]

    In: Proceedings of the 36th International Confer- ence on Neural Information Processing Systems

    Tan, Y ., Long, G., Ma, J., Liu, L., Zhou, T., Jiang, J.: Federated learning from pre-trained models: a contrastive learning approach. In: Proceedings of the 36th International Confer- ence on Neural Information Processing Systems. NIPS ’22, Curran Associates Inc., Red Hook, N...

    2024

  38. [46]

    Triantafillou, E., Larochelle, H., Zemel, R., Dumoulin, V .: Learning a universal tem- plate for few-shot dataset generalization (May 2021), http://proceedings.mlr.press/v139/ triantafillou21a.html

    2021

  39. [47]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5018–5027. Computer Vision and Pattern Recognition, IEEE (Jul 2017)...

    2017 arXiv

  40. [48]

    In: Lee, D., Sugiyama, M., Luxburg, U., Guyon, I., Garnett, R

    Vinyals, O., Blundell, C., Lillicrap, T., kavukcuoglu, k., Wierstra, D.: Matching networks for one shot learning. In: Lee, D., Sugiyama, M., Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in neural information processing systems. NIPS, vol. 29, pp. 3630–3638. Cur- ran Ass...

    2016

  41. [49]

    ACM Computing Surveys 53, 1–34 (Jun 2020)

    Wang, Y ., Yao, Q., Kwok, J., Ni, L.: Generalizing from a few examples: A survey on few-shot learning. ACM Computing Surveys 53, 1–34 (Jun 2020). https://doi.org/10.1145/3386252

    2020 doi

  42. [50]

    Wei, K., Li, J., Ding, M., Ma, C., Yang, H.H., Farokhi, F., Jin, S., Quek, T.Q.S., Vin- cent Poor, H.: Federated learning with differential privacy: Algorithms and performance analysis. Trans. Info. For. Sec.15, 3454–3469 (jan 2020). https://doi.org/10.1109/TIFS.2020. 2988575,...

    2020

  43. [51]

    Scientific Reports 14 (05 2024)

    Wu, S., Chen, J., Nie, X., Wang, Y ., Zhou, X., Lu, L., Peng, W., Nie, Y ., Menhaj, W.: Global prototype distillation for heterogeneous federated learning. Scientific Reports 14 (05 2024). https://doi.org/10.1038/s41598-024-62908-0

    2024 doi

  44. [52]

    IEEE Transactions on Computers 73(04), 1086–1098 (apr 2024)

    Yan, B., Zhang, H., Xu, M., Yu, D., Cheng, X.: Fedrfq: Prototype-based federated learn- ing with reduced redundancy, minimal failure, and enhanced quality. IEEE Transactions on Computers 73(04), 1086–1098 (apr 2024). https://doi.org/10.1109/TC.2024.3353455

    2024

  45. [53]

    IEEE Network 34, 16–22 (Sep 2020)

    Yang, K., Shi, Y ., Zhou, Y ., Yang, Z., Fu, L., Chen, W.: Federated machine learning for intelligent IoT via reconfigurable intelligent surface. IEEE Network 34, 16–22 (Sep 2020). https://doi.org/10.1109/mnet.011.2000045, http://arxiv.org/pdf/2004.05843

    2020 arXiv

  46. [54]

    ACM Transactions on Intelligent Systems and Technology 10(2), 1–19 (Jan 2019)

    Yang, Q., Liu, Y ., Chen, T., Tong, Y .: Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology 10(2), 1–19 (Jan 2019). https: //doi.org/10.1145/3298981, https://doi.org/10.1145/3298981

    2019 doi

  47. [55]

    Yang, T., Andrew, G., Eichner, H., Sun, H., Li, W., Kong, N., Ramage, D., Beaufays, F.: Applied federated learning: Improving google keyboard query suggestions (Dec 2018), https: //www.semanticscholar.org/paper/b97047c4dc75cbe8d6fc5cb3dd5a81d36458892d

    2018

  48. [56]

    In: Proceedings of the AAAI Conference on Artificial Intelligence (2024)

    Zhang, J., Liu, Y ., Hua, Y ., Cao, J.: Fedtgp: Trainable global prototypes with adaptive- margin-enhanced contrastive learning for data and model heterogeneity in federated learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2024)

    2024

  49. [58]

    IEEE Network 36(1), 159–165 (Jan 2022)

    Zhao, C., Sun, X., Yang, S., Ren, X., Zhao, P., McCann, J.: Exploration across small silos: Federated few-shot learning on network edge. IEEE Network 36(1), 159–165 (Jan 2022). https://doi.org/10.1109/MNET.111.2100329 FedAcross+ 27

    2022 doi

Pith tools

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