Pith. sign in

REVIEW 5 major objections 5 minor 52 references

Source-Free Domain Adaptation via Multi-view Contrastive Learning

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

Pith's one-line read A three-phase source-free domain adaptation method screens target samples with adaptive self-entropy thresholds and trains on variance-weighted contrastive views, claiming roughly 2–6 percent higher accuracy than 13 existing approaches on…

desk verdict Plausible SFUDA recipe, but a malformed loss and a 19-point discrepancy between the ablation and main tables make the headline numbers unreliable. read the letter →

arxiv 2507.03321 v1 pith:EKM6RBOL submitted 2025-07-04 cs.CV cs.AI

classification cs.CVcs.AI
keywords source-freedomainadaptationpseudo-labelingcontrastivelearningself-entropyprototypegenerationnoisylabelfilteringimageclassification
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

This paper tries to establish that source-free unsupervised domain adaptation—adapting a model to an unlabeled target domain without any access to the labeled source data—can be made markedly more accurate by attacking two bottlenecks: low-quality class prototypes and noisy pseudo-labels. The proposed method generates prototypes with a Reliable Sample Memory module that keeps, for each class, the lowest self-entropy samples under an adaptively chosen threshold, then assigns pseudo-labels through multi-view contrastive learning on variance-weighted feature views, and finally filters remaining noisy labels. The authors report that on VisDA-2017, Office-Home, and Office-31 the method improves classification accuracy by roughly 2 percent over the second-best competitor and roughly 6 percent over the average of 13 well-known approaches. If correct, the method would outperform all compared methods on these benchmarks, and it would strengthen the case that careful pseudo-label selection alone, without source data or generative modeling, can close much of the domain gap.

What carries the argument

The central machinery is a pair of adaptive thresholds built from self-entropy. Equation (1) defines $\eta$ as the maximum, over $M$ iterations, of the class-wise minimum normalized entropy: $\eta = \max_i \{ \min_j E_{i,j} \}$. This single number controls which target samples are 'reliable' enough to serve as prototypes (RSM) and, in the final phase, which pseudo-labels to keep; the authors argue a fixed threshold would discard informative samples from classes with typically higher entropy. The second piece is variance-weighted multi-view feature concatenation: each augmented view's features are scaled by the normalized average variance of that view's feature matrix, then joined into one vector that k-means clusters. The contrastive loss (Equation 8) and clustering loss (Equation 9) are added to a cross-entropy loss on the pseudo-labeled samples, with coefficients that sum to one and shift weight from pseudo-labels to the distribution-consistency terms as training proceeds.

What would settle it

Train the method on a source-target pair in which the source classifier's logits are deliberately miscalibrated (for example, by adding uniform noise to the logits while keeping the feature extractor fixed), and check whether the RSM-selected low-entropy samples still have above-chance true-label accuracy; if they do not, the entropy-threshold pipeline is not the source of the reported gains.

Watch

Extended reading notes

Core claim

The paper's central claim is that high-quality pseudo-labels are the key to source-free domain adaptation, and that they can be obtained by a three-stage pipeline. First, the Reliable Sample Memory (RSM) uses the per-class minimum of normalized self-entropy values, taken over iterations, as an adaptive threshold $\eta$ to pick a balanced set of representative target samples per class and build better prototypes. Second, Multi-View Contrastive Learning creates several augmented views of each target sample, extracts features per view, weights each view by the average variance of its feature matrix, and concatenates the weighted views into a single representation that is clustered with k-means to assign pseudo-labels. Third, an attention-weighted average of iteration-wise adaptive thresholds filters out labels whose self-entropy is too high. The paper reports that on VisDA-2017 the method reaches 89.23 percent average accuracy (against 87.8 for SHOT++), on Office-31 91.57 percent (against 91.25 for SHOT++), and on Office-Home 73.41 percent (against 72.93 for SHOT++), and that the ablation study shows each of the three phases adds a few points of accuracy.

Load-bearing premise

The load-bearing premise is that a per-class minimum of normalized self-entropy, tracked over iterations, reliably identifies trustworthy target samples, and that the variance-weighted concatenated features form class-consistent k-means clusters; if the frozen source features are miscalibrated on the target domain, the entropy threshold, the prototypes, and the pseudo-labels all degrade together.

Editorial extensions

If this is right

  • If the reported numbers hold, the method outperforms source-dependent approaches on Office-Home, reaching 73.41 percent against 67.84 for DAN and 71.24 for DANN, which would mean source-free adaptation need not sacrifice accuracy to privacy.
  • The ablation study indicates the gains are cumulative: adding pseudo-label assignment raises accuracy by roughly 24–28 points, adding prototype learning adds about 3.7–4.8 points, and noisy-label filtering adds another 3.4–4.0 points, so each phase carries a measurable share of the final result.
  • The adaptive entropy threshold stabilizes early: the paper reports $\eta$ fluctuates between about 0.1 and 0.8 in initial iterations and settles near zero after roughly 20 iterations, meaning the model converges to a consistent pool of high-confidence samples.
  • The method shows strong per-task results across dataset sizes, including 100 percent on the W→D task of Office-31 and 98.6 percent on the plane class of VisDA-2017, suggesting the gains are not driven by a single easy benchmark.

Reading between the lines

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

  • Editorial inference: the entropy-gating mechanism implicitly assumes the frozen source feature extractor is reasonably calibrated on target data; a domain shift that miscalibrates prediction confidence could make low self-entropy select confidently wrong samples, and both the prototype and pseudo-label stages would degrade together.
  • Editorial inference: treating per-view feature variance as a proxy for discriminative signal is a testable design choice; a direct comparison against uniform weighting or learned attention would show whether the reported gains depend specifically on the variance weighting.
  • Editorial inference: because all source-model weights stay frozen, the method's gains come almost entirely from target-side pseudo-label engineering, so it could plausibly be combined with other target-only regularizers without changing the privacy model.
  • Editorial inference: the claimed 6 percent edge over the 'average of 13 methods' depends on the composition of that baseline set; comparing only against the strongest source-free baselines (SHOT++, DPL, SAN) would likely show a smaller though still positive margin.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes a three-phase source-free unsupervised domain adaptation (SFUDA) method. Phase one uses a Reliable Sample Memory (RSM) module and self-entropy-based thresholds to select prototypes; phase two uses multi-view contrastive learning with variance-weighted feature concatenation and k-means clustering to assign pseudo-labels; phase three applies an adaptive entropy threshold to filter noisy labels. The authors report state-of-the-art results on VisDA-2017, Office-31, and Office-Home, claiming roughly 2% improvement over the second-best method and 6% over the average of 13 prior approaches. The method is described as operating with a frozen pre-trained source model and no access to source data.

Significance. If the reported results were valid and reproducible, the method would be a meaningful contribution to SFUDA, particularly the ideas of entropy-gated prototype selection and multi-view contrastive pseudo-label refinement. The paper is readable at a high level, gives a clear statement of the privacy motivation, and explicitly acknowledges increased computational cost as a limitation. However, the technical content as written does not support the central claim. The main contrastive loss in Eq. (8) is not a well-formed objective, the entropy normalization in Eq. (3) is not the self-entropy of a predictive distribution, and the implementation described in Section 4.4 contradicts the source-free and frozen-source premise by using a domain discriminator with a gradient reversal layer. Most seriously, the ablation results in Table 1 and the main results in Table 3 disagree by about 19 percentage points on the D→A and W→A Office-31 tasks, an inconsistency much larger than the claimed margins over baselines. No code or hyperparameter values are provided, so the headline claim rests entirely on the internal consistency of the manuscript, and that consistency fails at load-bearing points.

major comments (5)
  1. [Section 3.4, Eq. (8)] The contrastive loss in Eq. (8) is not a valid or implementable objective. The term w_{i,j} appears outside the logarithm but is never defined for the anchor i, and the denominator sums over all k ≠ i, which includes k = j, so the positive pair z_i, z_j is also treated as a negative. The accompanying sentence explaining the indicator function does not resolve this. Since this loss is one of the three terms in Eq. (10), the multi-view contrastive learning component cannot be evaluated as stated.
  2. [Section 3.3, Eq. (3)] Eq. (3) is labeled as self-entropy after min-max normalization, but it applies min-max normalization to the raw input x and then computes -Σ x' log x'. This is not the entropy of the model's predictive distribution defined in Eq. (2), and it depends on the arbitrary scale of x. Because Algorithm 1 and the threshold η in Eq. (1) are built on this quantity, the prototype-selection criterion is not actually well defined.
  3. [Section 4.4] The implementation details state that optimization is performed end-to-end on Fs, Cs, and D with a gradient reversal layer and a three-layer domain discriminator, but Sections 1 and 3.2 establish that no source data is available and that all source parameters are fixed. A domain discriminator that separates source and target features cannot be trained without source data, so the experimental pipeline described here is not the source-free, frozen-source method proposed in the rest of the paper.
  4. [Tables 1 and 3] For the same four Office-31 transfer tasks, the full method is reported at 98.5 (A→D), 94.3 (A→W), 97.7 (D→A), and 98.0 (W→A) in the ablation study, but at 97.6, 94.7, 78.9, and 79.9 in the main results table. The D→A and W→A discrepancies of roughly 19 percentage points are far larger than the claimed 0.3–2 point improvements over the second-best method, and no explanation is provided. This internal inconsistency means the ablation does not verify the model that produced the headline results.
  5. [Section 3.4, Eqs. (9)–(10)] The clustering loss in Eq. (9) is written as a plain k-means objective with no assignment variable and no mapping from clusters to classes, and the total loss in Eq. (10) includes L_ce, which is never defined, while the text says λ1, λ2, and λ3 control only the contrastive and clustering losses. The constraint that the weights sum to one is stated, but no numerical values are reported, so the overall objective is not reproducible.
minor comments (5)
  1. [Algorithm 2] Algorithm 2 uses inconsistent notation: the header says 'n2 input from eq(2)', and the matrix row size is N/n, while Algorithm 1 and Eq. (1) use an M × N matrix; these definitions should be reconciled.
  2. [Section 3.4] The symbol D is redefined as a labeled dataset D = {x_i, y_i} after Section 3.1 introduced D_T as the unlabeled target domain; this redefinition is confusing and should be avoided.
  3. [Tables 2 and 4] The acronym SFDA is used both for the field 'source-free domain adaptation' and for a specific method in Table 2 and the discussion around Figure 10, which makes the comparison difficult to follow.
  4. [Eq. (13)] The attention-weighted threshold in Eq. (13) is not followed by any algorithm or pseudo-code showing how θ_i, the attention weights α_i, and ρ are computed or where the weighted average is applied in training.
  5. [References] The citation to [49] in Section 4.4 for 'established practices' of data augmentation appears mismatched, as [49] is listed as a domain adaptation method, not an augmentation protocol.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper proposes a standard SFUDA self-training pipeline whose thresholds, pseudo-labels, and losses are computed from target data and frozen-model outputs, then evaluated against target ground-truth labels.

full rationale

The paper does not claim a first-principles derivation; it proposes an empirical three-phase SFUDA pipeline. Pseudo-labels are produced by the pre-trained source model and then used to train the same model, but this is the method's explicit self-training design, not a hidden reduction of a claimed result: final accuracy is measured against target ground-truth labels, and each loss term (contrastive, cross-entropy, clustering) is optimized on target features. The adaptive thresholds in Eqs. (1) and (11) are computed from the model's own entropy outputs; this is the stated operation of the algorithm, not a fitted quantity renamed as a prediction. No load-bearing self-citation appears: references [4,5,6] support background statements about transfer learning and privacy, and no uniqueness theorem is imported from the authors' prior work. The under-specified Eq. (8) and the conflicting Table 1/Table 3 numbers for D->A and W->A are serious correctness and reproducibility concerns, but they are not circularity. The conclusion's stated computational-cost limitation is acknowledged and does not create a circular step. Therefore no circular step can be exhibited under the required quote-and-reduction standard.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim depends on several unstated hyperparameters and data-dependent heuristics. The free parameter count is moderate for an empirical ML paper, but the lack of reported values for λ's and τ, plus the fitted threshold η, mean the method cannot be reproduced.

free parameters (5)
  • loss weights λ1, λ2, λ3 = not reported (sum constrained to 1)
    Trade-off parameters in Eq (10); sensitivity analysis shown in Figs 5-6 but concrete values not given.
  • temperature τ = not reported
    Scales similarity in contrastive loss Eq (8); value never specified.
  • number of views/augmentations M = not reported
    Number of transformations in Eq (4) and view count in Eq (7).
  • threshold η = computed from data (max-min entropy), not a constant
    Used to gate reliable samples in RSM; depends on the entropy normalization and is effectively fit to the target data.
  • learning rate = 0.001 and 0.0005
    Two learning rates are listed without specifying which component uses which.
assumptions (4)
  • domain assumption Self-entropy of the model's predictions reflects sample reliability in the target domain.
    Central to RSM and threshold η in Section 3.3.
  • domain assumption Features from the frozen source model are informative enough that variance-weighted concatenation and k-means produce class-consistent clusters.
    Loaded in Section 3.4, Eq (7) and the clustering step.
  • domain assumption The source model parameters are fixed (frozen) and only target-side modules are trainable.
    Stated in Section 3.2; contradicted later in Section 4.4.
  • standard math Standard contrastive learning assumptions: augmentations preserve semantics, and similarity in feature space corresponds to label similarity.
    Basis of MVCL loss in Eq (8).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Source-Free Domain Adaptation via Multi-view Contrastive Learning." pith.science (2026). https://pith.science/paper/EKM6RBOL

@misc{pith2026250703321,
  author       = {Pith},
  title        = {Pith review of: Source-Free Domain Adaptation via Multi-view Contrastive Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EKM6RBOL}},
  note         = {Machine review of arXiv:2507.03321}
}
read the original abstract

Domain adaptation has become a widely adopted approach in machine learning due to the high costs associated with labeling data. It is typically applied when access to a labeled source domain is available. However, in real-world scenarios, privacy concerns often restrict access to sensitive information, such as fingerprints, bank account details, and facial images. A promising solution to this issue is Source-Free Unsupervised Domain Adaptation (SFUDA), which enables domain adaptation without requiring access to labeled target domain data. Recent research demonstrates that SFUDA can effectively address domain discrepancies; however, two key challenges remain: (1) the low quality of prototype samples, and (2) the incorrect assignment of pseudo-labels. To tackle these challenges, we propose a method consisting of three main phases. In the first phase, we introduce a Reliable Sample Memory (RSM) module to improve the quality of prototypes by selecting more representative samples. In the second phase, we employ a Multi-View Contrastive Learning (MVCL) approach to enhance pseudo-label quality by leveraging multiple data augmentations. In the final phase, we apply a noisy label filtering technique to further refine the pseudo-labels. Our experiments on three benchmark datasets - VisDA 2017, Office-Home, and Office-31 - demonstrate that our method achieves approximately 2 percent and 6 percent improvements in classification accuracy over the second-best method and the average of 13 well-known state-of-the-art approaches, respectively.

Figures

Figures reproduced from arXiv: 2507.03321 by the authors.

Figure 1
Figure 1. Schematic Representation of the Three Phases of [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An Overviewe of the proposed method sample selection, it also depends on pseudo-labeling and self￾training techniques [52]. Vision Transformers have been uti￾lized in numerous research studies to adopt the self-attention mechanism, which is applicable for selecting more reliable features [33, 46, 45, 51]. Based on the gaps in these re￾searches, we propose a three-phase concept of source-free domain adaptation, as sh… view at source ↗
Figure 3
Figure 3. The prototype generation phase based on RMS [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: schematic of multi-view approach. 3.5 Pseudo-labels filtering Due to domain discrepancies, noisy and unreliable labels of￾ten appear in the model’s output, reducing accuracy and presenting a common challenge in pseudo-labeling. To mit￾igate this, we employ an adaptive …
Figure 7
Figure 7. Figure 7: Change of η according to the number of train￾ing iterations (left). Performance change with regard to fix entropy threshold η (right) 4.5 Sensitivity Analysis of Parameters The 3D scatter plot provides a clear visualization of how the weighting factors λ1, λ2, and λ3 a…
Figure 5
Figure 5. Figure 5: Change of Hyperparameters (λ1, λ2 and λ3) [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Correlation Matrix of Hyperparameters (λ1, λ2, λ3) the Office-31 and Office-Home benchmarks. A batch size of 32 is utilized across all tasks to balance computational efficiency and memory usage. We use a batch size of 32 for all tasks, which ensures efficient training …
Figure 8
Figure 8. Figure 8: feature space visualization (left) and the proposed [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 10
Figure 10. Figure 10: Confusion matrix of five state-of-the-art methods [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 50 canonical work pages

  1. [1]

    Weakly-supervised domain adaptation via gan and mesh model for estimating 3d hand poses interact- ing objects

    Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. “Weakly-supervised domain adaptation via gan and mesh model for estimating 3d hand poses interact- ing objects”. In: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition. 2020, pp. 6121–6131

  2. [2]

    Partial adversarial domain adap- tation

    Zhangjie Cao et al. “Partial adversarial domain adap- tation”. In: Proceedings of the European conference on computer vision (ECCV) . 2018, pp. 135–150

  3. [3]

    Homm: Higher-order moment match- ing for unsupervised domain adaptation

    Chao Chen et al. “Homm: Higher-order moment match- ing for unsupervised domain adaptation”. In: Proceed- ings of the AAAI conference on artificial intelligence . Vol. 34. 04. 2020, pp. 3422–3429

  4. [4]

    Leveraging meta-learning to improve unsupervised domain adap- tation

    Amirfarhad Farhadi and Arash Sharifi. “Leveraging meta-learning to improve unsupervised domain adap- tation”. In: The Computer Journal 67.5 (2024), pp. 1838– 1850

  5. [5]

    Domain adaptation in re- inforcement learning: a comprehensive and systematic study

    Amirfarhad Farhadi et al. “Domain adaptation in re- inforcement learning: a comprehensive and systematic study”. In: Frontiers of Information Technology & Elec- tronic Engineering 25.11 (2024), pp. 1446–1465

  6. [6]

    Leveraging Blockchain and ANFIS for Optimal Supply Chain Management

    Amirfarhad Farhadi et al. “Leveraging Blockchain and ANFIS for Optimal Supply Chain Management”. In: arXiv preprint arXiv:2408.17161 (2024)

  7. [7]

    Few-shot generative model for skeleton-based human action synthesis using cross- domain adversarial learning

    Kenichiro Fukushi et al. “Few-shot generative model for skeleton-based human action synthesis using cross- domain adversarial learning”. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision . 2024, pp. 3946–3955

  8. [8]

    Unsupervised domain adaptation by backpropagation

    Yaroslav Ganin and Victor Lempitsky. “Unsupervised domain adaptation by backpropagation”. In: Interna- tional conference on machine learning . PMLR. 2015, pp. 1180–1189

Show all 52 references
  1. [9]

    Domain-adversarial training of neural networks

    Yaroslav Ganin et al. “Domain-adversarial training of neural networks”. In: Journal of machine learning re- search 17.59 (2016), pp. 1–35

  2. [10]

    A kernel two-sample test

    Arthur Gretton et al. “A kernel two-sample test”. In: The Journal of Machine Learning Research13.1 (2012), pp. 723–773

  3. [11]

    Deep residual learning for image recognition

    Kaiming He et al. “Deep residual learning for image recognition”. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2016, pp. 770– 778

  4. [12]

    Model adaptation: Historical contrastive learning for unsupervised domain adap- tation without source data

    Jiaxing Huang et al. “Model adaptation: Historical contrastive learning for unsupervised domain adap- tation without source data”. In: Advances in neural information processing systems 34 (2021), pp. 3635– 3649. Table 4: Accuracy (%) on the Office-Home (ResNet-50) Method Sourc...

  5. [13]

    Structure preserv- ing cycle-GAN for unsupervised medical image domain adaptation

    Paolo Iacono and Naimul Khan. “Structure preserv- ing cycle-GAN for unsupervised medical image domain adaptation”. In: 2024 46th Annual International Con- ference of the IEEE Engineering in Medicine and Bi- ology Society (EMBC) . IEEE. 2024, pp. 1–5

  6. [14]

    Domain-invariant feature extrac- tion and fusion for cross-domain person re-identification

    Zhaoqian Jia et al. “Domain-invariant feature extrac- tion and fusion for cross-domain person re-identification”. In: The visual computer 39.3 (2023), pp. 1205–1216

  7. [15]

    Visually source-free domain adaptation via adversarial style matching

    Mengmeng Jing et al. “Visually source-free domain adaptation via adversarial style matching”. In: IEEE Transactions on Image Processing33 (2024), pp. 1032– 1044

  8. [16]

    Domain adaptation without source data

    Youngeun Kim et al. “Domain adaptation without source data”. In: IEEE Transactions on Artificial Intelligence 2.6 (2021), pp. 508–518

  9. [17]

    Attribute-aligned domain-invariant feature learning for unsupervised domain adaptation person re-identification

    Huafeng Li et al. “Attribute-aligned domain-invariant feature learning for unsupervised domain adaptation person re-identification”. In: IEEE Transactions on Information Forensics and Security 16 (2020), pp. 1480– 1494

  10. [18]

    A comprehensive survey on source- free domain adaptation

    Jingjing Li et al. “A comprehensive survey on source- free domain adaptation”. In: IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

  11. [19]

    Domain invariant and class discrim- inative feature learning for visual domain adaptation

    Shuang Li et al. “Domain invariant and class discrim- inative feature learning for visual domain adaptation”. In: IEEE transactions on image processing 27.9 (2018), pp. 4260–4273

  12. [20]

    Intelligent fault diagnosis by fusing domain adversarial training and maximum mean dis- crepancy via ensemble learning

    Yibin Li et al. “Intelligent fault diagnosis by fusing domain adversarial training and maximum mean dis- crepancy via ensemble learning”. In: IEEE Transac- tions on Industrial Informatics 17.4 (2020), pp. 2833– 2841

  13. [21]

    Do we re- ally need to access the source data? source hypothesis transfer for unsupervised domain adaptation

    Jian Liang, Dapeng Hu, and Jiashi Feng. “Do we re- ally need to access the source data? source hypothesis transfer for unsupervised domain adaptation”. In: In- ternational conference on machine learning . PMLR. 2020, pp. 6028–6039

  14. [22]

    Domain adaptation with auxiliary target domain-oriented clas- sifier

    Jian Liang, Dapeng Hu, and Jiashi Feng. “Domain adaptation with auxiliary target domain-oriented clas- sifier”. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021, pp. 16632– 16642

  15. [23]

    Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer

    Jian Liang et al. “Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer”. In: IEEE Transactions on Pattern Analysis and Machine Intelligence 44.11 (2021), pp. 8602– 8617

  16. [24]

    Multi- domain adversarial feature generalization for person re-identification

    Shan Lin, Chang-Tsun Li, and Alex C Kot. “Multi- domain adversarial feature generalization for person re-identification”. In: IEEE Transactions on Image Pro- cessing 30 (2020), pp. 1596–1607

  17. [25]

    An adaptive source-free unsuper- vised domain adaptation method for mechanical fault detection

    Jianing Liu et al. “An adaptive source-free unsuper- vised domain adaptation method for mechanical fault detection”. In: Mechanical Systems and Signal Pro- cessing 228 (2025), p. 112475

  18. [26]

    Deep transfer learning with joint adaptation networks

    Mingsheng Long et al. “Deep transfer learning with joint adaptation networks”. In: International confer- ence on machine learning . PMLR. 2017, pp. 2208– 2217

  19. [27]

    Learning transferable features with deep adaptation networks

    Mingsheng Long et al. “Learning transferable features with deep adaptation networks”. In:International con- ference on machine learning. PMLR. 2015, pp. 97–105

  20. [28]

    Unsupervised domain adapta- tion with residual transfer networks

    Mingsheng Long et al. “Unsupervised domain adapta- tion with residual transfer networks”. In: Advances in neural information processing systems 29 (2016)

  21. [29]

    Source-free domain adaptation via dynamic pseudo labeling and Self-supervision

    Qiankun Ma et al. “Source-free domain adaptation via dynamic pseudo labeling and Self-supervision”. In: Pattern Recognition 156 (2024), p. 110793

  22. [30]

    Overcoming learning bias via Pro- totypical Feature Compensation for source-free domain adaptation

    Zicheng Pan et al. “Overcoming learning bias via Pro- totypical Feature Compensation for source-free domain adaptation”. In: Pattern Recognition 158 (2025), p. 111025

  23. [31]

    Source-free domain adaptation via avatar prototype generation and adaptation

    Zhen Qiu et al. “Source-free domain adaptation via avatar prototype generation and adaptation”. In:arXiv preprint arXiv:2106.15326 (2021)

  24. [32]

    Semi-supervised domain adap- tation with auto-encoder via simultaneous learning

    Md Mahmudur Rahman, Rameswar Panda, and Mo- hammad Arif Ul Alam. “Semi-supervised domain adap- tation with auto-encoder via simultaneous learning”. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision . 2023, pp. 402– 411

  25. [33]

    Source-free domain adaptation for rgb-d semantic seg- mentation with vision transformers

    Giulia Rizzoli, Donald Shenaj, and Pietro Zanuttigh. “Source-free domain adaptation for rgb-d semantic seg- mentation with vision transformers”. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision . 2024, pp. 615–624

  26. [34]

    SS8: Source data-free domain adap- tation via deep clustering with weighted self-labelling

    Zihao Song et al. “SS8: Source data-free domain adap- tation via deep clustering with weighted self-labelling”. In: 2022 IEEE International Conference on Network- ing, Sensing and Control (ICNSC). IEEE. 2022, pp. 1– 6

  27. [35]

    A Two-Stage Multi-Target Domain Adaptation Framework for Prediction of Key Perfor- mance Indicators Based on Adversarial Network

    Cheng Su et al. “A Two-Stage Multi-Target Domain Adaptation Framework for Prediction of Key Perfor- mance Indicators Based on Adversarial Network”. In: IEEE Transactions on Emerging Topics in Computa- tional Intelligence 8.2 (2024), pp. 1772–1787

  28. [36]

    Enhancing aspect-based sentiment analysis using data augmentation based on back-translation

    Alireza Taheri, Azadeh Zamanifar, and Amirfarhad Farhadi. “Enhancing aspect-based sentiment analysis using data augmentation based on back-translation”. In: International Journal of Data Science and Analyt- ics (2024), pp. 1–26

  29. [37]

    Discriminative adversarial do- main adaptation

    Hui Tang and Kui Jia. “Discriminative adversarial do- main adaptation”. In: Proceedings of the AAAI confer- ence on artificial intelligence. Vol. 34. 04. 2020, pp. 5940– 5947

  30. [38]

    Nearest neighborhood-based deep clustering for source data-absent unsupervised domain adaptation

    Song Tang et al. “Nearest neighborhood-based deep clustering for source data-absent unsupervised domain adaptation”. In: arXiv preprint arXiv:2107.12585 (2021)

  31. [39]

    Source-free domain adaptation with frozen multimodal foundation model

    Song Tang et al. “Source-free domain adaptation with frozen multimodal foundation model”. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024, pp. 23711–23720

  32. [40]

    Robust source-free domain adaptation with anti-adversarial samples training

    Zhirui Wang, Liu Yang, and Yahong Han. “Robust source-free domain adaptation with anti-adversarial samples training”. In: Neurocomputing 614 (2025), p. 128777

  33. [41]

    Instance-invariant domain adaptive object detection via progressive disentanglement

    Aming Wu et al. “Instance-invariant domain adaptive object detection via progressive disentanglement”. In: IEEE Transactions on Pattern Analysis and Machine Intelligence 44.8 (2021), pp. 4178–4193

  34. [42]

    MI 2 GAN: generative adversar- ial network for medical image domain adaptation us- ing mutual information constraint

    Xinpeng Xie et al. “MI 2 GAN: generative adversar- ial network for medical image domain adaptation us- ing mutual information constraint”. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. Springer. 2020, pp. 516–525

  35. [43]

    Unraveling the Mysteries of Label Noise in Source-Free Domain Adaptation: Theory and Practice

    Gezheng Xu et al. “Unraveling the Mysteries of Label Noise in Source-Free Domain Adaptation: Theory and Practice”. In: IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

  36. [44]

    Adversarial domain adaptation with domain mixup

    Minghao Xu et al. “Adversarial domain adaptation with domain mixup”. In: Proceedings of the AAAI conference on artificial intelligence . Vol. 34. 04. 2020, pp. 6502–6509

  37. [45]

    Cdtrans: Cross-domain transformer for unsupervised domain adaptation

    Tongkun Xu et al. “Cdtrans: Cross-domain transformer for unsupervised domain adaptation”. In:arXiv preprint arXiv:2109.06165 (2021)

  38. [46]

    Transformer-based source-free domain adaptation

    Guanglei Yang et al. “Transformer-based source-free domain adaptation”. In: arXiv preprint arXiv:2105.14138 (2021)

  39. [47]

    Unsupervised domain adaptation without source data by casting a bait

    Shiqi Yang et al. “Unsupervised domain adaptation without source data by casting a bait”. In:arXiv preprint arXiv:2010.12427 1.2 (2020), p. 5

  40. [48]

    Source data-free unsupervised do- main adaptation for semantic segmentation

    Mucong Ye et al. “Source data-free unsupervised do- main adaptation for semantic segmentation”. In: Pro- ceedings of the 29th ACM international conference on multimedia. 2021, pp. 2233–2242

  41. [49]

    Unsupervised domain adaptation us- ing robust class-wise matching

    Lei Zhang et al. “Unsupervised domain adaptation us- ing robust class-wise matching”. In: IEEE Transac- tions on Circuits and Systems for Video Technology 29.5 (2018), pp. 1339–1349

  42. [50]

    Discriminative joint probability maximum mean discrepancy (DJP-MMD) for domain adaptation

    Wen Zhang and Dongrui Wu. “Discriminative joint probability maximum mean discrepancy (DJP-MMD) for domain adaptation”. In: 2020 international joint conference on neural networks (IJCNN) . IEEE. 2020, pp. 1–8

  43. [51]

    Universal source-free domain adaptation method for cross-domain fault diagnosis of machines

    Yongchao Zhang et al. “Universal source-free domain adaptation method for cross-domain fault diagnosis of machines”. In: Mechanical Systems and Signal Pro- cessing 191 (2023), p. 110159

  44. [52]

    Source-free domain adaptation with class prototype discovery

    Lihua Zhou et al. “Source-free domain adaptation with class prototype discovery”. In: Pattern recognition 145 (2024), p. 109974

Pith tools

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