Pith. sign in

REVIEW 4 major objections 4 minor 33 references

GAMA: Geometry-Aware Manifold Alignment via Structured Adversarial Perturbations for Robust Domain Adaptation

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read GAMA claims that splitting adversarial perturbations into on-manifold and off-manifold parts and explicitly aligning source and target manifolds with a geodesic loss tightens the domain adaptation bound and improves accuracy, robustness…

desk verdict GAMA is a coherent assembly of known manifold/adversarial ingredients, but the evidence and scholarship are too shaky to back the claimed consistent superiority. read the letter →

arxiv 2505.15194 v1 pith:Q3LOGUFA submitted 2025-05-21 cs.CV

Hana Satou , F Monkey This is my paper
classification cs.CV
keywords domainadaptationmanifoldalignmentadversarialperturbationstangentspacegeodesicdiscrepancyfew-shotdistributionshiftrobustness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that domain adaptation works better when the geometry of data manifolds is respected, not just their distributions. GAMA estimates a local tangent space at each point, pushes adversarial perturbations along that tangent space (semantic, on-manifold) and orthogonal to it (off-manifold), and adds a bidirectional geodesic loss that pulls source and target embeddings into structural alignment. The claim is that each of these pieces tightens a standard domain adaptation bound, and the combined objective beats existing adversarial and self-training methods on DomainNet, VisDA, and Office-Home in both zero-label and few-shot settings.

What carries the argument

The machinery is a local-geometry-guided perturbation generator coupled with a geodesic alignment term. A $k$-NN graph per point feeds a PCA step that defines the tangent basis $T_x(\mathcal{M})$; the input gradient is projected onto this basis to give $\delta_{\mathrm{on}}$ and the residual gives $\delta_{\mathrm{off}}$. Perturbed inputs $x_{\mathrm{on}} = x + \alpha \, \delta_{\mathrm{on}}/\|\delta_{\mathrm{on}}\|$ and $x_{\mathrm{off}} = x + \beta \, \delta_{\mathrm{off}}/\|\delta_{\mathrm{off}}\|$ are trained for consistency and robustness, while a temperature-controlled softmin over $k$-NN geodesic distances between source and target embeddings produces $\mathcal{L}_{\mathrm{geom}}$. The PCA tangent space is the load-bearing geometric estimator, and $\mathcal{L}_{\mathrm{geom}}$ is the mechanism that explicitly aligns the two domains.

What would settle it

On a benchmark where the true manifold geometry is known, such as synthetic data embedded on a low-dimensional submanifold, vary the weight of the geodesic loss and measure the geodesic discrepancy with an independent, much denser $k$-NN graph; if the discrepancy does not fall as the loss weight rises, the central alignment claim is wrong. Equivalently, replacing the PCA tangent projection with random orthogonal projections and observing no drop in accuracy or robustness would show that the geometric structure in the perturbations is not doing the work.

Watch

Extended reading notes

Core claim

The central discovery is that a single training objective can jointly control three sources of target error: on-manifold consistency, off-manifold smoothing, and manifold divergence. GAMA decomposes the input gradient into $\delta_{\mathrm{on}}$ (projection onto the PCA-estimated tangent space) and $\delta_{\mathrm{off}}$ (the residual), and it adds a geodesic discrepancy loss $\mathcal{L}_{\mathrm{geom}}$ that minimizes softmin nearest-neighbor distances between source and target embeddings. The theoretical section derives a final bound $$R_T(f) \le \hat{R}_S(f) + \epsilon_c + C/(\$epsilon^{2}$ n) + \mathrm{GeoD}(\mathcal{M}_S, \mathcal{M}_T) + \$\lambda$^*,$$ in which each term is matched to a corresponding loss component. Empirically, the paper reports that GAMA reaches 71.2% on Office-Home (Clipart to Product), 73.1% in the 1-shot VisDA setting, and 82.7% in the 5-shot setting, with the GeoAlign score dropping from 0.117 for the strongest baseline to 0.089.

Load-bearing premise

The method assumes the local tangent planes and graph-based geodesic distances it estimates are faithful enough to the true data manifolds that minimizing its alignment loss actually reduces the real geometric gap between domains.

Editorial extensions

If this is right

  • If the bound is tight, raising the weight on $\mathcal{L}_{\mathrm{geom}}$ should shrink an independently measured geodesic discrepancy between source and target embeddings.
  • On-manifold perturbations let the model learn invariance to natural intra-class variation, so clean target accuracy under domain shift should improve, not only accuracy under attack.
  • Off-manifold smoothing should flatten low-density regions of feature space, meaning robustness under PGD attacks should improve without sacrificing source accuracy.
  • The same objective applies in few-shot domain adaptation, where only one or five labeled target examples per class are available, and should still reduce manifold mismatch.
  • The unified objective gives a single set of hyperparameters $\lambda_{\mathrm{on}}$, $\lambda_{\mathrm{off}}$, and $\lambda_{\mathrm{geom}}$ that trade semantic consistency, robustness, and alignment; the ablations show that removing any one term hurts either accuracy or robustness.

Reading between the lines

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

  • A reader can infer that the quality of the PCA tangent-space estimate is the bottleneck: on manifolds with high curvature or disconnected components, $k$-NN graphs may misestimate tangents, and the on/off decomposition would then mix semantic and non-semantic directions; the paper does not quantify this sensitivity.
  • The geodesic discrepancy defined in the theory includes a curvature-gap term, but the implemented $\mathcal{L}_{\mathrm{geom}}$ only approximates it through pairwise nearest-neighbor distances; the paper does not directly verify that minimizing $\mathcal{L}_{\mathrm{geom}}$ reduces the curvature gap, and that link is the next natural test.
  • Because the method is agnostic to input modality once embeddings are obtained, the same geometry-aware alignment could be tested on structured prediction or scientific-machine-learning settings such as PDE solvers; the paper mentions these as future directions but reports no experiments there.
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 / 4 minor

Summary. The paper proposes GAMA, a domain adaptation method that combines on-manifold and off-manifold adversarial perturbations with a geodesic alignment loss. The central claims are that the method provides a principled geometric alignment of source and target manifolds, that its objective tightens a domain adaptation generalization bound, and that it consistently outperforms existing methods on DomainNet, VisDA-2017, and Office-Home in both unsupervised and few-shot settings. The manuscript contains a method description, a theoretical analysis section that states several bounds, and an experimental section reporting one Office-Home direction and two few-shot VisDA numbers.

Significance. If the central claims were supported, GAMA would be a useful contribution to geometry-aware domain adaptation, and the idea of combining structured on/off-manifold perturbations with explicit manifold alignment is reasonable and potentially relevant. However, the current manuscript does not provide the evidence needed to establish these claims. The experimental evaluation is far too thin to support the claimed consistent state-of-the-art performance, and the theoretical section does not actually derive the stated bounds. The paper also has substantial presentation problems, including duplicated Related Work sections and garbled references. The strengths of the paper are limited to the conceptual proposal and the clear statement of the loss components; there is no code, no machine-checked derivation, and no reproducible experimental protocol.

major comments (4)
  1. [Section 6, especially 6.4 and 6.5] The empirical evidence does not support the abstract's claim that GAMA 'consistently outperforms' existing methods on DomainNet, VisDA, and Office-Home. Section 6.4 reports exactly one Office-Home direction (Clipart→Product), with no standard deviations, no training details, and no backbone or hyperparameter information. The few-shot results in Section 6.5 are text-only numbers (73.1% and 82.7%) with no protocol, no source-only baseline, and no error bars. DomainNet, which is named in the abstract as an evaluation benchmark, is never reported anywhere. This is a load-bearing gap because the central claim is empirical superiority.
  2. [Section 5.5 and Section 4.3] The final generalization bound in Section 5.5 is asserted rather than derived, and it is constructed so that each term corresponds directly to a loss term in the objective: epsilon_c to L_on, C/(epsilon^2 n) to L_off, and GeoD to L_geom. This makes the 'analysis' largely a restatement of the design objective rather than an independent bound. Furthermore, the geometric alignment loss L_geom in Section 4.3 is an expectation of softmin distances over samples, while GeoD defined in Eq. (5.2) uses a supremum over manifolds and includes a CurvGap term. The paper does not show that minimizing L_geom reduces GeoD, so the link between the optimized loss and the claimed geometric divergence reduction is unestablished.
  3. [Sections 5.3 and 5.4] The bounds relating on-manifold consistency to Rademacher complexity and off-manifold smoothing to adversarial risk are stated without derivation or formal assumptions. Specifically, Section 5.3 claims that enforcing ||f(x_on)-f(x)|| <= epsilon_c implies a local Lipschitz condition that reduces the generalization gap by O(L_on/sqrt(n)), but no hypothesis class, metric, or theorem is given, and the dependence of Rademacher complexity on L_on is not shown. Section 5.4 uses an undefined quantity 'Vol(delta_off)' and asserts a bound R_adv(f) <= R_hat_adv(f) + O(Vol(delta_off)/(epsilon^2 n)) without proof. These are not minor omissions; they are the core of the paper's theoretical justification.
  4. [Sections 4.2 and 4.3] The method relies on PCA-based tangent space estimation from k-NN neighborhoods and graph/RBF-based geodesic distances in the embedding space, but the manuscript provides no analysis or experiment demonstrating that these quantities faithfully represent the true manifold geometry of source and target domains. The GeoAlign Score reported in Table 1 is an invented metric with no definition or reference, and no sensitivity analysis is given for k, the number of PCA components, or the softmin temperature. Since the claimed alignment benefit depends entirely on these geometric estimates, the absence of validation is a substantive concern.
minor comments (4)
  1. [Sections 2 and 3] The paper contains two consecutive sections both titled 'Related Work' with substantially overlapping content; Section 3.5 additionally discusses AI-safety policy considerations (Bengio et al.) that are not connected to the technical method. The manuscript should be reorganized and this duplication removed.
  2. [References] Several references are incomplete or garbled. For example, reference [15] has a malformed title and volume listing, reference [27] lists 'Qinxun Usman, Ninh Kaushik' which appear to be incorrect author names for the VisDA paper, and the MAADA method cited as [7] is actually attributed to Stutz et al., 'Disentangling adversarial robustness and generalization', which is not the method described. The reference list needs to be checked and corrected.
  3. [Section 6.2] The evaluation metrics are not fully specified. 'GeoAlign Score' is described only as average geodesic distance with lower being better, but the formula, the embedding space, and the number of samples used are not given. Robust Accuracy is reported under PGD-10 with epsilon = 4/255, but the model backbone, attack step size, and whether the attack is white-box or transfer-based are not stated.
  4. [Overall] The manuscript lacks training details such as the architecture, optimizer, learning rate schedule, batch size, number of iterations, and the values of the hyperparameters lambda_on, lambda_off, lambda_geom, alpha, beta, and k. These details are essential for reproducibility and for interpreting the reported results.

Circularity Check

2 steps flagged · score 6.0 of 10

The generalization-bound analysis and the GeoAlign evaluation are constructed from GAMA's own loss terms; the theoretical 'tightening' restates the objective rather than deriving it.

  1. self definitional [Section 5.5, 'Final Bound under GAMA' (Eq. 5.5); also Section 5.2]
    "RT (f ) ≤ ˆRS(f ) + ϵc + C/ϵ2n + GeoD(MS, MT ) + λ∗, where each term is controlled by a corresponding GAMA module: - ϵc: on-manifold consistency loss Lon, - C/(ϵ2n): off-manifold smoothing via Loff, - GeoD (·): geometric alignment via Lgeom."

    The proposed bound is assembled by declaring that its three new terms are the three GAMA losses (Lon, Loff, Lgeom). No independent derivation connects the expectation-of-softmin surrogate Lgeom to the supremum-based geodesic discrepancy defined in Section 5.2, and no proof shows that Loff yields the O(C/(ε^2 n)) term. Minimizing Ltotal therefore 'tightens' a bound whose terms are, by construction, the minimized losses themselves. The theoretical 'prediction' reduces to 'each GAMA loss appears in the bound,' making the analysis a restatement of the objective rather than a derived result.

  2. self definitional [Sections 6.2-6.4 and 6.6 (GeoAlign Score metric and ablation)]
    "GeoAlign Score: Average geodesic distance between source and target embeddings (lower is better)."

    This evaluation metric is the same quantity as the alignment loss Lgeom from Section 4.3, which is a bidirectional expectation of min over the other domain of dg(ϕS(x), ϕT(x')); only the softmin surrogate differs. GAMA is trained to minimize exactly this geodesic distance, yet Table 1 and the ablation ('w/o Lgeom ... increases GeoAlign from 0.089 to 0.154') are presented as confirming 'manifold alignment capability.' The 'lowest manifold discrepancy' result is therefore a low value of the trained objective, not an independent measure; the claim reduces to 'the method minimized its own alignment loss.'

full rationale

The Ben-David bound in Section 5.1 is standard and not circular. The circularity lies in Section 5.5: the proposed bound's terms (εc, C/(ε^2 n), GeoD) are declared, in the same passage, to be controlled by the three GAMA losses (Lon, Loff, Lgeom), so minimizing the objective is presented as a theorem without an independent derivation. The GeoAlign Score used to claim superior manifold alignment is the same geodesic discrepancy that Lgeom minimizes, so that evaluation is also self-referential. These issues are partial: the reported Office-Home accuracy (71.2%), PGD robustness (60.1%), and few-shot numbers are concrete and independent of the circular bound, and there is no load-bearing self-citation; the paper cites prior work but does not derive its central claim from its own earlier results. The missing DomainNet table, missing standard deviations, and omitted protocol are evidence gaps rather than circularity. Because the method's principal 'theoretical analysis' and its geometric evaluation reduce by construction to the objective, the partial-circularity score is 6, not 0.

Assumptions & free parameters 9 free parameters · 8 assumptions · 2 invented entities

GAMA rests on the manifold hypothesis and on several auxiliary assumptions linking local PCA geometry to true manifold geometry. The theory adds little independent support because the final bound is essentially a restatement of the objective: each loss term is identified with a term in the bound. The action of the method depends on several unspecified hyperparameters, and the experiments lack the detail needed to verify the claimed improvement.

free parameters (9)
  • on-manifold step size alpha
    Controls magnitude of xon = x + alpha * delta_on/||delta_on||; chosen by hand, no value or search range given.
  • off-manifold step size beta
    Controls magnitude of xoff; unspecified.
  • loss weight lambda_on
    Balancing weight for on-manifold consistency; no value given.
  • loss weight lambda_off
    Balancing weight for off-manifold robustness; no value given.
  • loss weight lambda_geom
    Balancing weight for geodesic alignment; no value given.
  • k for k-NN graph
    Number of neighbors for tangent estimation and geodesic distance; unspecified.
  • number of PCA tangent components
    Dimension of tangent subspace; unspecified.
  • softmin temperature
    Temperature for differentiability of min in Lgeom; unspecified.
  • RBF kernel bandwidth
    Alternative to k-NN for d_g; unspecified.
assumptions (8)
  • domain assumption High-dimensional data lies near a low-dimensional manifold (manifold hypothesis).
    Assumed in Section 4.1 and used to justify PCA tangent estimation and geodesic alignment.
  • domain assumption The PCA-computed tangent space of the k-NN neighborhood approximates the true tangent space of the data manifold.
    Used in Section 4.2 to define on-manifold perturbations; if inaccurate, the semantic/off-manifold split fails.
  • domain assumption The autoencoder reconstruction residual x - D(E(x)) approximates the off-manifold direction.
    Stated in Section 4.1 without verification; this residual is not actually used in the loss definitions later.
  • domain assumption The k-NN graph geodesic distance d_g approximates the true geodesic distance on the manifold.
    Section 4.3 defines Lgeom with d_g on a k-NN graph or RBF kernel; the bound in Section 5.2 uses true geodesic distance, so the two must coincide for the theory to apply.
  • ad hoc to paper Minimizing the softmin surrogate Lgeom reduces the true geodesic discrepancy GeoD(MS, MT).
    Section 5.2 asserts this link without proof; it is the central mechanism connecting the loss to the stated bound.
  • standard math The domain adaptation bound of Ben-David et al. is valid for the considered setting.
    Section 5.1 invokes it as background.
  • ad hoc to paper On-manifold consistency implies a local Lipschitz condition that reduces Rademacher complexity as O(Lon/sqrt(n)).
    Section 5.3 states this relationship without proof.
  • ad hoc to paper Off-manifold regularization flattens decision boundaries and yields the bound Radv <= Rhat_adv + O(Vol(delta_off)/(epsilon^2 n)).
    Section 5.4 cites prior theory [6] but provides no derivation of the specific volume term.
invented entities (2)
  • CurvGap(MS, MT)
    purpose: A term in the geodesic discrepancy definition in Section 5.2 intended to measure local curvature misalignment between source and target manifolds.
    Introduced without a mathematical definition or a way to compute it; it appears only in the final bound and disappears from the following equation, so it is not a substantive contribution.
  • GeoAlign Score
    purpose: Evaluation metric for manifold alignment (average geodesic distance between source and target embeddings).
    Used in experiments but no formula is given; it is computed from the same embedding space as the model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GAMA: Geometry-Aware Manifold Alignment via Structured Adversarial Perturbations for Robust Domain Adaptation." pith.science (2026). https://pith.science/paper/Q3LOGUFA

@misc{pith2026250515194,
  author       = {Pith},
  title        = {Pith review of: GAMA: Geometry-Aware Manifold Alignment via Structured Adversarial Perturbations for Robust Domain Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q3LOGUFA}},
  note         = {Machine review of arXiv:2505.15194}
}
read the original abstract

Domain adaptation remains a challenge when there is significant manifold discrepancy between source and target domains. Although recent methods leverage manifold-aware adversarial perturbations to perform data augmentation, they often neglect precise manifold alignment and systematic exploration of structured perturbations. To address this, we propose GAMA (Geometry-Aware Manifold Alignment), a structured framework that achieves explicit manifold alignment via adversarial perturbation guided by geometric information. GAMA systematically employs tangent space exploration and manifold-constrained adversarial optimization, simultaneously enhancing semantic consistency, robustness to off-manifold deviations, and cross-domain alignment. Theoretical analysis shows that GAMA tightens the generalization bound via structured regularization and explicit alignment. Empirical results on DomainNet, VisDA, and Office-Home demonstrate that GAMA consistently outperforms existing adversarial and adaptation methods in both unsupervised and few-shot settings, exhibiting superior robustness, generalization, and manifold alignment capability.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 24 canonical work pages

  1. [1]

    Deep learning

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436–444, 2015. 2

    work page 2015

  2. [2]

    Imagenet classification with deep convolutional neural net- works

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net- works. Advances in neural information processing systems , 25, 2012. 2

    2012

  3. [3]

    Visualiz- ing data using t-sne

    Laurens Van der Maaten and Geoffrey Hinton. Visualiz- ing data using t-sne. Journal of machine learning research, 9(11), 2008. 2

    work page 2008

  4. [4]

    Explaining and harnessing adversarial examples

    Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 2, 3

    arXiv 2014

  5. [5]

    Towards deep learning models resistant to adversarial attacks

    Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. International Confer- ence on Learning Representations (ICLR), 2018. 2, 3

    work page 2018

  6. [6]

    Theoretically prin- cipled trade-off between robustness and accuracy

    Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric P Xing, Laurent El Ghaoui, and Michael I Jordan. Theoretically prin- cipled trade-off between robustness and accuracy. InInterna- tional conference on machine learning (ICML), pages 7472– 7482, 2019. 2, 3, 5

    work page 2019

  7. [7]

    Disentan- gling adversarial robustness and generalization

    David Stutz, Matthias Hein, and Bernt Schiele. Disentan- gling adversarial robustness and generalization. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6976–6987, 2019. 2, 3, 5

    work page 2019

  8. [8]

    Manifold mixup: Better representations by interpolat- ing hidden states

    Vikas Verma, Alex Lamb, Christopher Beckham, Amir Na- jafi, Ioannis Mitliagkas, David Lopez-Paz, and Yoshua Ben- gio. Manifold mixup: Better representations by interpolat- ing hidden states. In International Conference on Machine Learning (ICML), pages 6438–6447. PMLR, 2019. 2, 3, 5

    work page 2019

Show all 33 references
  1. [9]

    Virtual adversarial training: a regularization method for supervised and semi-supervised learning

    Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(8):1979–1993, 2018. 2, 3, 5

    1979

  2. [10]

    Cross-modality perturbation synergy attack for person re-identification

    Yunpeng Gong, Zhun Zhong, Yansong Qu, Zhiming Luo, Rongrong Ji, and Min Jiang. Cross-modality perturbation synergy attack for person re-identification. In The Thirty- eighth Annual Conference on Neural Information Processing Systems, 2024. 2

    2024

  3. [11]

    Beyond dropout: Robust convolutional neural networks based on local feature masking

    Yunpeng Gong, Chuangliang Zhang, Yongjie Hou, Lifei Chen, and Min Jiang. Beyond dropout: Robust convolutional neural networks based on local feature masking. In 2024 In- ternational Joint Conference on Neural Networks (IJCNN) , pages 1–8. IEEE, 2024. 2, 3

    2024

  4. [12]

    Elimi- nate deviation with deviation for data augmentation and a general multi-modal data learning method

    Yunpeng Gong, Liqing Huang, and Lifei Chen. Elimi- nate deviation with deviation for data augmentation and a general multi-modal data learning method. arXiv preprint arXiv:2101.08533, 2021. 2, 3

    2021 arXiv

  5. [13]

    Adversarial learning for neural pde solvers with sparse data

    Yunpeng Gong, Yongjie Hou, Zhenzhong Wang, Zexin Lin, and Min Jiang. Adversarial learning for neural pde solvers with sparse data. arXiv preprint arXiv:2409.02431, 2024. 2, 3

    2024 arXiv

  6. [14]

    Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging

    Shekoofeh Azizi, Laura Culp, Jan Freyberg, Basil Mustafa, Sebastien Baur, Simon Kornblith, Ting Chen, Nenad Toma- sev, Jovana Mitrovi ´c, Patricia Strachan, et al. Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging. Nature Biome...

    2023

  7. [15]

    Learning internal representations by error prop- agation, parallel distributed processing, explorations in the microstructure of cognition, ed

    David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning internal representations by error prop- agation, parallel distributed processing, explorations in the microstructure of cognition, ed. de rumelhart and j. mcclel- land. vol. 1. 1986. Biometrika, 71(599-607):...

    1986

  8. [16]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2

    2009

  9. [17]

    Dropout: a simple way to prevent neural networks from overfitting

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929–1958, 2014. 3

    1929

  10. [18]

    Adversarial training for free! In Advances in Neu- ral Information Processing Systems (NeurIPS) , volume 32,

    Ali Shafahi, Mahyar Najibi, Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, and Tom Gold- stein. Adversarial training for free! In Advances in Neu- ral Information Processing Systems (NeurIPS) , volume 32,

  11. [19]

    Person re-identification method based on grayscale feature enhance- ment

    Yunpeng GONG, Zhiyong ZENG, and Feng YE. Person re-identification method based on grayscale feature enhance- ment. Journal of Computer Applications, 41(12):3590, 2021. 3

    2021

  12. [20]

    Cross-modality attack boosted by gradient-evolutionary multiform optimization

    Yunpeng Gong, Qingyuan Zeng, Dejun Xu, Zhenzhong Wang, and Min Jiang. Cross-modality attack boosted by gradient-evolutionary multiform optimization. arXiv preprint arXiv:2409.17977, 2024. 3

    2024 arXiv

  13. [21]

    Ex- ploring color invariance through image-level ensemble learn- ing

    Yunpeng Gong, Jiaquan Li, Lifei Chen, and Min Jiang. Ex- ploring color invariance through image-level ensemble learn- ing. arXiv preprint arXiv:2401.10512, 2024. 3

    2024 arXiv

  14. [22]

    Beyond augmentation: Empowering model robustness under extreme capture environments

    Yunpeng Gong, Yongjie Hou, Chuangliang Zhang, and Min Jiang. Beyond augmentation: Empowering model robustness under extreme capture environments. In 2024 International Joint Conference on Neural Networks (IJCNN) , pages 1–8. IEEE, 2024. 3

    2024

  15. [23]

    Cross-task attack: A self-supervision generative framework based on attention shift

    Qingyuan Zeng, Yunpeng Gong, and Min Jiang. Cross-task attack: A self-supervision generative framework based on attention shift. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2024. 3

    2024

  16. [24]

    A theory of learning from different domains

    Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. A theory of learning from different domains. In Ma- chine Learning, volume 79, pages 151–175. Springer, 2010. 3, 4

    2010

  17. [25]

    Managing ex- treme ai risks amid rapid progress

    Yoshua Bengio, Geoffrey Hinton, Andrew Yao, Dawn Song, Pieter Abbeel, Trevor Darrell, Yuval Noah Harari, Ya-Qin Zhang, Lan Xue, Shai Shalev-Shwartz, et al. Managing ex- treme ai risks amid rapid progress. Science, 384(6698):842– 845, 2024. 3

    2024

  18. [26]

    Moment matching for multi-source domain adaptation

    Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, and Kate Saenko. Moment matching for multi-source domain adaptation. Proceedings of the IEEE/CVF International 7 Conference on Computer Vision (ICCV), pages 1406–1415,

  19. [27]

    Visda: The visual domain adaptation challenge

    Xingchao Peng, Qinxun Usman, Ninh Kaushik, Judy Hoff- man, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. In arXiv preprint arXiv:1710.06924, 2017. 5

    2017 arXiv

  20. [28]

    Deep hashing network for unsupervised domain adaptation

    Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5018–5027, 2017. 5

    2017

  21. [29]

    Domain-adversarial train- ing of neural networks

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pas- cal Germain, Hugo Larochelle, Franc ¸ois Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial train- ing of neural networks. In Journal of Machine Learning Re- search, volume 17, pages 2096–2030, 2016. 5

    2016

  22. [30]

    Maximum classifier discrepancy for unsu- pervised domain adaptation

    Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tat- suya Harada. Maximum classifier discrepancy for unsu- pervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3723–3732, 2018. 5

    2018

  23. [31]

    Bridging theory and algorithm for domain adap- tation

    Weichen Zhang, Wanli Ouyang, Wen Li, Dong Xu, and Xi- aokang Liu. Bridging theory and algorithm for domain adap- tation. In International Conference on Machine Learning (ICML), pages 7404–7413. PMLR, 2019. 5

    2019

  24. [32]

    Mutual mean-teaching: Pseudo la- bel refinery for unsupervised domain adaptation on person re-identification

    Yixiao Li, Tianwei Yang, Zesen Wang, Longhui Wang, Hong Liu, and Shengjin Gao. Mutual mean-teaching: Pseudo la- bel refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations (ICLR), 2020. 5

    2020

  25. [33]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Zhou, Tao Liu, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations (ICLR), 2021. 5 8

    2021

Pith tools

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