Pith. sign in

REVIEW 4 major objections 7 minor 51 references

Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read By framing training as a dynamical system and using the Lyapunov exponent to steer the learning rate, LEAwareSGD keeps optimization near the edge of chaos and improves single-domain generalization on PACS, OfficeHome, and DomainNet.

desk verdict LE-guided learning rate for SDG shows consistent gains, but the core estimator is missing and the hyperparameters conflict; worth a serious referee. read the letter →

arxiv 2507.04302 v1 pith:JPWB7A3G submitted 2025-07-06 cs.CV cs.LG

classification cs.CVcs.LG
keywords singledomaingeneralizationLyapunovexponentedgeofchaosadversarialdataaugmentationlearningrateschedulingshiftdynamicalsystems
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 proposes LEAwareSGD, an optimizer that treats neural-network training as a discrete dynamical system and uses the Lyapunov exponent of the parameter updates to decide how fast to learn. When the exponent rises, the learning rate shrinks, keeping the system slightly short of chaos; the paper argues this 'edge of chaos' state balances stability with adaptability. Combined with adversarial data augmentation, the method reports average accuracies of 69.46% on PACS, 54.38% on OfficeHome, and 28.15% on DomainNet, ahead of the baselines compared, and up to 9.47 percentage points of absolute improvement over AdvST when only 10% of PACS training data is used. A sympathetic reader would take the claim to be that steering optimization by a stability measure discovers more generalizable features than fixed-schedule or augmentation-only training.

What carries the argument

The central object is the Lyapunov exponent of the parameter-update dynamics, defined as the long-run growth rate of an infinitesimal perturbation: $LE = \lim_{t\to\infty} \frac{1}{t} \ln\left(\frac{\|\delta\theta_t\|}{\|\delta\theta_0\|}\right)$, with $\delta\theta_{t+1} = (I - \eta_t H[L(\theta_t)]) \delta\theta_t$ obtained by first-order expansion. Because the exponent is tied to the product history of $I - \eta_i H[L(\theta_i)]$, the learning rate can steer it; the rule $\eta_{t+1} = \eta_t e^{-\beta \Delta LE_t}$ for $\Delta LE_t > 0$ is what moves the dynamics toward the edge of chaos. A weight-decay term is added to keep the Hessian approximately positive definite, which biases the LE negative and keeps training stable.

What would settle it

Run LEAwareSGD on PACS with the learning-rate rule driven by the gradient norm instead of the LE; if the control reproduces the reported gains, the Lyapunov measurement is not causally responsible. Alternatively, compute a finite-time Lyapunov exponent from Jacobian-vector products over an actual training trajectory and check whether it is near zero exactly when the rule says the model is at the edge of chaos.

Watch

Extended reading notes

Core claim

The paper's central claim is that the generalization failure of prior adversarial augmentation methods comes from their localized perturbations, which under-explore parameter space; LEAwareSGD fixes this by measuring the Lyapunov exponent of the training dynamics and lowering the learning rate whenever the exponent increases, guiding the model into a slightly stable, edge-of-chaos regime. The reported evidence is consistent accuracy gains across three benchmarks, with the largest gains in low-data PACS, plus auxiliary results showing that the measured LE stays closer to zero during training and that the optimizer also improves ADA, ME-ADA, and AdvST when they are run with it. The paper also claims the method works across ResNet backbones and in leave-one-domain-out DG, at training cost comparable to existing augmentation methods.

Load-bearing premise

The load-bearing premise is that a Lyapunov exponent of the parameter-update dynamics can be computed online at each iteration, but the defining equations require the full Hessian history and an infinite-time limit; without a practical estimator, the update rule reduces to an arbitrary monotone learning-rate decay.

Editorial extensions

If this is right

  • LEAwareSGD beats the compared SDG baselines on all three benchmarks, reporting 69.46% on PACS, 54.38% on OfficeHome, and 28.15% on DomainNet.
  • The largest reported gain is +9.47 percentage points over AdvST when only 10% of the PACS source data is available, suggesting the mechanism helps most in low-data regimes.
  • Adding LEAwareSGD's LE-aware learning-rate rule to existing adversarial augmentation methods improves them: +7.00 for ME-ADA and +5.15 for ADA on OfficeHome.
  • The performance gains hold across ResNet-34/50/101/152 and in leave-one-domain-out DG, so the effect is not tied to one architecture or training protocol.
  • Training stays near the edge of chaos, with the measured LE close to zero during runs, matching the paper's account of why the method generalizes.

Reading between the lines

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

  • The paper never specifies a finite-time estimator for the LE; a practical implementation may effectively be checking how much the Hessian or loss landscape is changing, so the LE framing may be replaceable by curvature-aware learning-rate decay.
  • If the same LE-guided rule works for other augmentation families, it suggests the benefit is in the optimization dynamics rather than in the augmentation itself, which would make it a general tool for out-of-distribution training.
  • A natural test is to compare LEAwareSGD against a simple rule that decays the learning rate whenever the gradient norm fails to decrease; if accuracy matches, the Lyapunov measurement is not the active ingredient.
  • The edge-of-chaos principle suggests the same LE signal could also modulate augmentation strength, not just learning rate, which the paper does not explore.
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 / 7 minor

Summary. The paper proposes LEAwareSGD, an optimizer that modulates the learning rate using a Lyapunov exponent (LE) of the parameter-update dynamics, integrated with adversarial data augmentation for single domain generalization (SDG). The authors frame neural network training as a discrete-time dynamical system, derive a perturbation-propagation formula involving the Hessian, and define an LE that they propose to control via Eq. 9. They report accuracy improvements over prior SDG methods on PACS, OfficeHome, and DomainNet, including a 9.47% gain in a low-data PACS setting. The central claim is that training near the 'edge of chaos' improves generalization, and that LE guidance achieves this state.

Significance. If the results held and the method were implementable, the dynamical-systems perspective could be a novel contribution to SDG and to optimizer design. The paper includes extensive comparisons on three benchmarks, ablations with different optimizers and backbones, and a sensitivity analysis. However, the core algorithm is not executable as written because the LE is defined through an infinite-horizon limit over products of Hessians, with no practical estimator provided. This undermines the reproducibility of all reported results and the causal attribution to LE guidance. The theoretical relation used to justify the learning-rate rule also relies on unstated conditions. Given that the central mechanism is not well-defined, the current manuscript does not support its claims.

major comments (4)
  1. [Section 3.5, Algorithm 1 (lines 6-7); Section 3.2, Eqs. (6)-(7)] The algorithm requires computing LE_t at each iteration, but Eq. (6) expresses δθ_t as a product of (I − η_i H[L(θ_i)]) over every past step, and Eq. (7) defines the LE as an infinite-horizon limit. No finite-time estimator, Hessian-vector product approximation, tangent-space method, or other practical surrogate is provided. For a ResNet-18 model, forming or storing the full Hessian at each iteration is computationally infeasible, and the limit cannot be evaluated online. Consequently, LEAwareSGD is not a well-defined training algorithm, and the accuracies in Tables 1-9 cannot be reproduced or causally attributed to Lyapunov-exponent guidance.
  2. [Section 3.3, Eq. (9)] The learning-rate update decreases η when ΔLE_t > 0 and leaves it unchanged otherwise, so the learning rate is non-increasing over time. This rule cannot amplify perturbations from the stable side of the edge of chaos, and there is no mechanism to increase LE if it becomes too negative. The claimed behavior of 'guiding the model toward the edge of chaos' is therefore not supported by the update rule; in the absence of a practical LE estimator, the procedure reduces to a monotone learning-rate decay schedule.
  3. [Section 3.2, Eq. (8)] The bounds relating the LE to the Hessian and learning rate require unstated assumptions, such as η_i ||H[L(θ_i)]|| < 1 for the lower bound and the validity of neglecting the higher-order term in Eq. (4). Without these conditions, the inequality chain is not justified. Additionally, Eq. (7) defines the LE as a t→∞ limit, but the right-hand sides of Eq. (8) involve finite-horizon products; the passage to the limit is not addressed. This is a load-bearing gap because Eq. (8) is the only theoretical link between the LE and the learning-rate rule.
  4. [Sections 3.5, 4.3, 4.5 and repeated 'Suppl. Sec-??' references] The manuscript repeatedly defers key material to supplementary sections that are not present: the 'detailed theoretical analysis' promised after Section 3.5, the OfficeHome optimizer comparison, the learning-rate analysis for optimizers, the OfficeHome backbone experiments, and the OfficeHome low-data results. These placeholders make it impossible to verify the completeness of the experiments and the theoretical claims, which are central to the paper's contribution.
minor comments (7)
  1. [Abstract and Section 4.5, Table 7] The abstract's 'up to 9.47% improvement' refers to the 10% data ratio in Table 7, not the main protocol; this should be stated clearly to avoid overgeneralization.
  2. [Section 3.4, Eq. (10)] The adversarial data augmentation objective is introduced independently of the LE computation; the connection between the feature distance dθ(·) and the LE-based learning-rate modulation is not explained, so it is unclear how the augmentation and the LE guidance interact.
  3. [Section 4.1, hyperparameter tuning] The sentence 'we fix all training settings and tune only β and γ' does not specify the validation protocol used to select the best values; without a clear selection rule, the reported results may be overfitted to the test domains.
  4. [Section 4.5, Figure 3] The LE curves in Figure 3 are plotted without error bars or confidence intervals, and the method used to compute the plotted LE values is not specified; the figure cannot be quantitatively interpreted.
  5. [Table 5] The multi-source DG results in Table 5 report no standard deviations, unlike Tables 1-3; the authors should either add them or explain why they are omitted.
  6. [Throughout] The placeholder 'Suppl. Sec-??' appears multiple times (e.g., after Section 3.5 and in Sections 4.3 and 4.5); these should be resolved to actual section numbers or the material should be included in the main text.
  7. [Section 3.2, sentence following Eq. (2)] The sentence 'Parameters after perturbation is eθt := θt + δθt,where δ is a tiny real-valued scalar hyperparameter.The perturbation at step t+1propagates according to the following formula:' has missing spaces and punctuation; it should be corrected for readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported accuracy gains are external benchmark results and no equation in the paper reduces to its own input; the LE-based control loop is self-referential only in its internal validation, not circular.

full rationale

The paper's claimed derivation chain is: (i) define the Lyapunov exponent of the linearized parameter-update dynamics (Eqs. 5-7), (ii) bound LE in terms of learning rate and Hessian (Eq. 8), (iii) use the sign of LE_t - LE_{t-1} to decay the learning rate (Eq. 9), and (iv) measure generalization on PACS, OfficeHome, and DomainNet. The headline accuracies (69.46%, 54.38%, 28.15%) are external held-out benchmark numbers and are not algebraically equal to, or fitted from, the LE values or the hyperparameters. The 'edge of chaos' statement is definitional - LE near zero is the paper's own criterion for edge of chaos - but it is used as a mechanism story, not as a proof that forces the accuracy numbers. No self-citation carries the argument: references [44,45] on edge-of-chaos training are by different authors and are used only as background motivation. There is a serious non-circular deficiency: Algorithm 1 line 7 asks to 'Calculate LE using Eq. 7' online, while Eq. 7 defines LE as an infinite-horizon limit and Eq. 6 requires the full Hessian history; this makes the algorithm not well-defined and not reproducible, and the causal attribution to LE guidance is unverifiable. That is a correctness and reproducibility risk, not a circularity step, because the claimed benchmark improvements do not reduce by construction to the LE definition or to a fitted parameter.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central mechanism rests on an internal LE signal whose practical estimator is unspecified; the LR control rule is ad hoc; the regularization coefficients β, γ, and the perturbation scale δ are freely chosen per benchmark; λ in the augmentation objective is never reported. This is a high free-parameter count relative to the claimed first-principles motivation.

free parameters (5)
  • β (LE adjustment sensitivity) = 1e-1 in main experiments; sensitivity peaks at 1e-3 (PACS) and 1e-2 (OfficeHome)
    Controls learning-rate modulation in Eq. 9. Tuned per dataset, but the value used in the tables does not match the reported optimal range.
  • γ (weight decay in Eq. 10) = 5e-4 (PACS), 1e-4 (OfficeHome), 1e-5 (DomainNet)
    Claimed to keep the Hessian near positive definite. OfficeHome value contradicts the sensitivity analysis that favors γ=1e-5.
  • λ (adversarial/consistency balance) = not given
    Balances classification loss and feature distance in Eq. 10; no value is reported anywhere in the paper.
  • δ (perturbation magnitude) = not given
    Scalar hyperparameter for the initial parameter perturbation in the LE derivation; never quantified.
  • LE estimator parameters (window, Hessian approximation) = not given
    Algorithm 1 instructs to compute LE from Eq. 7, a limit; any practical finite-time estimator needs a window size and Hessian approximation, none of which are specified.
assumptions (5)
  • standard math First-order Taylor expansion of the gradient (Eq. 4) with higher-order terms neglected is valid.
    Used to derive the perturbation update Eq. 5; ignores the o(||δθ||^2) term without justification in the presence of non-smooth deep-network losses.
  • standard math Norm inequalities in Eq. 8 hold; specifically that ||I - ηH|| and the lower bound ln(1 - η||H||) are valid, requiring η||H||<1.
    The lower bound is only valid under η_i||H_i||<1 and for the smallest singular value; this condition is not stated or enforced.
  • domain assumption Training near the edge of chaos (LE≈0) improves generalization.
    Borrowed from citing refs [44,45]; the current paper does not test this hypothesis independently.
  • domain assumption The adversarial augmentation objective (Eq. 10) is an effective SDG framework.
    Adopted from AdvST [48] and prior work; the paper does not ablate this choice.
  • ad hoc to paper The learning-rate update rule (Eq. 9) drives the model toward the edge of chaos.
    The rule (decrease η when ΔLE>0) is asserted; it is not derived from the LE bounds (Eq. 8) and no closed-loop stability analysis is provided.
invented entities (1)
  • Finite-time Lyapunov exponent of the parameter-update trajectory as a control signal
    purpose: Measure proximity to the edge of chaos and modulate the learning rate
    The LE is computed from the model's own Hessian history (Eq. 6-7); the estimator is unspecified and cannot be validated externally. No independent observable predicts the claimed edge-of-chaos state.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization." pith.science (2026). https://pith.science/paper/JPWB7A3G

@misc{pith2026250704302,
  author       = {Pith},
  title        = {Pith review of: Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JPWB7A3G}},
  note         = {Machine review of arXiv:2507.04302}
}
read the original abstract

Single Domain Generalization (SDG) aims to develop models capable of generalizing to unseen target domains using only one source domain, a task complicated by substantial domain shifts and limited data diversity. Existing SDG approaches primarily rely on data augmentation techniques, which struggle to effectively adapt training dynamics to accommodate large domain shifts. To address this, we propose LEAwareSGD, a novel Lyapunov Exponent (LE)-guided optimization approach inspired by dynamical systems theory. By leveraging LE measurements to modulate the learning rate, LEAwareSGD encourages model training near the edge of chaos, a critical state that optimally balances stability and adaptability. This dynamic adjustment allows the model to explore a wider parameter space and capture more generalizable features, ultimately enhancing the model's generalization capability. Extensive experiments on PACS, OfficeHome, and DomainNet demonstrate that LEAwareSGD yields substantial generalization gains, achieving up to 9.47\% improvement on PACS in low-data regimes. These results underscore the effectiveness of training near the edge of chaos for enhancing model generalization capability in SDG tasks.

Figures

Figures reproduced from arXiv: 2507.04302 by the authors.

Figure 1
Figure 1. t-SNE visualization of parameter trajectories in initial [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Parameter sensitivity study. LEAwareSGD’s effectiveness in enhancing model general￾ization. On the PACS dataset, integrating LEAwareSGD with AdvST led to a 2.40% increase in average accu￾racy, reaching 69.46%. This improvement highlights LEAwareSGD’s ability to facilitate more effective param￾eter space exploration, enhancing AdvST’s generalization capacity. For ADA and ME-ADA, LEAwareSGD achieved gains of 0.52% and… view at source ↗
Figure 3
Figure 3. LE dynamics over training epochs across four domains [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 37 canonical work pages

  1. [1]

    Chaos: an introduction to dynamical systems

    Kathleen T Alligood, Tim D Sauer, James A Yorke, and David Chillingworth. Chaos: an introduction to dynamical systems. SIAM Review, 40(3):732–732, 1998. 1

  2. [2]

    Adversarial bayesian augmentation for single-source domain generaliza- tion

    Sheng Cheng, Tejas Gokhale, and Yezhou Yang. Adversarial bayesian augmentation for single-source domain generaliza- tion. In Proceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 11400–11410, 2023. 2

  3. [3]

    Randaugment: Practical automated data augmenta- tion with a reduced search space

    Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmenta- tion with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702–703, 2020. 6

  4. [4]

    Attention consistency on visual corruptions for single-source domain generalization

    Ilke Cugu, Massimiliano Mancini, Yanbei Chen, and Zeynep Akata. Attention consistency on visual corruptions for single-source domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4165–4174, 2022. 6

  5. [5]

    Improved regularization of convo- lutional neural networks with cutout

    Terrance DeVries. Improved regularization of convo- lutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 6

  6. [6]

    Optimal ma- chine intelligence at the edge of chaos

    Ling Feng, Lin Zhang, and Choy Heng Lai. Optimal ma- chine intelligence at the edge of chaos. arXiv preprint arXiv:1909.05176, 2019. 1

  7. [7]

    Sharpness-aware minimization for efficiently improving generalization

    Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412,

  8. [8]

    Lya- punov stable learning laws for multilayer recurrent neural networks

    Alejandro Guarneros-Sandoval, Mariana Ballesteros, Ivan Salgado, Julia Rodr´ıguez-Santill´an, and Isaac Chairez. Lya- punov stable learning laws for multilayer recurrent neural networks. Neurocomputing, 491:644–657, 2022. 3

Show all 51 references
  1. [9]

    Train faster, generalize better: Stability of stochastic gradient descent

    Moritz Hardt, Ben Recht, and Yoram Singer. Train faster, generalize better: Stability of stochastic gradient descent. In International Conference on Machine Learning , pages 1225–1234. PMLR, 2016. 2

  2. [10]

    Chaotic nature of the electroencephalo- gram during shallow and deep anesthesia: From analysis of the lyapunov exponent

    Kazuko Hayashi. Chaotic nature of the electroencephalo- gram during shallow and deep anesthesia: From analysis of the lyapunov exponent. Neuroscience, 557:116–123, 2024. 1

  3. [11]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016. 1, 4

  4. [12]

    Augmix: A simple data processing method to improve robustness and uncertainty

    Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781, 2019. 6

  5. [13]

    Self-challenging improves cross-domain generalization

    Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In ECCV, pages 124–140, 2020. 5, 6

  6. [14]

    Adam: A method for stochastic opti- mization

    Diederik P Kingma. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980, 2014. 2, 5, 6

  7. [15]

    Deeper, broader and artier domain generaliza- tion

    Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generaliza- tion. In Proceedings of the IEEE International Conference on Computer Vision, pages 5542–5550, 2017. 2, 4

  8. [16]

    Pro- gressive domain expansion network for single domain gen- eralization

    Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xi- aoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Pro- gressive domain expansion network for single domain gen- eralization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 224–233,

  9. [17]

    Deep learning via dynamical systems: An approximation perspective

    Qianxiao Li, Ting Lin, and Zuowei Shen. Deep learning via dynamical systems: An approximation perspective. Journal of the European Mathematical Society , 25(5):1671–1709,

  10. [18]

    Deep learning for hy- perspectral image classification: An overview

    Shutao Li, Weiwei Song, Leyuan Fang, Yushi Chen, Pedram Ghamisi, and Jon Atli Benediktsson. Deep learning for hy- perspectral image classification: An overview. IEEE Trans- actions on Geoscience and Remote Sensing , 57(9):6690– 6709, 2019. 1

  11. [19]

    Deep learn- ing theory review: An optimal control and dynamical sys- tems perspective

    Guan-Horng Liu and Evangelos A Theodorou. Deep learn- ing theory review: An optimal control and dynamical sys- tems perspective. arXiv preprint arXiv:1908.10920 , 2019. 3

  12. [20]

    Decoupled weight decay regularization

    I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 2, 5, 6

  13. [21]

    Reducing domain gap by reduc- ing style bias

    Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reduc- ing style bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 8690– 8699, 2021. 5

  14. [22]

    A method for solving the convex program- ming problem with convergence rate o (1/k2)

    Yurii Nesterov. A method for solving the convex program- ming problem with convergence rate o (1/k2). In Proceed- ings of the USSR Academy of Sciences , pages 543–547,

  15. [23]

    A survey on transfer learn- ing

    Sinno Jialin Pan and Qiang Yang. A survey on transfer learn- ing. IEEE Transactions on Knowledge and Data Engineer- ing, 22(10):1345–1359, 2009. 1

  16. [24]

    Moment matching for multi-source domain adaptation

    Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 1406–1415,

  17. [25]

    Learning to learn single domain generalization

    Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12556–12565, 2020. 1, 2

  18. [26]

    A stochastic approxi- mation method

    Herbert Robbins and Sutton Monro. A stochastic approxi- mation method. The annals of mathematical statistics, pages 400–407, 1951. 2, 5, 6

  19. [27]

    Transfer learning for visual categorization: A survey

    Ling Shao, Fan Zhu, and Xuelong Li. Transfer learning for visual categorization: A survey. IEEE Transactions on Neu- ral Networks and Learning Systems, 26(5):1019–1034, 2014. 1

  20. [28]

    Gradi- ent matching for domain generalization

    Yuge Shi, Jeffrey Seely, Philip HS Torr, N Siddharth, Awni Hannun, Nicolas Usunier, and Gabriel Synnaeve. Gradi- ent matching for domain generalization. arXiv preprint arXiv:2104.09937, 2021. 5

  21. [29]

    Introduction to focus issue: When machine learn- ing meets complex systems: Networks, chaos, and nonlinear dynamics

    Yang Tang, J ¨urgen Kurths, Wei Lin, Edward Ott, and Ljupco Kocarev. Introduction to focus issue: When machine learn- ing meets complex systems: Networks, chaos, and nonlinear dynamics. Chaos: An Interdisciplinary Journal of Nonlinear Science, 30(6), 2020. 3 9

  22. [30]

    Rmsprop: Divide the gradient by a running average of its recent magnitude

    Tijmen Tieleman and Geoffrey Hinton. Rmsprop: Divide the gradient by a running average of its recent magnitude. coursera: Neural networks for machine learning. COURS- ERA Neural Networks Mach. Learn, 17, 2012. 2, 5, 6

  23. [31]

    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 , pages 5018–5027, 2017. 2, 4

  24. [32]

    On lyapunov exponents for rnns: Under- standing information propagation using dynamical systems tools

    Ryan V ogt, Maximilian Puelma Touzel, Eli Shlizerman, and Guillaume Lajoie. On lyapunov exponents for rnns: Under- standing information propagation using dynamical systems tools. Frontiers in Applied Mathematics and Statistics , 8: 818799, 2022. 3

  25. [33]

    Generalizing to unseen domains via adversarial data augmentation

    Riccardo V olpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. In Pro- ceedings of the 32nd International Conference on Neural In- formation Processing Systems, pages 5339–53...

  26. [34]

    Meta convolutional neural networks for single domain generalization

    Chaoqun Wan, Xu Shen, Yonggang Zhang, Zhiheng Yin, Xinmei Tian, Feng Gao, Jianqiang Huang, and Xian-Sheng Hua. Meta convolutional neural networks for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4682...

  27. [35]

    Sharpness-aware gradient matching for domain generaliza- tion

    Pengfei Wang, Zhaoxiang Zhang, Zhen Lei, and Lei Zhang. Sharpness-aware gradient matching for domain generaliza- tion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 3769–3778,

  28. [36]

    Learning to diversify for single do- main generalization

    Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single do- main generalization. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision, pages 834–843,

  29. [37]

    Simde: A simple domain expan- sion approach for single-source domain generalization

    Qinwei Xu, Ruipeng Zhang, Yi-Yan Wu, Ya Zhang, Ning Liu, and Yanfeng Wang. Simde: A simple domain expan- sion approach for single-source domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 4798–4808, 2023. 1, 2, 5, 6

  30. [38]

    Robust and generalizable visual represen- tation learning via random convolutions

    Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual represen- tation learning via random convolutions. arXiv preprint arXiv:2007.13003, 2020. 5

  31. [39]

    Improve unsupervised domain adaptation with mixup training

    Shen Yan, Huan Song, Nanxiang Li, Lincan Zou, and Liu Ren. Improve unsupervised domain adaptation with mixup training. arXiv preprint arXiv:2001.00677, 2020. 5, 6

  32. [40]

    Causality- inspired domain expansion network for single domain gener- alization

    Shuai Yang, Zhen Zhang, and Lichuan Gu. Causality- inspired domain expansion network for single domain gener- alization. Knowledge-Based Systems, 301:112269, 2024. 1, 2

  33. [41]

    Practical single domain generalization via training-time and test-time learn- ing

    Shuai Yang, Zhen Zhang, and Lichuan Gu. Practical single domain generalization via training-time and test-time learn- ing. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages 3794– 3805, 2024. 4, 5, 6

  34. [42]

    Gradient surgery for multi-task learning

    Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Pro- cessing Systems, 33:5824–5836, 2020. 5

  35. [43]

    Generalizing deep learning for medical image segmentation to unseen do- mains via deep stacked transformation

    Ling Zhang, Xiaosong Wang, Dong Yang, Thomas Sanford, Stephanie Harmon, Baris Turkbey, Bradford J Wood, Holger Roth, Andriy Myronenko, Daguang Xu, et al. Generalizing deep learning for medical image segmentation to unseen do- mains via deep stacked transformation. IEEE Transac...

  36. [44]

    Edge of chaos as a guiding principle for modern neural network training

    Lin Zhang, Ling Feng, Kan Chen, and Choy Heng Lai. Edge of chaos as a guiding principle for modern neural network training. arXiv preprint arXiv:2107.09437, 2021. 1, 3

  37. [45]

    Asymptotic edge of chaos as guiding principle for neural net- work training

    Lin Zhang, Ling Feng, Kan Chen, and Choy Heng Lai. Asymptotic edge of chaos as guiding principle for neural net- work training. International Journal of Artificial Intelligence and Robotics Research, 1(01):2350001, 2024. 3

  38. [46]

    Flatness-aware minimization for domain generalization

    Xingxuan Zhang, Renzhe Xu, Han Yu, Yancheng Dong, Pengfei Tian, and Peng Cui. Flatness-aware minimization for domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5189– 5202, 2023. 2

  39. [47]

    Maximum-entropy adversarial data augmentation for im- proved generalization and robustness

    Long Zhao, Ting Liu, Xi Peng, and Dimitris Metaxas. Maximum-entropy adversarial data augmentation for im- proved generalization and robustness. Advances in Neural Information Processing Systems, 33:14435–14447, 2020. 1, 2, 4, 5, 6

  40. [48]

    Advst: Revisiting data augmentations for single domain generaliza- tion

    Guangtao Zheng, Mengdi Huai, and Aidong Zhang. Advst: Revisiting data augmentations for single domain generaliza- tion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21832–21840, 2024. 2, 4, 5, 6

  41. [49]

    Mixstyle neural networks for domain generalization and adaptation

    Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Mixstyle neural networks for domain generalization and adaptation. International Journal of Computer Vision , 132 (3):822–836, 2024. 5, 6

  42. [50]

    A comprehensive survey on transfer learning

    Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43–76, 2020. 1

  43. [51]

    Surrogate gap minimization improves sharpness-aware training

    Juntang Zhuang, Boqing Gong, Liangzhe Yuan, Yin Cui, Hartwig Adam, Nicha Dvornek, Sekhar Tatikonda, James Duncan, and Ting Liu. Surrogate gap minimization improves sharpness-aware training. arXiv preprint arXiv:2203.08065,

Pith tools

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