Pith. sign in

REVIEW 3 major objections 5 minor 41 references

General and Estimable Learning Bound Unifying Covariate and Concept Shifts

T0 review · 3 major / 5 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 This paper proves that target-domain error can be bounded by source error plus two entropic-optimal-transport shift terms, and that both terms can be estimated from finite samples with exponential concentration.

desk verdict Theorem 3.10 is a correct and genuinely general Lipschitz triangle inequality, but the paper's advertised estimability guarantee rests on two concentration theorems that are unproved in the version I read, and Eq. (19) is printed with the inequality reversed. read the letter →

arxiv 2506.12829 v2 pith:SFHCJSZM submitted 2025-06-15 stat.ML cs.LG

Hongbo Chen , Li Charlie Xia This is my paper
classification stat.MLcs.LG MSC 68T0562G0549Q22
keywords distributionshiftcovariateconceptentropicoptimaltransportlearningboundSinkhornalgorithmdebiasedestimationdomainadaptation
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

Existing learning bounds under distribution shift break when the source and target covariates occupy different regions: the concept-shift term becomes undefined and therefore cannot be measured from data. This paper replaces the covariate-shift measure with entropic optimal transport and defines concept shift as the Wasserstein distance between source and target label distributions averaged over the transport plan between covariates. On that basis it proves a bound on target error as source error plus the covariate-shift distance scaled by the hypothesis and loss Lipschitz constants, plus the concept-shift term scaled by the loss Lipschitz constant. It also proposes a split-sample debiased estimator for the covariate-shift distance and a label-based estimator for concept shift, with exponential concentration guarantees, and packages them as the DataShifts algorithm. If the proof holds, distribution-shift analysis becomes a quantifiable pre-deployment step for regression, classification, and stochastic labels alike.

What carries the argument

The load-bearing object is entropic optimal transport $W_\beta(P,Q)=\inf_{\gamma\in\Gamma(P,Q)}\{ \int \rho\,d\gamma + \beta H(\gamma\,|\,P\otimes Q)\}$, whose optimal plan $\gamma^*$ is unique for $\beta>0$ and has full product support. This plan simultaneously defines the covariate shift $S_{\rm Cov}=W_\beta(D_S^X,D_T^X)$ and the total pair concept shift $S^{\gamma^*}_{\rm Cpt}=\mathbb{E}_{(x_S,x_T)\sim\gamma^*}[W_1(D_S^{Y|X=x_S},D_T^{Y|X=x_T})]$, so no pointwise comparison of conditional label laws is needed outside the source support. The estimation side rests on a split-sample debiased estimator $W^{\rm deb}_\beta(\cdot,\cdot)$ that subtracts within-domain empirical transport distances from cross-domain ones, together with the Sinkhorn algorithm used to compute discrete entropic transport plans.

What would settle it

Run the debiased estimator on independent Gaussian samples in, say, $d=100$ with identical source and target distributions; the true $W_\beta$ is zero, so any upward bias is pure estimation error. If the error follows the known $O(N^{-1/d})$ decay of empirical Wasserstein distances instead of the claimed exponential decay, Theorem 4.3 is false. A complementary check: on a regression dataset with a known Lipschitz hypothesis, compute the estimated bound and compare it with the measured target error on many target domains; the paper's claim fails if the bound is routinely violated beyond numerical slack.

Watch

Extended reading notes

Core claim

The paper's central discovery is that entropic optimal transport gives a support-agnostic way to separate covariate shift from concept shift. For an $L_h$-Lipschitz hypothesis and a separately $(L_\ell,L'_\ell)$-Lipschitz loss, the target error satisfies $\epsilon_T(h) \le \epsilon_S(h) + L_h L'_\ell S_{\rm Cov} + L_\ell S^{\gamma^*}_{\rm Cpt}$, where $S_{\rm Cov}=W_\beta(D_S^X,D_T^X)$ and $S^{\gamma^*}_{\rm Cpt}=\mathbb{E}_{(x_S,x_T)\sim\gamma^*}[W_1(D_S^{Y|X=x_S},D_T^{Y|X=x_T})]$ with $\gamma^*$ the optimal entropic transport plan between the covariate distributions. Because $\gamma^*$ couples every source point to every target point when $\beta>0$, the concept-shift term stays well-defined even when $\operatorname{supp}(D_S^X)$ and $\operatorname{supp}(D_T^X)$ are disjoint. The paper further claims that split-sample debiasing of the plug-in entropic transport estimator removes the high-dimensional overestimation and yields exponential concentration, making the bound directly estimable from finite samples.

Load-bearing premise

The practical claim collapses if the split-sample debiased estimator does not actually eliminate the high-dimensional bias of empirical entropic transport distances, because the proof of its exponential concentration (Theorem 4.3) is the only bridge from the bound to finite-sample estimability.

Editorial extensions

If this is right

  • Target error can be bounded and, given samples, estimated for regression, multiclass classification, and stochastic label noise, not just binary deterministic classification.
  • Support overlap between source and target no longer needs to be assumed; the bound is meaningful when the target covariate distribution has mass outside the source support.
  • The split-sample debiased estimator removes the curse-of-dimensionality overestimation of empirical Wasserstein distances, so shift quantification remains meaningful in high-dimensional covariate spaces.
  • The DataShifts algorithm outputs concrete numbers: an estimated covariate shift, concept shift, and an error-bound estimate, usable as a pre-deployment check.
  • The bound's tightness on the enzyme-stability experiment indicates that estimated bounds can track actual test error across many target domains.

Reading between the lines

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

  • Beyond the paper, the bound can be read as a pre-deployment diagnostic: compute the estimated bound on candidate target domains and withhold deployment where it exceeds the tolerable risk envelope.
  • An extension the authors do not pursue is to use the same split-sample debiasing idea on heavy-tailed or non-Euclidean covariate spaces, where the squared-exponential-moment assumption and the exponential concentration of Theorem 4.3 may fail.
  • Because the bound is linear in $L_h$, representation learning could directly optimize $\hat\epsilon_S + L_h L'_\ell \hat S_{\rm Cov} + L_\ell \hat S_{\rm Cpt}$ as a trainable surrogate, turning the theory into a training objective rather than only an evaluation tool.
  • When target labels are unavailable, the concept-shift estimator cannot be evaluated; replacing them with pseudo-labels from the source-trained model is a natural testable extension that this paper does not cover.
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

3 major / 5 minor

Summary. The paper proposes new, support-agnostic definitions of covariate and concept shift based on entropic optimal transport. Its main theoretical result, Theorem 3.10, states that for a separately Lipschitz loss and an L_h-Lipschitz hypothesis, the target error satisfies eps_T(h) <= eps_S(h) + L_h L'_l S_Cov + L_l S^{gamma*}_Cpt, where S_Cov is the entropic OT distance between source and target covariate marginals and S^{gamma*}_Cpt is the expected conditional-label Wasserstein distance under the optimal transport plan. The paper then introduces a split-sample debiased estimator for S_Cov and a label-pair estimator for S^{gamma*}_Cpt, claims exponential concentration inequalities for both (Theorems 4.3 and 4.6), and packages these into the DataShifts algorithm. Section 5 applies DataShifts to an enzyme-stability regression benchmark.

Significance. If all claims held, the paper would make a useful contribution: Theorem 3.10 is a clean, structurally sound triangle inequality that unifies covariate and concept shift for stochastic labeling, general metric label spaces, and general separately Lipschitz losses, and the critique of prior Y|X-shift definitions under support mismatch is correct. The estimability half, however, is not established. Theorem 4.3 is stated without proof despite asserting a surprisingly strong exponential concentration result, and Theorem 4.6 is false as printed because its inequality direction is reversed. Since the paper's central practical claim is that the bound can be estimated with concentration guarantees, the present manuscript does not yet support that claim. A proof or correction of the two concentration theorems is essential before the paper can be accepted.

major comments (3)
  1. [Section 4.1, Theorem 4.3] This theorem is the sole evidence for the claim that S_Cov can be estimated with exponential concentration, but no proof is given. The asserted result is much stronger than the O(N^{-1/d}) bias of the plug-in Wasserstein estimator that the authors themselves cite from Fournier and Guillin in the same section. The statement is of the form 'there exists N such that for NS, NT > N', so it does not provide explicit finite-sample rates, and it cannot by itself justify the high-dimensional accuracy shown in Figure 1 or the use of the debiased estimator in DataShifts. The authors need to supply a proof with explicit constants or replace the claim by standard finite-sample bounds and adjust the claims in the abstract and Section 4 accordingly.
  2. [Section 4.2, Theorem 4.6] As printed, the inequality direction is reversed: the theorem states P(|hat_S_Cpt - S^{gamma*}_Cpt - Delta| < eps) <= RHS, where RHS tends to 0 as N_S, N_T go to infinity, while the left-hand side tends to 1 for a consistent estimator. The intended statement is presumably P(|...| >= eps) <= RHS. In addition, the theorem is stated without proof, the constant Delta is not defined before the theorem, and Algorithm 1 uses hat_S_Cpt directly as an estimate of S^{gamma*}_Cpt even though the theorem is centered on S^{gamma*}_Cpt + Delta. The claimed bound 0 <= Delta <= sqrt(I(D_S_XY)) + sqrt(I(D_T_XY)) also requires proof. Without a corrected and proved statement, the estimability of the concept-shift term is unsupported.
  3. [Section 4.1, Definition 4.2 and Algorithm 1] Theorem 4.3 is stated only for beta = 0, whereas Definition 4.2, Figure 1, and the DataShifts algorithm use positive beta (Figure 1 reports beta = 0.001). The paper itself emphasizes that beta > 0 gives a unique optimal transport plan. Consequently, the concentration guarantee does not apply to the estimator actually used in the experiments and algorithm. The authors should either prove a version of Theorem 4.3 for beta > 0, set beta = 0 in the algorithm and experiments, or clearly state that the practical estimator is not covered by the theorem.
minor comments (5)
  1. [Section 3.2, Theorem 3.10] The main learning bound is stated without proof. The proof is an elementary triangle inequality plus a coupling argument, but it should be included so that readers can verify the role of the optimal transport plan and the measurable selection of optimal label couplings.
  2. [Section 4.1, Eq. (15)] The debiased estimator uses an absolute value inside the square root. If the bracketed expression is negative in expectation, the estimator is not a plug-in estimator of a squared Wasserstein quantity, and its bias behavior should be discussed.
  3. [Section 4.3, Algorithm 1] The text has a typo: 'followingig Algorithm 1' should be 'following Algorithm 1'.
  4. [Section 5.1, Figure 2] The sentence 'The results is shown in Fig.2' should be 'The results are shown in Fig.2'. The figure should also clearly label which points represent the estimated bound and which represent the test error, since the text's description of 'lies just below the diagonal' is ambiguous for an upper bound.
  5. [Abstract and Section 4] The abstract claims the method applies 'in most applications', but Theorems 4.3 and 4.6 require finite squared-exponential moments on X and a bounded label space; these assumptions should be stated in the abstract or at least in the contributions section.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the learning bound is a direct Lipschitz/triangle consequence of the shift definitions, and the estimability shortfalls are correctness risks, not circular reductions.

full rationale

The claimed derivation chain is not circular. S_Cov in Definition 3.1 and S^gamma*_Cpt in Definition 3.6 are defined solely from the source and target distributions; in particular, the transport plan gamma* is chosen from the covariate marginals D_S^X and D_T^X, not from the target error. Theorem 3.10 then follows structurally from the separately Lipschitz loss and the L_h-Lipschitz hypothesis: one bounds |ell(y_T,h(x_T)) - ell(y_S,h(x_S))| by L_ell rho_Y(y_S,y_T) + L'_ell rho'_Y(h(x_S),h(x_T)), controls the first term by the conditional-label Wasserstein distance under gamma*, and controls the second by the transport cost of gamma*, which is at most W_beta(D_S^X,D_T^X). No fitted parameter is renamed as a prediction, and no term in the bound is calibrated to epsilon_T. The estimators in Definitions 4.2 and 4.4 are plug-in or debiasing constructions from source and target samples and are not fit to test error. There are no load-bearing self-citations: Corollary 2.7 rests on standard entropic optimal transport properties and is not a self-citation chain. The genuine concerns in the paper are correctness and rigor issues, not circularity: Theorem 4.3 asserts exponential concentration for the debiased Wasserstein estimator under only finite squared-exponential moments, but no proof is supplied and the claim is in tension with known n^{-1/d} empirical Wasserstein rates; Theorem 4.6 states P(|hat S_Cpt - S^gamma*_Cpt - Delta| < eps) <= RHS, where the RHS tends to 0 although the left side should tend to 1 for a consistent estimator, so the inequality as printed cannot be correct. These issues are about the soundness of the estimability half, not about a definition or fit being circular, so the circularity score remains 0.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

The bound itself introduces no fitted parameter for the target error, but the estimability claims rely on domain-specific assumptions about entropic OT uniqueness, Lipschitz continuity, moment conditions, and conditional label regularity. The main free hyperparameter is beta.

free parameters (1)
  • beta (entropic regularization strength) = 0.001 in experiments
    The entropic optimal transport distance S_Cov and the total pair Y|X shift both depend on beta; the paper chooses a small positive value but gives no selection rule, and the bound and estimators are beta-dependent.
assumptions (6)
  • standard math Entropic optimal transport has a unique minimizer with full product support when beta > 0 (Corollary 2.7).
    Used in Theorem 3.8 to make the total pair Y|X shift well-defined and in the estimator's convergence argument.
  • standard math Conditional probabilities are only unique almost everywhere under the source measure, so expectations under a target measure with new support are ill-defined (Section 2.2).
    This measure-theoretic convention underlies the paper's critique of existing concept shift definitions and motivates the total pair shift.
  • domain assumption The loss is separately (L_l, L'_l)-Lipschitz and the hypothesis is L_h-Lipschitz (Definition 3.9, Theorem 3.10).
    The main bound requires both the loss and the hypothesis to be Lipschitz, which may fail for unbounded losses or non-smooth models without finite constants.
  • domain assumption Source and target covariate distributions have finite squared-exponential moments (Theorem 4.3).
    The concentration inequality for the debiased estimator assumes this strong tail condition; it is stated but not verified in experiments.
  • domain assumption Conditional label distributions are total-variation Lipschitz in x and the label space is bounded (Theorem 4.6).
    The concentration bound for the total pair shift estimator relies on these additional Lipschitz and boundedness assumptions.
  • domain assumption The learner's Lipschitz constant L_h can be computed by existing solvers (Algorithm 1).
    The DataShifts algorithm depends on accurate Lipschitz constant estimates, which are NP-hard to compute exactly for general neural networks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of General and Estimable Learning Bound Unifying Covariate and Concept Shifts." pith.science (2026). https://pith.science/paper/SFHCJSZM

@misc{pith2026250612829,
  author       = {Pith},
  title        = {Pith review of: General and Estimable Learning Bound Unifying Covariate and Concept Shifts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SFHCJSZM}},
  note         = {Machine review of arXiv:2506.12829}
}
read the original abstract

Generalization under distribution shift remains a core challenge in modern machine learning, yet existing learning bound theory is limited to narrow, idealized settings and is non-estimable from samples. In this paper, we bridge the gap between theory and practical applications. We first show that existing bounds become loose and non-estimable because their concept shift definition breaks when the source and target supports mismatch. Leveraging entropic optimal transport, we propose new support-agnostic definitions for covariate and concept shifts, and derive a novel unified error bound that applies to broad loss functions, label spaces, and stochastic labeling. We further develop estimators for these shifts with concentration guarantees, and the DataShifts algorithm, which can quantify distribution shifts and estimate the error bound in most applications -- a rigorous and general tool for analyzing learning error under distribution shift.

Figures

Figures reproduced from arXiv: 2506.12829 by the authors.

Figure 1
Figure 1. (a)–(c) show the empirical distance from entropic optimal transport (β = 0.001) versus sample size N, dimension d, and true Wasserstein-1 distance. In (a)–(b), ηˆ and ηˆ ′ are independent empirical measures from high-dimensional standard normals, thus the true distance is zero. The traditional estimator Wβ(ˆη, ηˆ ′ ) greatly overestimates as d increases, as shown in (b), and even a much larger N only brings a small … view at source ↗
Figure 2
Figure 2. Test Error v.s. Estimated Error Bound on 179 Domains 6. Conclusion We have presented a unified theoretical and algorithmic framework for analyzing generalization under distribution shift. By redefining covariate and concept shift using en￾tropic optimal transport, we derived a new learning bound that is both tighter and estimable from finite samples. Our proposed estimators come with strong concentration guar￾antees… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 28 canonical work pages

  1. [1]

    Invariant risk minimization

    Arjovsky, M., Bottou, L., Gulrajani, I., and Lopez-Paz, D. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019

    arXiv 1907

  2. [2]

    Analysis of representations for domain adaptation

    Ben-David, S., Blitzer, J., Crammer, K., and Pereira, F. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006

    work page 2006

  3. [3]

    Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., and Vaughan, J. W. A theory of learning from different domains. Machine learning, 79: 0 151--175, 2010

    2010

  4. [4]

    Berger, J. O. Statistical decision theory and Bayesian analysis. Springer Science & Business Media, 2013

    work page 2013

  5. [5]

    Boyd, S. P. and Vandenberghe, L. Convex optimization. Cambridge university press, 2004

    2004

  6. [6]

    T., Namkoong, H., and Yadlowsky, S

    Cai, T. T., Namkoong, H., and Yadlowsky, S. Diagnosing model performance under distribution shift. arXiv preprint arXiv:2303.02011, 2023

    arXiv 2023

  7. [7]

    Convergence of entropic schemes for optimal transport and gradient flows

    Carlier, G., Duval, V., Peyr \'e , G., and Schmitzer, B. Convergence of entropic schemes for optimal transport and gradient flows. SIAM Journal on Mathematical Analysis, 49 0 (2): 0 1385--1418, 2017

    work page 2017

  8. [8]

    Convergence rate of general entropic optimal transport costs

    Carlier, G., Pegon, P., and Tamanini, L. Convergence rate of general entropic optimal transport costs. Calculus of Variations and Partial Differential Equations, 62 0 (4): 0 116, 2023

    work page 2023

Show all 41 references
  1. [9]

    and Marchand, M

    Chen, Q. and Marchand, M. Algorithm-dependent bounds for representation learning of multi-source domain adaptation. In International Conference on Artificial Intelligence and Statistics, pp.\ 10368--10394. PMLR, 2023

    2023

  2. [10]

    Cole, S. R. and Frangakis, C. E. The consistency statement in causal inference: a definition or an assumption? Epidemiology, 20 0 (1): 0 3--5, 2009

    2009

  3. [11]

    Joint distribution optimal transportation for domain adaptation

    Courty, N., Flamary, R., Habrard, A., and Rakotomamonjy, A. Joint distribution optimal transportation for domain adaptation. Advances in neural information processing systems, 30, 2017

    2017

  4. [12]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Cuturi, M. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013

    2013

  5. [13]

    Duncan, T. E. On the absolute continuity of measures. The Annals of Mathematical Statistics, 41 0 (1): 0 30--38, 1970

    1970

  6. [14]

    Efficient and accurate estimation of lipschitz constants for deep neural networks

    Fazlyab, M., Robey, A., Hassani, H., Morari, M., and Pappas, G. Efficient and accurate estimation of lipschitz constants for deep neural networks. Advances in neural information processing systems, 32, 2019

    2019

  7. [15]

    and Guillin, A

    Fournier, N. and Guillin, A. On the rate of convergence in wasserstein distance of the empirical measure. Probability theory and related fields, 162 0 (3): 0 707--738, 2015

    2015

  8. [16]

    and McCann, R

    Gangbo, W. and McCann, R. J. The geometry of optimal transportation. 1996

    1996

  9. [17]

    Domain-adversarial training of neural networks

    Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., March, M., and Lempitsky, V. Domain-adversarial training of neural networks. Journal of machine learning research, 17 0 (59): 0 1--35, 2016

    2016

  10. [18]

    and Meng, X.-L

    Gelman, A. and Meng, X.-L. Simulating normalizing constants: From importance sampling to bridge sampling to path sampling. Statistical science, pp.\ 163--185, 1998

    1998

  11. [19]

    Stochastic optimization for large-scale optimal transport

    Genevay, A., Cuturi, M., Peyr \'e , G., and Bach, F. Stochastic optimization for large-scale optimal transport. Advances in neural information processing systems, 29, 2016

    2016

  12. [20]

    Domain adaptation for large-scale sentiment classification: A deep learning approach

    Glorot, X., Bordes, A., and Bengio, Y. Domain adaptation for large-scale sentiment classification: A deep learning approach. In Proceedings of the 28th international conference on machine learning (ICML-11), pp.\ 513--520, 2011

    2011

  13. [21]

    What conditional probability could not be

    H \'a jek, A. What conditional probability could not be. Synthese, 137 0 (3): 0 273--323, 2003

    2003

  14. [22]

    An introduction to statistical learning, volume 112

    James, G., Witten, D., Hastie, T., Tibshirani, R., et al. An introduction to statistical learning, volume 112. Springer, 2013

    2013

  15. [23]

    and Kallenberg, O

    Kallenberg, O. and Kallenberg, O. Foundations of modern probability, volume 2. Springer, 1997

    1997

  16. [24]

    Out-of-distribution generalization via risk extrapolation (rex)

    Krueger, D., Caballero, E., Jacobsen, J.-H., Zhang, A., Binas, J., Zhang, D., Le Priol, R., and Courville, A. Out-of-distribution generalization via risk extrapolation (rex). In International conference on machine learning, pp.\ 5815--5826. PMLR, 2021

    2021

  17. [25]

    Towards out-of-distribution generalization: A survey

    Liu, J., Shen, Z., He, Y., Zhang, X., Xu, R., Yu, H., and Cui, P. Towards out-of-distribution generalization: A survey. arXiv preprint arXiv:2108.13624, 2021

    2021 arXiv

  18. [26]

    Learning transferable features with deep adaptation networks

    Long, M., Cao, Y., Wang, J., and Jordan, M. Learning transferable features with deep adaptation networks. In International conference on machine learning, pp.\ 97--105. PMLR, 2015

    2015

  19. [27]

    G., Raeder, T., Alaiz-Rodr \' guez, R., Chawla, N

    Moreno-Torres, J. G., Raeder, T., Alaiz-Rodr \' guez, R., Chawla, N. V., and Herrera, F. A unifying view on dataset shift in classification. Pattern recognition, 45 0 (1): 0 521--530, 2012

    2012

  20. [28]

    Novozymes Enzyme Stability Prediction , 2023

    Novozymes. Novozymes Enzyme Stability Prediction , 2023. URL https://www.kaggle.com/competitions/novozymes-enzyme-stability-prediction/

    2023

  21. [29]

    Panaretos, V. M. and Zemel, Y. Statistical aspects of wasserstein distances. Annual review of statistics and its application, 6 0 (1): 0 405--431, 2019

    2019

  22. [30]

    Multi-adversarial domain adaptation

    Pei, Z., Cao, Z., Long, M., and Wang, J. Multi-adversarial domain adaptation. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  23. [31]

    Fishr: Invariant gradient variances for out-of-distribution generalization

    Rame, A., Dancette, C., and Cord, M. Fishr: Invariant gradient variances for out-of-distribution generalization. In International Conference on Machine Learning, pp.\ 18347--18377. PMLR, 2022

    2022

  24. [32]

    A stochastic gradient method with an exponential convergence \_rate for finite training sets

    Roux, N., Schmidt, M., and Bach, F. A stochastic gradient method with an exponential convergence \_rate for finite training sets. Advances in neural information processing systems, 25, 2012

    2012

  25. [33]

    Wasserstein distance guided representation learning for domain adaptation

    Shen, J., Qu, Y., Zhang, W., and Yu, Y. Wasserstein distance guided representation learning for domain adaptation. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  26. [34]

    Covariate shift adaptation by importance weighted cross validation

    Sugiyama, M., Krauledat, M., and M \"u ller, K.-R. Covariate shift adaptation by importance weighted cross validation. Journal of Machine Learning Research, 8 0 (5), 2007

    2007

  27. [35]

    and Saenko, K

    Sun, B. and Saenko, K. Deep coral: Correlation alignment for deep domain adaptation. In Computer vision--ECCV 2016 workshops: Amsterdam, the Netherlands, October 8-10 and 15-16, 2016, proceedings, part III 14, pp.\ 443--450. Springer, 2016

    2016

  28. [36]

    and Fran c ois, D

    Verleysen, M. and Fran c ois, D. The curse of dimensionality in data mining and time series prediction. In International work-conference on artificial neural networks, pp.\ 758--770. Springer, 2005

    2005

  29. [37]

    Towards a theoretical framework of out-of-distribution generalization

    Ye, H., Xie, C., Cai, T., Li, R., Li, Z., and Wang, L. Towards a theoretical framework of out-of-distribution generalization. Advances in Neural Information Processing Systems, 34: 0 23519--23531, 2021

    2021

  30. [38]

    Nico++: Towards better benchmarking for domain generalization

    Zhang, X., He, Y., Xu, R., Yu, H., Shen, Z., and Cui, P. Nico++: Towards better benchmarking for domain generalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 16036--16047, 2023

    2023

  31. [39]

    Bridging theory and algorithm for domain adaptation

    Zhang, Y., Liu, T., Long, M., and Jordan, M. Bridging theory and algorithm for domain adaptation. In International conference on machine learning, pp.\ 7404--7413. PMLR, 2019

    2019

  32. [40]

    Unsupervised multi-class domain adaptation: Theory, algorithms, and practice

    Zhang, Y., Deng, B., Tang, H., Zhang, L., and Jia, K. Unsupervised multi-class domain adaptation: Theory, algorithms, and practice. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44 0 (5): 0 2775--2792, 2020

    2020

  33. [41]

    T., Zhang, K., and Gordon, G

    Zhao, H., Des Combes, R. T., Zhang, K., and Gordon, G. On learning invariant representations for domain adaptation. In International conference on machine learning, pp.\ 7523--7532. PMLR, 2019

    2019

Pith tools

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