Pith. sign in

REVIEW 3 major objections 4 minor 38 references

On the Hardness of Unsupervised Domain Adaptation: Optimal Learners and Information-Theoretic Perspective

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper proves that, under a distribution over source-target-classifier triples, the optimal unsupervised domain adaptation learner is the hardened posterior-aggregate classifier, and the entropy of the posterior target label (PTLU)…

desk verdict A genuinely new average-case framework for UDA hardness, with a sound k>2 lower bound, but the empirical PTLU concentration proof does not hold as written. read the letter →

arxiv 2507.06552 v1 pith:3TA54DOM submitted 2025-07-09 stat.ML cs.ITcs.LGmath.IT

Zhiyi Dong , Zixuan Liu , Yongyi Mao This is my paper · ORCID
classification stat.MLcs.ITcs.LGmath.IT MSC 68T0594A1768Q32
keywords unsuperviseddomainadaptationcovariateshiftposteriordistributionoverclassifiersBayesianoptimallearnerrisklowerboundtargetlabeluncertaintyempiricalentropytransferdifficulty
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to define how hard an unsupervised domain adaptation (UDA) problem is before seeing data. It models the learner's uncertainty as a prior distribution over ground-truth triples (source distribution, target distribution, labeling function), and measures success by average target risk over that distribution. The main result characterizes the optimal learner: it outputs the hardened posterior-aggregate classifier, and that learner's risk is the natural difficulty measure for the problem. The paper then proves that a new quantity, the posterior target label uncertainty (PTLU), lower-bounds the risk of any learner, so PTLU and its sample-based version EPTLU can be used as computable hardness proxies.

What carries the argument

The central object is the posterior distribution $\rho(\cdot|s)$ over classifiers induced by the prior $\pi$ after observing a sample $s$. Its aggregated soft classifier is $\rho^A(y|x,s) = \Pr_{F\sim\rho}(F(x)=y)$, and the hardened version $(\rho^A)^H(x|s) = \arg\max_y \rho^A(y|x,s)$ is the Bayes-optimal output. The argument is carried by two mechanisms: a decomposition showing that overall risk is an expectation of sample-wise risks, and an entropy-error lower bound (Fano's inequality) that converts posterior uncertainty into a risk lower bound. For finite samples, EPTLU replaces the $q$-expectation with an average over $x_t$, with Hoeffding-type and posterior-concentration arguments controlling the gap.

What would settle it

Take a synthetic UDA class with finite input space, fixed $p$ and $q$, and a prior over two classifiers that agree on the source support but differ on the target. Compute the exact optimal risk $R^*$ and PTLU for two such classes; if the class with smaller PTLU has larger $R^*$, then PTLU does not order UDA difficulty as a proxy. The paper's inequality would still hold, since it is a lower bound, so this observation targets the proxy claim rather than Theorem 1.

Watch

Extended reading notes

Core claim

The central claim is that the UDA learning problem under a prior $\pi$ over $(p,q,f)$ has an explicit optimal learner. For any observed sample $s_{m,n}$, the learner that returns $(\rho^A)^H(\cdot|s_{m,n})$ — the hardened version of the aggregated posterior classifier — achieves the sample-wise optimal risk $e^*(s_{m,n},q) = 1 - \mathbb{E}_{x\sim q}\max_y \rho^A(y|x,s_{m,n})$, and averaging these sample-wise risks gives the optimal overall risk $R^*_{m,n}$. Consequently, the difficulty of a UDA class is the optimal learner's expected risk, not the worst case over rare hard instances. The paper also establishes that for any learner, the entropy $U(s_{m,n},q) = \mathbb{E}_{x\sim q} H(\rho^A(\cdot|x,s_{m,n}))$ lower-bounds the sample-wise risk via a Fano-type entropy-error bound, with different forms for binary and multiclass labels. The empirical version EPTLU $\tilde{U}(s_{m,n})$ replaces the target expectation by an average over the unlabeled target sample, and a high-probability lower bound is proved for it as well.

Load-bearing premise

The prior $\pi$ over $(p,q,f)$ is assumed to be the true distribution generating the problem; if it is misspecified, the claimed optimal learner and PTLU-based difficulty numbers no longer track actual transfer difficulty.

Editorial extensions

If this is right

  • For any UDA class $\pi$, the optimal learner is fully specified by the posterior; designing a UDA algorithm can be reframed as approximating the posterior over classifiers rather than minimizing a divergence between source and target features.
  • PTLU and EPTLU give a principled way to compare source domains: a candidate source is preferable when the posterior it induces leads to lower target-label entropy.
  • The lower bounds identify an irreducible risk that no UDA method can beat, so a practitioner can decide that adaptation is not worth attempting when EPTLU is high.
  • Rare hard UDA instances no longer dominate the analysis; difficulty is an average-case quantity governed by the prior $\pi$.

Reading between the lines

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

  • One implication the paper leaves implicit is that disagreement among an ensemble of source-consistent classifiers could serve as a practical estimator of PTLU, since the entropy of $\rho^A$ is exactly the average ensemble disagreement.
  • The formulation suggests a testable protocol: on benchmark UDA tasks, estimate EPTLU from labeled source plus unlabeled target data and check whether it predicts actual transfer accuracy across tasks; the paper does not run such an experiment.
  • Because the optimal learner is posterior-based, the framework connects naturally to Bayesian deep learning, where the posterior over classifiers is approximated; treating PTLU as an acquisition criterion for active UDA would be a natural extension.
  • The prior-sensitivity caveat implies that PTLU should be reported with uncertainty over $\pi$, not as a single number.
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 / 4 minor

Summary. This paper proposes a Bayesian average-case formulation of unsupervised domain adaptation. A prior π over ground-truth triples (p, q, f), called a UDA class, determines the learner's uncertainty; the learner sees m labeled source points and n unlabeled target points and is evaluated by the expected target-domain 0/1 risk averaged over π. The paper proves (Lemma 1) that the overall risk decomposes into sample-wise risks, characterizes the optimal learner (Theorem 1) as the hardened posterior-aggregate classifier, and uses the resulting optimal risk as a measure of UDA hardness. It then introduces Posterior Target Label Uncertainty (PTLU) and an empirical version EPTLU, proves Fano-type lower bounds in terms of these quantities (Theorems 2–5), and presents four synthetic examples comparing PTLU with f-divergence, Wasserstein distance, HΔH divergence, Y-discrepancy, and the transfer exponent.

Significance. The average-case Bayesian viewpoint is a genuine departure from the worst-case analyses that dominate the UDA literature, and it usefully de-emphasizes rare hard instances. Lemma 1 and Theorem 1 are clean and correct, and the four examples are instructive: they show concretely that common divergence measures can disagree with actual transfer difficulty in either direction, while PTLU tracks posterior uncertainty. The k>2 Fano lower bound is a valid information-theoretic argument. The paper's central weakness is that its two load-bearing empirical claims are not adequately proved: the EPTLU concentration bound in Theorem 4 uses an invalid Hoeffding step, and the k=2 branch of the lower bounds is not derived in the supplied proofs. Because the practical message of the paper rests on EPTLU as a computable hardness proxy, these gaps must be repaired before the claims are supported.

major comments (3)
  1. [§3.5, Theorem 4; Appendix G] The proof applies Hoeffding's inequality to the n quantities H(ρA(·|x_i, sm,n)) as if they were i.i.d. draws of a fixed bounded function, but they are not: ρA(·|·, sm,n) is the posterior given the entire target sample, so replacing one target point changes the posterior and hence all n summands. The target expectation U(sm,n) is itself a random function of the sample, so the classical Hoeffding bound does not apply and the claimed probability 1−exp(−2nt²/(log k)²) is unsupported. A McDiarmid analysis with only the boundedness of each entropy term would give a much weaker exp(−2t²/(n log² k)) rate unless a Lipschitz or stability condition on the posterior is imposed, and no such condition appears. Since Theorem 4 is the only result that makes EPTLU a finite-sample lower bound on the risk of any learner, the central practical claim of the paper is not established. A repair would be to evaluate EPTLU on an independent target sample after forming the posterior from a separate sample, or to prove a stability inequality for the posterior entropy.
  2. [§3.3, Theorems 2–5; Appendices D and E] For k=2, the displayed lower bounds have the form e ≥ U²/(4+e*²), but the proofs do not establish this. Appendix D derives only U ≤ H_b(e) ≤ 2√(e(1−e)); no e* appears in the derivation, and the displayed denominator 4+e*² is not obtained. The same unproved k=2 expression is propagated to Corollary 2, Theorem 3, Theorem 4, and Theorem 5, and in Theorem 5 the denominator uses e*(sm,n,q) while Corollary 2 uses e*(p,q,fp), a further mismatch that is not explained. Because all examples in Section 3.4 are binary, the lower-bound support for the binary demonstrations rests on this missing step. Please either give a correct proof of the k=2 statement or restrict the stated bounds to k>2 and adapt the binary examples accordingly.
  3. [§3.4 and §3.5] The four examples in Section 3.4 use continuous input spaces and infinite hypothesis classes, but Theorem 5, which is the paper's finite-sample estimation result for EPTLU, requires a finite input space and a finite consistent-classifier set F[fp] (K<∞) together with assumptions (A1)–(A3). The paper therefore proves EPTLU consistency in a setting disjoint from the settings in which PTLU is demonstrated. This is not by itself an error, but the paper should state explicitly that the examples are exact infinite-sample calculations and are outside the scope of Theorem 5; as written, the abstract and introduction invite the reader to view the examples as supporting the EPTLU proxy claim, which they do not.
minor comments (4)
  1. [§3.4, Table 2] Table 2 appears inconsistent with the text for Example 3: the text reports PTLU values 0.72 and 0.36 and states that HΔH divergence and Y-discrepancy are lower in the second instance, but the table lists PTLU as 0 and 0.5 and lists HΔH and Y-discrepancy as equal across the two instances. Please correct the table or the prose.
  2. [§4] In the concluding remarks, the quantity is called 'PLTU' instead of 'PTLU', and the phrase 'requires the precise knowledge' contains a typo in an earlier section; a proofreading pass is needed.
  3. [§3.3, Theorem 3] The probability space in Theorem 3 is not specified: the statement says 'with probability at least 1−δ' but the proof applies Chebyshev to the posterior draw of f given the sample. Please state explicitly that the probability is over the posterior distribution of the ground-truth classifier given sm,n.
  4. [§2, Eq. (9) and general notation] The notation π_{P,Q,F|S} is used as if it were a single conditional distribution, but the conditioning object is a sample whose distribution depends on (p,q,f); this is standard in Bayesian analysis, yet the paper should define the joint law carefully enough that the Markov property used in the Appendix B decomposition is transparent.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the optimal-learner and PTLU bounds follow from the stated Bayesian model and Fano's inequality; the Wang & Mao self-citations are contextual and not load-bearing.

full rationale

The paper's central claims are derived from the Bayesian formulation rather than from their conclusions. The optimal learner in Theorem 1 is the pointwise Bayes classifier under the posterior ρ(·|s), and the proof in Appendix C directly minimizes e(A;s,q)=1−E_x E_g ρA(g(x)|x,s), so the result is a mathematical consequence of the risk definition, not an input-equivalent ansatz. PTLU is defined as the conditional entropy U(s,q)=E_{x∼q}H(ρA(·|x,s)); Theorem 2 lower-bounds sample-wise risk by U via Fano's inequality, an external information-theoretic theorem, so using PTLU as a hardness proxy is a genuine lower bound rather than a renamed risk or a fitted prediction. The EPTLU results are plug-in concentration statements; although Appendix G's Hoeffding application raises a possible dependence issue (the entropy terms depend on the full sample), that is a correctness concern, not circularity, because the claimed bound does not reduce to its own input by construction. Theorem 5 invokes Ghosal et al. (2000) and states regularity conditions (A1)–(A3); no load-bearing assumption is justified solely by the authors' own prior work. The self-citations to Wang & Mao (2022, 2025) appear only in the introduction's literature review and are not used in any proof. The concluding remark that specifying π is difficult is an acknowledged practical limitation, not a circular step.

Assumptions & free parameters 0 free parameters · 5 assumptions · 1 invented entities

The central claim relies on knowing the true prior π, the covariate-shift assumption, and finiteness of the consistent-classifier set for the empirical concentration theorem.

assumptions (5)
  • domain assumption The learner has access to a fixed prior π over (p,q,f) that is the true data-generating distribution.
    The entire Bayesian formulation, including the optimal learner and PTLU, is defined relative to π; a misspecified π yields suboptimal learners and misleading PTLU (Section 1, Model of Nature).
  • domain assumption Covariate shift: source and target share the same labeling function f.
    Stated in Section 2: 'We assume that the source and target domains share the same classifier f, namely, the setting of covariate shift.'
  • domain assumption The output space Y is finite (classification), and X is a measurable subset of R^d.
    Section 2: 'We will primarily consider the setting where Y is finite... and X is a measurable subset of R^d.'
  • ad hoc to paper For Theorem 5, X is finite and the set of source-consistent classifiers F[fp] is finite (K<∞).
    Appendix H explicitly assumes K<∞ and finite alphabet to apply posterior concentration (Lemma 2) and discrete distribution arguments.
  • standard math The mapping f -> R(g|q,f) and the entropy H(ρA(·|x,s)) are bounded and measurable, allowing Fubini and Hoeffding.
    Used implicitly in swaps of expectations (Appendix B) and Hoeffding (Appendix G).
invented entities (1)
  • Posterior Target Label Uncertainty (PTLU) and its empirical version EPTLU
    purpose: Quantify the average posterior entropy of target labels and serve as a proxy for UDA hardness.
    The value of PTLU as a hardness predictor is established within this paper via lower-bound theorems and examples; no external test of its predictive power is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Hardness of Unsupervised Domain Adaptation: Optimal Learners and Information-Theoretic Perspective." pith.science (2026). https://pith.science/paper/3TA54DOM

@misc{pith2026250706552,
  author       = {Pith},
  title        = {Pith review of: On the Hardness of Unsupervised Domain Adaptation: Optimal Learners and Information-Theoretic Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3TA54DOM}},
  note         = {Machine review of arXiv:2507.06552}
}
abstract

This paper studies the hardness of unsupervised domain adaptation (UDA) under covariate shift. We model the uncertainty that the learner faces by a distribution $\pi$ in the ground-truth triples $(p, q, f)$ -- which we call a UDA class -- where $(p, q)$ is the source -- target distribution pair and $f$ is the classifier. We define the performance of a learner as the overall target domain risk, averaged over the randomness of the ground-truth triple. This formulation couples the source distribution, the target distribution and the classifier in the ground truth, and deviates from the classical worst-case analyses, which pessimistically emphasize the impact of hard but rare UDA instances. In this formulation, we precisely characterize the optimal learner. The performance of the optimal learner then allows us to define the learning difficulty for the UDA class and for the observed sample. To quantify this difficulty, we introduce an information-theoretic quantity -- Posterior Target Label Uncertainty (PTLU) -- along with its empirical estimate (EPTLU) from the sample , which capture the uncertainty in the prediction for the target domain. Briefly, PTLU is the entropy of the predicted label in the target domain under the posterior distribution of ground-truth classifier given the observed source and target samples. By proving that such a quantity serves to lower-bound the risk of any learner, we suggest that these quantities can be used as proxies for evaluating the hardness of UDA learning. We provide several examples to demonstrate the advantage of PTLU, relative to the existing measures, in evaluating the difficulty of UDA learning.

Figures

Figures reproduced from arXiv: 2507.06552 by the authors.

Figure 1
Figure 1. Distributions p, q1 , q2 and classifiers f c of two UDA classes π 1 (left) and π 2 (right) in Example 1. As computed in Appendix F, the optimal overall target domain risks R∗ ∞ of two UDA classes are 0 and 0.125, respectively, implying that π 1 is the easiest while π 2 is relatively challenging. We then select one of the hardest UDA instances from each UDA class: (p, q1 , f 0 ) from π 1 and (p, q2 , f 0 ) from π 2 .… view at source ↗
Figure 2
Figure 2. Distributions p, q and the classifier f 0 of two UDA classes π 1 (left) and π 2 (right) in Example 2. As computed in Appendix F, we have R1∗ ∞ = 0 for π 1 while R2∗ ∞ = 0.25 for π 2 , indicating that π 2 is harder. We then select one of the hardest UDA instances, (p, q, f 0 ) and (q, p, f 0 ), from π 1 and π 2 , respectively. In the UDA instance (p, q, f 0 ), only f 0 achieves zero error on the source domain, result… view at source ↗
Figure 3
Figure 3. Visualization of distributions p 1 , p2 , q and the classifier f 0 of two UDA instances in Example 3. Example 3. Again, we define two UDA classes, π 1 and π 2 , with the latter being more difficult. In this example, we show that the novel measure—transfer exponent—fails to distinguish between these two UDA classes. Let Ω(π 1 P,Q) := {(p 1 , q)} and Ω(π 2 P,Q) := {(p 2 , q)}, where p 1 is uniformly distributed over i… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of distributions p, qc and classifiers f c in Example 4. Let H := {f c , ¯f c : c ∈ R}. Then, the overall target do￾main risk of the optimal learner is 0, implying that this UDA class has the lowest transfer difficulty. We then draw one of the hardest UDA…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 27 canonical work pages

  1. [1]

    f-domain adversarial learning: Theory and algorithms

    David Acuna, Guojun Zhang, Marc T Law, and Sanja Fidler. f-domain adversarial learning: Theory and algorithms. In International Conference on Machine Learning, pp.\ 66--75. PMLR, 2021

    work page 2021

  2. [2]

    A model of inductive bias learning

    Jonathan Baxter. A model of inductive bias learning. Journal of artificial intelligence research, 12: 0 149--198, 2000

    work page 2000

  3. [3]

    On the hardness of domain adaptation and the utility of unlabeled target samples

    Shai Ben-David and Ruth Urner. On the hardness of domain adaptation and the utility of unlabeled target samples. In International Conference on Algorithmic Learning Theory, pp.\ 139--153. Springer, 2012

    work page 2012

  4. [4]

    Domain adaptation--can quantity compensate for quality? Annals of Mathematics and Artificial Intelligence, 70: 0 185--202, 2014

    Shai Ben-David and Ruth Urner. Domain adaptation--can quantity compensate for quality? Annals of Mathematics and Artificial Intelligence, 70: 0 185--202, 2014

    work page 2014

  5. [5]

    Analysis of representations for domain adaptation

    Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006

    2006

  6. [6]

    A theory of learning from different domains

    Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79: 0 151--175, 2010

    work page 2010

  7. [7]

    Learning bounds for domain adaptation

    John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman. Learning bounds for domain adaptation. Advances in neural information processing systems, 20, 2007

    work page 2007

  8. [8]

    Learning bounds for importance weighting

    Corinna Cortes, Yishay Mansour, and Mehryar Mohri. Learning bounds for importance weighting. Advances in neural information processing systems, 23, 2010

    work page 2010

Show all 38 references
  1. [9]

    Elements of information theory

    Thomas M Cover. Elements of information theory. John Wiley & Sons, 1999

    1999

  2. [10]

    Eine information’s theoretische ungleichung und ihre anwendung auf den beweis der ergodizitat von markoschen ketten, magyar tud

    I Csiszar. Eine information’s theoretische ungleichung und ihre anwendung auf den beweis der ergodizitat von markoschen ketten, magyar tud. Akad. Mat, 1963

    1963

  3. [11]

    The value of out-of-distribution data

    Ashwin De Silva, Rahul Ramesh, Carey Priebe, Pratik Chaudhari, and Joshua T Vogelstein. The value of out-of-distribution data. In International Conference on Machine Learning, pp.\ 7366--7389. PMLR, 2023

    2023

  4. [12]

    Class notes for course 6.574: Transmission of information

    RM Fano. Class notes for course 6.574: Transmission of information. Lecture Notes, 1952

    1952

  5. [13]

    Unsupervised domain adaptation by backpropagation

    Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pp.\ 1180--1189. PMLR, 2015

    2015

  6. [14]

    Ghosh, and Aad W

    Subhashis Ghosal, Jayanta K. Ghosh, and Aad W. van der Vaart. Convergence rates of posterior distributions . The Annals of Statistics, 28 0 (2): 0 500 -- 531, 2000

    2000

  7. [15]

    On the value of target data in transfer learning

    Steve Hanneke and Samory Kpotufe. On the value of target data in transfer learning. Advances in Neural Information Processing Systems, 32, 2019

    2019

  8. [16]

    Algorithms and theory for multiple-source adaptation

    Judy Hoffman, Mehryar Mohri, and Ningshan Zhang. Algorithms and theory for multiple-source adaptation. Advances in neural information processing systems, 31, 2018

    2018

  9. [17]

    Correcting sample selection bias by unlabeled data

    Jiayuan Huang, Arthur Gretton, Karsten Borgwardt, Bernhard Sch \"o lkopf, and Alex Smola. Correcting sample selection bias by unlabeled data. Advances in neural information processing systems, 19, 2006

    2006

  10. [18]

    Unsupervised domain adaptation based on source-guided discrepancy

    Seiichi Kuroki, Nontawat Charoenphakdee, Han Bao, Junya Honda, Issei Sato, and Masashi Sugiyama. Unsupervised domain adaptation based on source-guided discrepancy. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pp.\ 4122--4129, 2019

    2019

  11. [19]

    Coupled generative adversarial networks

    Ming-Yu Liu and Oncel Tuzel. Coupled generative adversarial networks. Advances in neural information processing systems, 29, 2016

    2016

  12. [20]

    Domain adaptation: Learning bounds and algorithms

    Yishay Mansour, Mehryar Mohri, and Afshin Rostamizadeh. Domain adaptation: Learning bounds and algorithms. In Proceedings of the 22nd Annual Conference on Learning Theory (COLT), 2009

    2009

  13. [21]

    Multiple source adaptation and the r \'e nyi divergence

    Yishay Mansour, Mehryar Mohri, and Afshin Rostamizadeh. Multiple source adaptation and the r \'e nyi divergence. arXiv preprint arXiv:1205.2628, 2012

    2012 arXiv

  14. [22]

    Some pac-bayesian theorems

    David A McAllester. Some pac-bayesian theorems. In Proceedings of the eleventh annual conference on Computational learning theory, pp.\ 230--234, 1998

    1998

  15. [23]

    New analysis and algorithm for learning with drifting distributions

    Mehryar Mohri and Andres Mu \ n oz Medina. New analysis and algorithm for learning with drifting distributions. In Algorithmic Learning Theory: 23rd International Conference, ALT 2012, Lyon, France, October 29-31, 2012. Proceedings 23, pp.\ 124--138. Springer, 2012

    2012

  16. [24]

    Kl guided domain adaptation

    A Tuan Nguyen, Toan Tran, Yarin Gal, Philip HS Torr, and At l m G \"u ne s Baydin. Kl guided domain adaptation. arXiv preprint arXiv:2106.07780, 2021

    2021 arXiv

  17. [25]

    Statistical aspects of wasserstein distances

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

    2019

  18. [26]

    A pac analysis of a bayesian estimator

    John Shawe-Taylor and Robert C Williamson. A pac analysis of a bayesian estimator. In Proceedings of the tenth annual conference on Computational learning theory, pp.\ 2--9, 1997

    1997

  19. [27]

    Wasserstein distance guided representation learning for domain adaptation

    Jian Shen, Yanru Qu, Weinan Zhang, and Yong Yu. Wasserstein distance guided representation learning for domain adaptation. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  20. [28]

    Improving predictive inference under covariate shift by weighting the log-likelihood function

    Hidetoshi Shimodaira. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of statistical planning and inference, 90 0 (2): 0 227--244, 2000

    2000

  21. [29]

    A novel domain adaptation theory with jensen--shannon divergence

    Changjian Shui, Qi Chen, Jun Wen, Fan Zhou, Christian Gagn \'e , and Boyu Wang. A novel domain adaptation theory with jensen--shannon divergence. Knowledge-Based Systems, 257: 0 109808, 2022

    2022

  22. [30]

    Bridging the gap between f-gans and wasserstein gans

    Jiaming Song and Stefano Ermon. Bridging the gap between f-gans and wasserstein gans. In International Conference on Machine Learning, pp.\ 9078--9087. Pmlr, 2020

    2020

  23. [31]

    A theory of the learnable

    Leslie G Valiant. A theory of the learnable. Communications of the ACM, 27 0 (11): 0 1134--1142, 1984

    1984

  24. [32]

    Information-theoretic analysis of unsupervised domain adaptation

    Ziqiao Wang and Yongyi Mao. Information-theoretic analysis of unsupervised domain adaptation. arXiv preprint arXiv:2210.00706, 2022

    2022 arXiv

  25. [33]

    On f -divergence principled domain adaptation: An improved framework

    Ziqiao Wang and Yongyi Mao. On f -divergence principled domain adaptation: An improved framework. Advances in Neural Information Processing Systems, 37: 0 6711--6748, 2025

    2025

  26. [34]

    Bridging theory and algorithm for domain adaptation

    Yuchen Zhang, Tianle Liu, Mingsheng Long, and Michael Jordan. Bridging theory and algorithm for domain adaptation. In International conference on machine learning, pp.\ 7404--7413. PMLR, 2019

    2019

  27. [35]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  28. [36]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  29. [37]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  30. [38]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

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