REVIEW 5 major objections 6 minor 125 references
On the Transferability and Discriminability of Repersentation Learning in Unsupervised Domain Adaptation
T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read This paper proves that adversarial domain adaptation needs an explicit target-discriminability loss, and instantiates that requirement as RLGLC, which pairs a relaxed Wasserstein alignment with a contrastive estimator of conditional…
desk verdict The empirical package is worth a look, but the theoretical core does not hold: the key definitions need a joint distribution that the experiments never provide, and the proofs contain internal contradictions. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The engine of the argument is the conditional-mutual-information pair of Definitions 4.1 and 4.2 together with the inequality chain in the proof of Theorem 8.1. Discriminability of $Z_u$ is measured by $I(X_s; X_t \mid Z_u)$ and transferability by $I(X_t; Z_t \mid X_s)$ and $I(X_s; Z_s \mid X_t)$; the proof shows both transferability terms are bounded above by the KL divergence between the feature distributions, so the alignment term of the standard objective decreases them, while the source classification loss acts as a surrogate lower bound on $I(Z_s; Y)$ and thereby, through Proposition 8.1, drives $I(X_s; X_t \mid Z_s)$ toward zero. On the method side, two instruments carry the instantiation: AR-WWD, a Wasserstein distance whose ground metric is itself a Wasserstein distance and whose alignment constraint is relaxed from 'equal distributions' to 'target contained in source' to absorb class imbalance, and the Local Consistency Module, which estimates the missing term $I(X_s; X_t \mid Z_t)$ with a conditional noise-contrastive estimator that Proposition 4.1 shows to be a lower bound, saturated by the log-ratio score function, and tight as the number of negative samples $K$ grows.
What would settle it
Keep the model frozen and compute the CNCE estimator of Equation (11) on the same mini-batch under two different pairing rules for source and target samples — the nearest-neighbor pairing the implementation uses versus a random pairing — and if the estimator's value shifts materially with the pairing, then the target discriminability the Local Consistency Module is built to maximize is an artifact of the pairing heuristic rather than a well-defined property of the two domains, and the necessity theorem built on it does not apply to the unpaired setting.
Extended reading notes
Core claim
The central claim, stated as Theorem 4.1, is that representations obtained by minimizing the standard UDA objective — a distributional discrepancy $D(P_\phi^s(Z_s), P_\phi^t(Z_t))$ plus a source classification loss $L_{cl}$ — have increased transferability in both domains but increased discriminability only in the source domain. The proof bounds the transferability terms $I(X_s; Z_s \mid X_t)$ and $I(X_t; Z_t \mid X_s)$ by the KL divergence between the induced feature distributions, so the alignment term of the objective reduces them; at the same time, the source loss acts as a surrogate lower bound on $I(Z_s; Y)$, and the paper's Proposition 8.1 ties full source discriminability, $I(X_s; X_t \mid Z_s)=0$, to $I(Z_s; Y)=I(X_s; Y)$, so the source loss drives source discriminability as a by-product. No term in the objective touches $I(X_s; X_t \mid Z_t)$, so the target representation can satisfy the alignment constraint while retaining only part of the task-relevant information — transferable but class-blind. From this the paper concludes that an explicit target-discriminability loss is necessary rather than heuristic, and that a good representation for UDA should be defined as one minimizing all four information terms simultaneously, which is exactly what the proposed framework in its Equation (4) optimizes.
Load-bearing premise
The load-bearing premise is that a joint distribution linking source and target samples exists, so that quantities like $I(X_s; X_t \mid Z_u)$ are defined; the paper's motivating example pairs MRI with X-ray from the same patient, but the paper never states a joint distribution for the unpaired benchmarks where the method is actually tested.
Editorial extensions
If this is right
- Any adversarial UDA method that stops at distribution alignment plus source risk is, by Theorem 4.1, provably leaving target-domain discriminability unconstrained, so adding a target-oriented loss is a requirement rather than a heuristic.
- The Bayes-error bounds of Theorems 5.1 and 5.2 express the average target error as an upper bound driven by the four information terms, so RLGLC's objective is minimizing the exact quantities that bound error.
- The CNCE estimator is a reusable component — a lower bound on $I(X_s; X_t \mid Z_t)$ that is tight as the number of negatives grows — and the paper's '+LM' experiments show it can be grafted onto existing adversarial baselines with consistent gains.
- AR-WWD's containment relaxation implies strict distributional equality is actively harmful when class ratios differ across domains, because exact alignment forces target samples across decision boundaries.
Reading between the lines
- Taken at full strength, the theorem unifies a family of prior additions to UDA: pseudo-labeling, entropy minimization, batch-spectral penalization, and gradient harmonization all look like indirect ways of supplying the same missing $I(X_s; X_t \mid Z_t)$ term, which would make this paper's contribution an explanation of why those tricks work rather than another trick.
- A testable consequence the authors do not draw: the benefit of a target-discriminability term should grow with the class-imbalance gap between domains, since that is the regime where alignment alone most aggressively destroys target structure; a controlled synthetic shift could verify that ordering.
- Because the definitions require a joint distribution over $(X_s, X_t)$ that the paper never supplies, a natural sharpening would be to define the coupling explicitly — for instance, through the transport plan that AR-WWD itself estimates — which would turn the formal mutual information into an empirical and checkable quantity.
- The CNCE estimator is a conditional variant of the contrastive objective used in self-supervised representation learning, so the optimality result in Proposition 4.1 effectively imports the log-ratio critic into UDA; the local consistency term appears to operate by 'explaining away' domain-specific information once the target feature is observed.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes an information-theoretic framework for unsupervised domain adaptation (UDA). It defines 'discriminability' and 'transferability' via conditional mutual information terms (Definitions 4.1 and 4.2), claims Theorem 4.1 that the standard adversarial objective (distribution alignment plus source-domain classification loss) improves source discriminability but only target transferability, and concludes that an explicit target-domain discriminability loss is necessary. It then introduces the RLGLC method, which combines an asymmetrically-relaxed Wasserstein distance (AR-WWD) with a conditional noise-contrastive-estimation-based local consistency module, and reports strong empirical results on Office-31, Office-Home, VisDA-2017, DomainNet, Digits, and on semantic segmentation and object detection tasks.
Significance. If the theoretical claims were correct, the paper would make a substantial contribution by explaining why target-domain discriminability losses help adversarial UDA, and RLGLC would be a strong practical instantiation. The experimental program is broad, with many benchmarks, ablations, and statistical tests, and the proposed method shows consistent improvements. However, the theoretical core has fundamental flaws that are load-bearing for the main claim: the definitions require a joint distribution over unpaired variables that is never specified, the proof of Theorem 4.1 contains algebraic errors, Proposition 4.1 is self-contradictory, and the Bayes-error bounds in Theorem 5.1 are not valid probability bounds. These issues cannot be fixed by local revision, so the paper cannot be recommended for acceptance.
major comments (5)
- [Section 4.1 (Definitions 4.1, 4.2) and Assumption 4.1] The quantities I(Xs;Xt|Zu), I(Xs;Zs|Xt), and I(Xt;Zt|Xs) are not well-defined because the manuscript never specifies a joint distribution over (Xs, Xt). Section 3.1 defines only the marginals Ps(X) and Pt(X), and the experiments use standard unpaired UDA benchmarks. Under the canonical product coupling that is consistent with unsupervised sampling, Xs and Xt are independent, so I(Xs;Xt|Zu)=0 for every representation Zu; this would make every target representation perfectly discriminative by Definition 4.1 and would contradict Assumption 4.1, which requires I(Xs;Xt)=H(Y)>0. The motivating paired MRI/X-ray example in Section 4.1 is informal and never formalized, and the proof of Theorem 8.1 in the appendix relies on the undefined expectation E_{P(Xs,Xt)}. Without a specified coupling, Theorem 4.1 and all subsequent information-theoretic results are not well-posed for the UDA setting studied.
- [Appendix Theorem 8.1, Eq. (36)] The chain of inequalities bounding I(Xs;Zs|Xt) contains an algebraic error: the line 'E_{P(Xs,Xt)} log P(Xt)/P(Xs) = -KL(P(Xs)||P(Xt))' is false for marginals Ps and Pt; the left-hand side equals H(Ps)-H(Pt) under any coupling with those marginals, which is not the negative KL divergence. In addition, the paper's claim in Section 4.2 that minimizing KL(P(Zs)||P(Zt)) and KL(P(Zt)||P(Zs)) is 'equivalent' to minimizing the Wasserstein distance between P(Zs) and P(Zt) is unsupported. Consequently, the proof of Theorem 4.1 does not establish that the standard objective improves only the transferability of Zt.
- [Proposition 4.1 (Section 4.3.2)] Part (2) of Proposition 4.1 asserts that for any finite K there exists a critic φ* with sup_φ I_CNCE = I(Xs;Xt|Zt) exactly. The paper's own proof in Appendix Theorem 8.3 contradicts this: Step 5 there states that for finite K the quantity is at most I and that 'equality holds in the limit' as K→∞. Standard InfoNCE-style variational bounds are strictly loose for finite K. This matters because Eq. (16) uses I_CNCE with finite mini-batch K as a stand-in for the discriminability term I(Xs;Xt|Zt), so the claimed exactness is false and only an approximation is available.
- [Theorem 5.1 and Appendix Theorem 8.2] The claimed upper bounds in Eqs. (18)-(21) are not valid probability bounds. The exponent -H(Y)+I can be positive when I>H(Y), making the right-hand side negative and hence not an upper bound on the nonnegative error probability. The proof in Appendix Theorem 8.2 says 'we get a form: \bar{P}_e ≥ exp[-H(Y)+I] ... Inverting this inequality gives the upper bound: \bar{P}_e ≤ 1 - exp[-H(Y)+I]', which is a logical non sequitur. The proof also defers the crucial steps to 'detailed technical steps' that are never given. Thus, the claim that minimizing the RLGLC objective reduces the target Bayes error is unsupported.
- [Abstract and Theorem 4.1] The central claim that an additional target-domain discriminability loss is 'necessary' is largely a definitional artifact. Once discriminability is defined as I(Xs;Xt|Zu) (Definition 4.1), any framework that includes I(Xs;Xt|Zt) in its objective (Eq. (4)) will increase the metric by construction, and the standard objective (2), which omits it, will not. Theorem 4.1 does not demonstrate operational necessity; it only shows that the standard objective does not optimize the paper's chosen metric. An independent characterization of target discriminability (for example, in terms of Bayes error) would be needed to support the necessity claim.
minor comments (6)
- [Title] The title contains a typo: 'Repersentation' should be 'Representation'.
- [Section 4.1 and Section 4.3.1] Section 4.1 uses 'Defination' instead of 'Definition', and Section 4.3.1 writes 'per-given hyper-parameter' where 'pre-given' is intended.
- [Table 2] Table 2, row TCPL, column R→P: the value 55.8 is markedly lower than the values in neighboring rows and columns and appears to be a typographical error.
- [Section 4.3.2] The sentence 'Because that {Zt, Xt} is a pair, thus, we implement P(Zt, Xs, Xt) as {Zt, Xt, Xs,j}...' is grammatically unclear and technically ad hoc; the probabilistic model relating the positive and negative samples in Eq. (11) should be specified precisely.
- [Section 5, final paragraph] The claim that RLGLC 'offers a more precise measurement' of I(Xs;Zs|Xt) and I(Xt;Zt|Xs) is not defined or derived anywhere, and these quantities are not estimated by the proposed objective.
- [Section 6.4] The Friedman test is reported with many critical values; a single summary table or a concise statement of the test outcome for each dataset would make the statistical analysis easier to verify.
Circularity Check
The claimed necessity of a target-discriminability loss is substantially built into the paper's own definitions: target discriminability is defined as I(Xs;Xt|Zt), and the proposed framework Eq. (5) inserts exactly that term as the 'additional' loss, so Theorem 4.1's central conclusion is partly definitional rather than independently derived.
-
self definitional
[Section 4.1, Definition 4.1 and Theorem 4.1; Section 4.2, Eq. (4)-(5); Appendix Theorem 8.1]
"Definition 4.1. Discriminability: The amount of discriminative information contained in the representation Zu of Xu can be defined as I(Xs;Xt∣Zu), u∈{s,t}. The smaller I(Xs;Xt∣Zu) is, the more discriminative information Zu contains. ... However, there is no obvious term to constrain the learned sample feature representations of the target domain to be with discriminability."
The central conclusion -- that objective (2) needs an additional target-discriminability loss -- follows directly from the paper's own definitions. 'Discriminability' of the target is defined as I(Xs;Xt|Zt), and the framework Eq. (4) is assembled to minimize exactly the four quantities I(Xs;Zs|Xt), I(Xt;Zt|Xs), I(Xs;Xt|Zs), I(Xs;Xt|Zt). Eq. (5) then rewrites the first three as existing UDA terms D and Lcl, leaving I(Xs;Xt|Zt) as the residual 'needed' term. Thus the proof of Theorem 8.1 shows only that alignment reduces the transferability terms and that source Lcl is related to I(Xs;Xt|Zs); the assertion that target discriminability is missing is tautological once target discriminability is defined as I(Xs;Xt|Zt) and no such term appears in Eq. (2).
full rationale
The paper's formal definition of 'good representation learning' is Definition 4.1/4.2, and Eq. (4) is constructed so that the four minimized information terms are exactly the transferability and discriminability measures. Consequently Theorem 4.2 is true by definition, and Eq. (5) reveals I(Xs;Xt|Zt) as the 'necessary' addition. This makes the headline theoretical claim substantially self-definitional: the conclusion that a target-discriminability loss is necessary is essentially the statement that the authors' chosen discriminability measure is absent from the standard objective. The paper does contain non-circular components: the CNCE estimator is a standard variational lower bound, the +LM ablations (DANN+LM, SWD+LM, CAN+LM, MDD+LM) provide independent empirical support for the value of the local module, and AR-WWD is a concrete algorithmic contribution. A separate non-circular concern is that the information-theoretic quantities require a joint distribution P(Xs,Xt) that is never specified for the unpaired UDA benchmarks, which is a correctness gap rather than a circularity. On balance, the central 'necessity' claim is partly baked into the definitions, so a moderate score of 6 is appropriate; the independent ablations and algorithmic work prevent this from being a fully circular paper.
Assumptions & free parameters
free parameters (3)
- alpha (regularization weight in Eq 16) =
1
- beta (AR-WWD relaxation hyperparameter) =
0.4
- lambda (gradient penalty coefficient in Eq 9) =
10
assumptions (5)
- domain assumption Assumption 4.1: I(Xs∪t;Y) = I(Xs;Y) = I(Xt;Y) = H(Y)
- ad hoc to paper Existence of a joint distribution P(Xs, Xt) with a paired interpretation
- domain assumption Markov chain Zt <- Xt <-> Y <-> Xs -> Zs (Lemma 8.1)
- ad hoc to paper Minimizing KL(P(Zs)||P(Zt)) is equivalent to minimizing the Wasserstein distance
- ad hoc to paper Decoupling Xu into label-related XY and label-unrelated X_U^u_Y, with XY shared across domains and independent of X_U^u_Y
Cite this review
Pith. "Pith review of On the Transferability and Discriminability of Repersentation Learning in Unsupervised Domain Adaptation." pith.science (2026). https://pith.science/paper/DI4PO2D4
@misc{pith2026250522099,
author = {Pith},
title = {Pith review of: On the Transferability and Discriminability of Repersentation Learning in Unsupervised Domain Adaptation},
year = {2026},
howpublished = {\url{https://pith.science/paper/DI4PO2D4}},
note = {Machine review of arXiv:2505.22099}
}
read the original abstract
In this paper, we addressed the limitation of relying solely on distribution alignment and source-domain empirical risk minimization in Unsupervised Domain Adaptation (UDA). Our information-theoretic analysis showed that this standard adversarial-based framework neglects the discriminability of target-domain features, leading to suboptimal performance. To bridge this theoretical-practical gap, we defined "good representation learning" as guaranteeing both transferability and discriminability, and proved that an additional loss term targeting target-domain discriminability is necessary. Building on these insights, we proposed a novel adversarial-based UDA framework that explicitly integrates a domain alignment objective with a discriminability-enhancing constraint. Instantiated as Domain-Invariant Representation Learning with Global and Local Consistency (RLGLC), our method leverages Asymmetrically-Relaxed Wasserstein of Wasserstein Distance (AR-WWD) to address class imbalance and semantic dimension weighting, and employs a local consistency mechanism to preserve fine-grained target-domain discriminative information. Extensive experiments across multiple benchmark datasets demonstrate that RLGLC consistently surpasses state-of-the-art methods, confirming the value of our theoretical perspective and underscoring the necessity of enforcing both transferability and discriminability in adversarial-based UDA.
Figures
Figures from the paper (1 more)
Reference graph
Works this paper leans on
-
[1]
Domain-adversarial training of neural networks,
Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Lavi- olette, M. Marchand, and V . Lempitsky, “Domain-adversarial training of neural networks,”The journal of machine learning research, vol. 17, no. 1, pp. 2096–2030, 2016
2016
-
[2]
Unsupervised multi-class domain adaptation: Theory, algorithms, and practice,
Y . Zhang, B. Deng, H. Tang, L. Zhang, and K. Jia, “Unsupervised multi-class domain adaptation: Theory, algorithms, and practice,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020
2020
-
[3]
Unified optimal transport framework for universal domain adaptation,
W. Chang, Y . Shi, H. D. Tuan, and J. Wang, “Unified optimal transport framework for universal domain adaptation,”Conference on Neural Information Processing Systems, 2022
2022
-
[4]
Transvqa: Transferable vector quantization alignment for unsupervised domain adaptation,
Y . Sun, W. Dong, X. Li, L. Dong, G. Shi, and X. Xie, “Transvqa: Transferable vector quantization alignment for unsupervised domain adaptation,”IEEE Transactions on Image Processing, 2024
2024
-
[5]
Learning transferable conceptual prototypes for interpretable unsupervised domain adaptation,
J. Gao, X. Ma, and C. Xu, “Learning transferable conceptual prototypes for interpretable unsupervised domain adaptation,”IEEE Transactions on Image Processing, 2024
2024
-
[6]
Wasserstein distance guided representation learning for domain adaptation,
J. Shen, Y . Qu, W. Zhang, and Y . Yu, “Wasserstein distance guided representation learning for domain adaptation,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018
2018
-
[7]
Wasserstein generative adversarial networks,
M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein generative adversarial networks,” inInternational conference on machine learning. PMLR, 2017, pp. 214–223
2017
-
[8]
Bridging theory and algorithm for domain adaptation,
Y . Zhang, T. Liu, M. Long, and M. Jordan, “Bridging theory and algorithm for domain adaptation,” inInternational Conference on Machine Learning. PMLR, 2019, pp. 7404–7413
2019
Show all 125 references
-
[9]
Probability- polarized optimal transport for unsupervised domain adaptation,
Y . Wang, C.-X. Ren, Y .-M. Zhai, Y .-W. Luo, and H. Yan, “Probability- polarized optimal transport for unsupervised domain adaptation,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 14, 2024, pp. 15 653–15 661
2024
-
[10]
Prompt-based distribution alignment for unsupervised domain adaptation,
S. Bai, M. Zhang, W. Zhou, S. Huang, Z. Luan, D. Wang, and B. Chen, “Prompt-based distribution alignment for unsupervised domain adaptation,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 2, 2024, pp. 729–737
2024
-
[11]
Unsupervised and semi-supervised robust spherical space domain adaptation,
X. Gu, J. Sun, and Z. Xu, “Unsupervised and semi-supervised robust spherical space domain adaptation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 3, pp. 1757–1774, 2022
2022
-
[12]
Pseudo-calibration: Improving predictive uncertainty estimation in unsupervised domain adaptation,
D. Hu, J. Liang, X. Wang, and C.-S. Foo, “Pseudo-calibration: Improving predictive uncertainty estimation in unsupervised domain adaptation,” in Forty-first International Conference on Machine Learning, 2024
2024
-
[13]
Class-incremental unsupervised domain adaptation via pseudo-label distillation,
K. Wei, X. Yang, Z. Xu, and C. Deng, “Class-incremental unsupervised domain adaptation via pseudo-label distillation,”IEEE Transactions on Image Processing, 2024
2024
-
[14]
A versatile framework for unsupervised domain adaptation based on instance weighting,
J. Zhu, F. Ye, Q. Xiao, P. Guo, Y . Zhang, and Q. Yang, “A versatile framework for unsupervised domain adaptation based on instance weighting,”IEEE Transactions on Image Processing, 2024
2024
-
[15]
Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation,
D. Jiahua, C. Yang, Z. Fang, and D. Zhengming, “Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022
2022
-
[16]
Contrastive adap- tation network for unsupervised domain adaptation,
G. Kang, L. Jiang, Y . Yang, and A. G. Hauptmann, “Contrastive adap- tation network for unsupervised domain adaptation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 4893–4902
2019
-
[17]
Robust local preserving and global aligning network for adversarial domain adaptation,
W. Qiang, J. Li, C. Zheng, B. Su, and H. Xiong, “Robust local preserving and global aligning network for adversarial domain adaptation,”IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 3, pp. 3014–3029, 2021
2021
-
[18]
Auxiliary task guided mean and covariance alignment network for adversarial domain adaptation,
W. Qiang, J. Li, C. Zheng, and B. Su, “Auxiliary task guided mean and covariance alignment network for adversarial domain adaptation,” Knowledge-Based Systems, vol. 223, p. 107066, 2021
2021
-
[19]
Representation learning with contrastive predictive coding,
A. v. d. Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,”arXiv preprint arXiv:1807.03748, 2018
2018 arXiv
-
[20]
A simple framework for contrastive learning of visual representations,
T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” inInternational conference on machine learning. PMLR, 2020, pp. 1597–1607
2020
-
[21]
Intriguing properties of contrastive losses,
T. Chen and L. Li, “Intriguing properties of contrastive losses,”arXiv preprint arXiv:2011.02803, 2020
2011 arXiv
-
[22]
Understanding self-training for gradual domain adaptation,
A. Kumar, T. Ma, and P. Liang, “Understanding self-training for gradual domain adaptation,” inInternational Conference on Machine Learning. PMLR, 2020, pp. 5468–5479
2020
-
[23]
Margin-aware adversarial domain adaptation with optimal transport,
S. Dhouib, I. Redko, and C. Lartizien, “Margin-aware adversarial domain adaptation with optimal transport,” inInternational Conference on Machine Learning. PMLR, 2020, pp. 2514–2524
2020
-
[24]
Robust optimal transport with applications in generative modeling and domain adaptation,
Y . Balaji, R. Chellappa, and S. Feizi, “Robust optimal transport with applications in generative modeling and domain adaptation,”arXiv preprint arXiv:2010.05862, 2020
2010 arXiv
-
[25]
Gradually vanishing bridge for adversarial domain adaptation,
S. Cui, S. Wang, J. Zhuo, C. Su, Q. Huang, and Q. Tian, “Gradually vanishing bridge for adversarial domain adaptation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 12 455–12 464
2020
-
[26]
Domain adaptation with conditional distribution matching and generalized label shift,
R. T. d. Combes, H. Zhao, Y .-X. Wang, and G. Gordon, “Domain adaptation with conditional distribution matching and generalized label shift,”arXiv preprint arXiv:2003.04475, 2020
2003 arXiv
-
[27]
Heuristic domain adaptation,
S. Cui, X. Jin, S. Wang, Y . He, and Q. Huang, “Heuristic domain adaptation,”arXiv preprint arXiv:2011.14540, 2020
2011 arXiv
-
[28]
Unsupervised domain adaptation with hierarchical gradient synchronization,
L. Hu, M. Kan, S. Shan, and X. Chen, “Unsupervised domain adaptation with hierarchical gradient synchronization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 4043–4052
2020
-
[29]
Pixel- level cycle association: A new perspective for domain adaptive semantic segmentation,
G. Kang, Y . Wei, Y . Yang, Y . Zhuang, and A. G. Hauptmann, “Pixel- level cycle association: A new perspective for domain adaptive semantic segmentation,”arXiv preprint arXiv:2011.00147, 2020
2011 arXiv
-
[30]
Unsupervised domain adaptation via structurally regularized deep clustering,
H. Tang, K. Chen, and K. Jia, “Unsupervised domain adaptation via structurally regularized deep clustering,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 8725– 8735. SUBMITTED TO IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHI...
2020
-
[31]
A kernel two-sample test,
A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Schölkopf, and A. Smola, “A kernel two-sample test,”The Journal of Machine Learning Research, vol. 13, no. 1, pp. 723–773, 2012
2012
-
[32]
Deep domain confusion: Maximizing for domain invariance,
E. Tzeng, J. Hoffman, N. Zhang, K. Saenko, and T. Darrell, “Deep domain confusion: Maximizing for domain invariance,”arXiv preprint arXiv:1412.3474, 2014
2014 arXiv
-
[33]
Return of frustratingly easy domain adaptation,
B. Sun, J. Feng, and K. Saenko, “Return of frustratingly easy domain adaptation,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 30, no. 1, 2016
2016
-
[34]
Unsupervised domain adaptation based on source-guided discrepancy,
S. Kuroki, N. Charoenphakdee, H. Bao, J. Honda, I. Sato, and M. Sugiyama, “Unsupervised domain adaptation based on source-guided discrepancy,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 33, no. 01, 2019, pp. 4122–4129
2019
-
[35]
Sliced wasserstein discrepancy for unsupervised domain adaptation,
C.-Y . Lee, T. Batra, M. H. Baig, and D. Ulbricht, “Sliced wasserstein discrepancy for unsupervised domain adaptation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 10 285–10 295
2019
-
[36]
Domain adaptation with asymmetrically-relaxed distribution alignment,
Y . Wu, E. Winston, D. Kaushik, and Z. Lipton, “Domain adaptation with asymmetrically-relaxed distribution alignment,” inInternational Conference on Machine Learning. PMLR, 2019, pp. 6872–6881
2019
-
[37]
Reliable weighted optimal transport for unsupervised domain adaptation,
R. Xu, P. Liu, L. Wang, C. Chen, and J. Wang, “Reliable weighted optimal transport for unsupervised domain adaptation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 4394–4403
2020
-
[38]
Enhanced transport distance for unsupervised domain adaptation,
M. Li, Y .-M. Zhai, Y .-W. Luo, P.-F. Ge, and C.-X. Ren, “Enhanced transport distance for unsupervised domain adaptation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 13 936–13 944
2020
-
[39]
Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation,
X. Chen, S. Wang, M. Long, and J. Wang, “Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation,” inProceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, ser. P...
2019
-
[40]
Gradient harmonization in unsuper- vised domain adaptation,
F. Huang, S. Song, and L. Zhang, “Gradient harmonization in unsuper- vised domain adaptation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024
2024
-
[41]
Cycle self-training for domain adaptation,
H. Liu, J. Wang, and M. Long, “Cycle self-training for domain adaptation,”Advances in Neural Information Processing Systems, vol. 34, pp. 22 968–22 981, 2021
2021
-
[42]
Domain adaptation: Learning bounds and algorithms,
Y . Mansour, M. Mohri, and A. Rostamizadeh, “Domain adaptation: Learning bounds and algorithms,”arXiv preprint arXiv:0902.3430, 2009
2009 arXiv
-
[43]
A theory of learning from different domains,
S. Ben-David, J. Blitzer, K. Crammer, A. Kulesza, F. Pereira, and J. W. Vaughan, “A theory of learning from different domains,”Machine learning, vol. 79, no. 1, pp. 151–175, 2010
2010
-
[44]
New analysis and algorithm for learning with drifting distributions,
M. Mohri and A. M. Medina, “New analysis and algorithm for learning with drifting distributions,” inInternational Conference on Algorithmic Learning Theory. Springer, 2012, pp. 124–138
2012
-
[45]
A pac-bayesian approach for domain adaptation with specialization to linear classifiers,
P. Germain, A. Habrard, F. Laviolette, and E. Morvant, “A pac-bayesian approach for domain adaptation with specialization to linear classifiers,” inInternational conference on machine learning. PMLR, 2013, pp. 738–746
2013
-
[46]
Adaptation algorithm and theory based on generalized discrepancy,
C. Cortes, M. Mohri, and A. Muñoz Medina, “Adaptation algorithm and theory based on generalized discrepancy,” inProceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2015, pp. 169–178
2015
-
[47]
Theoretical analysis of domain adaptation with optimal transport,
I. Redko, A. Habrard, and M. Sebban, “Theoretical analysis of domain adaptation with optimal transport,” inJoint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 2017, pp. 737–753
2017
-
[48]
On learn- ing invariant representations for domain adaptation,
H. Zhao, R. T. Des Combes, K. Zhang, and G. Gordon, “On learn- ing invariant representations for domain adaptation,” inInternational Conference on Machine Learning. PMLR, 2019, pp. 7523–7532
2019
-
[49]
The information bottleneck method,
N. Tishby, F. C. Pereira, and W. Bialek, “The information bottleneck method,”arXiv preprint physics/0004057, 2000
2000 arXiv
-
[50]
An information theoretic framework for multi-view learning,
K. Sridharan and S. M. Kakade, “An information theoretic framework for multi-view learning,” inCOLT, no. 114, 2008, pp. 403–414
2008
-
[51]
A survey on multi-view learning,
C. Xu, D. Tao, and C. Xu, “A survey on multi-view learning,”arXiv preprint arXiv:1304.5634, 2013
2013 arXiv
-
[52]
Self- supervised learning from a multi-view perspective,
Y .-H. H. Tsai, Y . Wu, R. Salakhutdinov, and L.-P. Morency, “Self- supervised learning from a multi-view perspective,”arXiv preprint arXiv:2006.05576, 2020
2006 arXiv
-
[53]
T. M. Cover and J. A. Thomas,Elements of Information Theory. John Wiley & Sons, 2012
2012
-
[54]
Representation learning: A review and new perspectives,
Y . Bengio, A. Courville, and P. Vincent, “Representation learning: A review and new perspectives,”IEEE transactions on pattern analysis and machine intelligence, vol. 35, no. 8, pp. 1798–1828, 2013
2013
-
[55]
Learning deep features for discriminative localization,
B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learning deep features for discriminative localization,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 2921–2929
2016
-
[56]
Focal loss for dense object detection,
T. Lin, “Focal loss for dense object detection,”arXiv preprint arXiv:1708.02002, 2017
2017 arXiv
-
[57]
Efficient estimation of word representations in vector space,
T. Mikolov, “Efficient estimation of word representations in vector space,” arXiv preprint arXiv:1301.3781, vol. 3781, 2013
2013 arXiv
-
[58]
Earth mover’s distances on discrete surfaces,
J. Solomon, R. Rustamov, L. Guibas, and A. Butscher, “Earth mover’s distances on discrete surfaces,”ACM Transactions on Graphics, p. 1–12, Jul 2014. [Online]. Available: http: //dx.doi.org/10.1145/2601097.2601175
2014
-
[59]
The earth mover’s distance as a metric for image retrieval,
Y . Rubner, “The earth mover’s distance as a metric for image retrieval,” International Journal of Computer Vision, p. 99–121, Jan 2000. [Online]. Available: http://dx.doi.org/10.1023/a:1026543900054
-
[60]
Geometry of probability simplex via optimal transport,
W. Li, “Geometry of probability simplex via optimal transport,”arXiv preprint arXiv:1803.06360, vol. 2, no. 4, p. 13, 2018
2018 arXiv
-
[61]
Wasserstein of wasserstein loss for learning generative models,
Y . Dukler, W. Li, A. Lin, and G. Montúfar, “Wasserstein of wasserstein loss for learning generative models,” inInternational Conference on Machine Learning. PMLR, 2019, pp. 1716–1725
2019
-
[62]
Villaniet al.,Optimal transport: old and new
C. Villaniet al.,Optimal transport: old and new. Springer, 2009, vol. 338
2009
-
[63]
Improved training of wasserstein gans,
I. Gulrajani, F. Ahmed, M. Arjovsky, V . Dumoulin, and A. C. Courville, “Improved training of wasserstein gans,”Advances in neural information processing systems, vol. 30, 2017
2017
-
[64]
On the regularization of wasserstein gans,
H. Petzka, A. Fischer, and D. Lukovnikov, “On the regularization of wasserstein gans,”International Conference on Learning Representa- tions,International Conference on Learning Representations, Sep 2017
2017
-
[65]
Relations between entropy and error probability,
M. Feder and N. Merhav, “Relations between entropy and error probability,”IEEE Transactions on Information Theory, vol. 40, no. 1, pp. 259–266, 1994
1994
-
[66]
Deep hashing network for unsupervised domain adaptation,
H. Venkateswara, J. Eusebio, S. Chakraborty, and S. Panchanathan, “Deep hashing network for unsupervised domain adaptation,” in2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017
2017
-
[67]
Adapting visual category models to new domains,
K. Saenko, B. Kulis, M. Fritz, and T. Darrell, “Adapting visual category models to new domains,” inComputer Vision - ECCV 2010, 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV, ser. Lecture Notes in Computer Sci...
2010
-
[68]
Visda: The visual domain adaptation challenge,
X. Peng, B. Usman, N. Kaushik, J. Hoffman, D. Wang, and K. Saenko, “Visda: The visual domain adaptation challenge,”arXiv preprint arXiv:1710.06924, 2017
2017 arXiv
-
[69]
A database for handwritten text recognition research,
J. J. Hull, “A database for handwritten text recognition research,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 16, no. 5, pp. 550–554, 2002
2002
-
[70]
Reading digits in natural images with unsupervised feature learning,
Y . Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, and A. Ng, “Reading digits in natural images with unsupervised feature learning,”NIPS, 01 2011
2011
-
[71]
Gradient-based learning applied to document recognition,
Y . Lecun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,”Proceedings of the IEEE, vol. 86, pp. 2278 – 2324, 12 1998
1998
-
[72]
Moment matching for multi-source domain adaptation,
X. Peng, Q. Bai, X. Xia, Z. Huang, K. Saenko, and B. Wang, “Moment matching for multi-source domain adaptation,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 1406– 1415
2019
-
[73]
Deep residual learning for image recognition,
K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778
2016
-
[74]
Learning transferable features with deep adaptation networks,
M. Long, Y . Cao, J. Wang, and M. I. Jordan, “Learning transferable features with deep adaptation networks,” inProceedings of the 32nd In- ternational Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, ser. JMLR Workshop and Conference Proceedings, F. R....
2015
-
[75]
Deep transfer learning with joint adaptation networks,
M. Long, H. Zhu, J. Wang, and M. I. Jordan, “Deep transfer learning with joint adaptation networks,” inInternational conference on machine learning. PMLR, 2017, pp. 2208–2217
2017
-
[76]
Generate to adapt: Aligning domains using generative adversarial networks,
S. Sankaranarayanan, Y . Balaji, C. D. Castillo, and R. Chellappa, “Generate to adapt: Aligning domains using generative adversarial networks,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 8503–8512
2018
-
[77]
Adversarial discrimi- native domain adaptation,
E. Tzeng, J. Hoffman, K. Saenko, and T. Darrell, “Adversarial discrimi- native domain adaptation,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 7167–7176
2017
-
[78]
Unsupervised image-to-image transla- tion networks,
M. Liu, T. Breuel, and J. Kautz, “Unsupervised image-to-image transla- tion networks,” inAdvances in Neural Information Processing Systems SUBMITTED TO IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 16 30: Annual Conference on Neural Information Processing Sys...
2017
-
[79]
Cycada: Cycle-consistent adversarial domain adaptation,
J. Hoffman, E. Tzeng, T. Park, J. Zhu, P. Isola, K. Saenko, A. A. Efros, and T. Darrell, “Cycada: Cycle-consistent adversarial domain adaptation,” inProceedings of the 35th International Conference on Machine Learning, ICML 2018, Stockholmsmässan, Stockholm, Sweden, July 10-15...
2018
-
[80]
Conditional adversarial domain adaptation,
M. Long, Z. Cao, J. Wang, and M. I. Jordan, “Conditional adversarial domain adaptation,”arXiv preprint arXiv:1705.10667, 2017
2017 arXiv
-
[81]
Maximum classifier discrepancy for unsupervised domain adaptation,
K. Saito, K. Watanabe, Y . Ushiku, and T. Harada, “Maximum classifier discrepancy for unsupervised domain adaptation,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 3723–3732
2018
-
[82]
Bridging theory and algorithm for domain adaptation,
Y . Zhang, T. Liu, M. Long, and M. I. Jordan, “Bridging theory and algorithm for domain adaptation,” inProceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, ser. Proceedings of Machine Learning Research, K...
2019
-
[83]
Invariant and transportable representations for anti-causal domain shifts,
Y . Jiang and V . Veitch, “Invariant and transportable representations for anti-causal domain shifts,”Conference on Neural Information Processing Systems, 2022
2022
-
[84]
Transporting causal mechanisms for unsupervised domain adaptation,
Z. Yue, Q. Sun, X.-S. Hua, and H. Zhang, “Transporting causal mechanisms for unsupervised domain adaptation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 8599–8608
2021
-
[85]
Connect, not collapse: Explaining contrastive learning for unsupervised domain adaptation,
K. Shen, R. M. Jones, A. Kumar, S. M. Xie, J. Z. HaoChen, T. Ma, and P. Liang, “Connect, not collapse: Explaining contrastive learning for unsupervised domain adaptation,” inInternational Conference on Machine Learning. PMLR, 2022, pp. 19 847–19 878
2022
-
[86]
Identifiability conditions for domain adaptation,
I. Gulrajani and T. Hashimoto, “Identifiability conditions for domain adaptation,” inInternational Conference on Machine Learning. PMLR, 2022, pp. 7982–7997
2022
-
[87]
Partial disentanglement for domain adaptation,
L. Kong, S. Xie, W. Yao, Y . Zheng, G. Chen, P. Stojanov, V . Akinwande, and K. Zhang, “Partial disentanglement for domain adaptation,” in International Conference on Machine Learning. PMLR, 2022, pp. 11 455–11 472
2022
-
[88]
Dcan: Dual channel-wise alignment networks for unsupervised scene adaptation,
Z. Wu, X. Han, Y .-L. Lin, M. G. Uzunbas, T. Goldstein, S. N. Lim, and L. S. Davis, “Dcan: Dual channel-wise alignment networks for unsupervised scene adaptation,” inProceedings of the European Conference on Computer Vision (ECCV), 2018, pp. 518–534
2018
-
[89]
Unsupervised domain adaptation for semantic segmentation via class-balanced self-training,
Y . Zou, Z. Yu, B. Kumar, and J. Wang, “Unsupervised domain adaptation for semantic segmentation via class-balanced self-training,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 289–305
2018
-
[90]
Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation,
T.-H. Vu, H. Jain, M. Bucher, M. Cord, and P. Pérez, “Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 2517–2526
2019
-
[91]
Semantic-transferable weakly- supervised endoscopic lesions segmentation,
J. Dong, Y . Cong, G. Sun, and D. Hou, “Semantic-transferable weakly- supervised endoscopic lesions segmentation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 10 712–10 721
2019
-
[92]
Domain adaptation for semantic segmentation with maximum squares loss,
M. Chen, H. Xue, and D. Cai, “Domain adaptation for semantic segmentation with maximum squares loss,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 2090– 2099
2019
-
[93]
Constructing self-motivated pyramid curriculums for cross-domain semantic segmentation: A non- adversarial approach,
Q. Lian, F. Lv, L. Duan, and B. Gong, “Constructing self-motivated pyramid curriculums for cross-domain semantic segmentation: A non- adversarial approach,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 6758–6767
2019
-
[94]
Contextual-relation consistent domain adaptation for semantic segmentation,
J. Huang, S. Lu, D. Guan, and X. Zhang, “Contextual-relation consistent domain adaptation for semantic segmentation,” inEuropean conference on computer vision. Springer, 2020, pp. 705–722
2020
-
[95]
Cscl: Critical semantic- consistent learning for unsupervised domain adaptation,
J. Dong, Y . Cong, G. Sun, Y . Liu, and X. Xu, “Cscl: Critical semantic- consistent learning for unsupervised domain adaptation,” inComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VIII 16. Springer, 2020, pp. 745–762
2020
-
[96]
Learning from scale-invariant examples for domain adaptation in semantic segmentation,
M. N. Subhani and M. Ali, “Learning from scale-invariant examples for domain adaptation in semantic segmentation,” inComputer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXII 16. Springer, 2020, pp. 290–306
2020
-
[97]
Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation,
H. Wang, T. Shen, W. Zhang, L.-Y . Duan, and T. Mei, “Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation,” inEuropean conference on computer vision. Springer, 2020, pp. 642–659
2020
-
[98]
Unsupervised intra- domain adaptation for semantic segmentation through self-supervision,
F. Pan, I. Shin, F. Rameau, S. Lee, and I. S. Kweon, “Unsupervised intra- domain adaptation for semantic segmentation through self-supervision,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 3764–3773
2020
-
[99]
Cross-domain semantic segmenta- tion via domain-invariant interactive relation transfer,
F. Lv, T. Liang, X. Chen, and G. Lin, “Cross-domain semantic segmenta- tion via domain-invariant interactive relation transfer,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 4334–4343
2020
-
[100]
What can be transferred: Unsupervised domain adaptation for endoscopic lesions segmentation,
J. Dong, Y . Cong, G. Sun, B. Zhong, and X. Xu, “What can be transferred: Unsupervised domain adaptation for endoscopic lesions segmentation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 4023–4032
2020
-
[101]
Effective comparative prototype hashing for unsupervised domain adaptation,
H. Cui, L. Zhao, F. Li, L. Zhu, X. Han, and J. Li, “Effective comparative prototype hashing for unsupervised domain adaptation,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 8, 2024, pp. 8329–8337
2024
-
[102]
Semantic-aware message broadcast- ing for efficient unsupervised domain adaptation,
X. Li, C. Lan, G. Wei, and Z. Chen, “Semantic-aware message broadcast- ing for efficient unsupervised domain adaptation,”IEEE Transactions on Image Processing, 2024
2024
-
[103]
The use of ranks to avoid the assumption of normality implicit in the analysis of variance,
M. Friedman, “The use of ranks to avoid the assumption of normality implicit in the analysis of variance,”Journal of the american statistical association, vol. 32, no. 200, pp. 675–701, 1937
1937
-
[104]
The pascal visual object classes (voc) challenge,
M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman, “The pascal visual object classes (voc) challenge,”International journal of computer vision, vol. 88, no. 2, pp. 303–338, 2010
2010
-
[105]
Cross-domain weakly- supervised object detection through progressive domain adaptation,
N. Inoue, R. Furuta, T. Yamasaki, and K. Aizawa, “Cross-domain weakly- supervised object detection through progressive domain adaptation,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 5001–5009
2018
-
[106]
Playing for data: Ground truth from computer games,
S. R. Richter, V . Vineet, S. Roth, and V . Koltun, “Playing for data: Ground truth from computer games,” inEuropean conference on computer vision. Springer, 2016, pp. 102–118
2016
-
[107]
The cityscapes dataset for semantic urban scene understanding,
M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele, “The cityscapes dataset for semantic urban scene understanding,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 3213–3223
2016
-
[108]
The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes,
G. Ros, L. Sellart, J. Materzynska, D. Vazquez, and A. M. Lopez, “The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 3234–3243
2016
-
[109]
Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation,
J. Dong, Y . Cong, G. Sun, Z. Fang, and Z. Ding, “Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 3, pp. 1664–1681, 2021
2021
-
[110]
Category contrast for unsupervised domain adaptation in visual tasks,
J. Huang, D. Guan, A. Xiao, S. Lu, and L. Shao, “Category contrast for unsupervised domain adaptation in visual tasks,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1203–1214
2022
-
[111]
Multi-granularity alignment domain adaptation for object detection,
W. Zhou, D. Du, L. Zhang, T. Luo, and Y . Wu, “Multi-granularity alignment domain adaptation for object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 9581–9590
2022
-
[112]
Kl guided domain adaptation,
A. T. Nguyen, T. Tran, Y . Gal, P. H. Torr, and A. G. Baydin, “Kl guided domain adaptation,”arXiv preprint arXiv:2106.07780, 2022
2022 arXiv
-
[113]
Graph-relational domain adaptation,
Z. Xu, G.-H. Lee, Y . Wang, H. Wanget al., “Graph-relational domain adaptation,”arXiv preprint arXiv:2202.03628, 2022
2022 arXiv
-
[114]
Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs,
L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, “Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs,”IEEE transactions on pattern analysis and machine intelligence, vol. 40, no. 4, pp. 834–848, 2...
2017
-
[116]
For any choice ofϕand anyK≥1, ICNCE(Xs;X t ∣Z t, ϕ, K)≤I(X s;X t ∣Z t).(39)
-
[117]
There exists a function ϕ∗ that attains the supremum. Specifically, if ϕ∗(Xs, Xt, Zt)=log P(X t ∣X s, Zt) P(X t ∣Z t) +c(X s, Zt),(40) wherec(X s, Zt)is a function independent ofX t, then we have: supϕ ICNCE(Xs;X t ∣Z t, ϕ, K) =I CNCE(Xs;X t ∣Z t, ϕ∗, K) =I(X s;X t ∣Z t). (41)
-
[118]
In particular, we have: lim K→∞ sup ϕ ICNCE(Xs;X t ∣Z t, ϕ, K)=I(Xs;X t ∣Z t).(42) Proof
As the number of negative samples K grows, the approxima- tion tightens. In particular, we have: lim K→∞ sup ϕ ICNCE(Xs;X t ∣Z t, ϕ, K)=I(Xs;X t ∣Z t).(42) Proof. Step 1: Start from the definition of conditional mutual information. Recall that: I(X s;X t ∣Z t)= EP(Z t)[EP(X s,...
-
[119]
Bound withI(X s;Z s ∣X t): ¯Pe ≤1−exp [−H(Y)+I(X s;Z s ∣X t)].(58)
-
[120]
Bound withI(X t;Z t ∣X s): ¯Pe ≤1−exp [−H(Y)+I(X t;Z t ∣X s)].(59)
-
[121]
Bound withI(X s;X t ∣Z s): ¯Pe ≤1−exp [−H(Y)+I(X s;X t ∣Z s)].(60)
-
[122]
Step 1: Relating Error Probability to Entropy and Mutual Information We begin with a fundamental link between error probability and the entropy of the label Y
Bound withI(X s;X t ∣Z t)+∆(ϕ, ψ): ¯Pe ≤1−exp [−H(Y)+I(X s;X t ∣Z t)+∆(ϕ, ψ)].(61) Proof. Step 1: Relating Error Probability to Entropy and Mutual Information We begin with a fundamental link between error probability and the entropy of the label Y . Consider a classifier or d...
-
[123]
Consider that to perfectly predict Y , one must have sufficient information about Xs andX t, or their representations
Start from a scenario where the learner attempts to predict Y using information from the conditioned variables. Consider that to perfectly predict Y , one must have sufficient information about Xs andX t, or their representations
-
[124]
One can rearrange a Fano-type inequality or a Chernoff bound on the probability of error to link it to conditional mutual information
If the representation Zs given Xt does not reduce uncertainty much, this implies that the classifier’s error probability cannot be arbitrarily small. One can rearrange a Fano-type inequality or a Chernoff bound on the probability of error to link it to conditional mutual information
-
[125]
Using standard inequalities (e.g., from the proof techniques in related information-theoretic generalization bounds), we get a form: ¯Pe ≥exp[−H(Y)+I(X s;Z s∣Xt)],(67) or similarly for the other terms. Inverting this inequality gives the upper bound: ¯Pe ≤1−exp[−H(Y)+I(X s;Z s...
-
[6314]
Springer, 2010, pp. 213–226
2010
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.