Pith. sign in

REVIEW 4 major objections 4 minor 55 references

Deep Adversarial Transition Learning using Cross-Grafted Generative Stacks

T0 review · 4 major / 4 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read Cross-grafted decoder stacks beat domain adaptation benchmarks.

desk verdict Cross-grafted decoder stacks are a genuine variation on UNIT, but the paper lacks evidence that adversarial target transitions preserve labels, and the experimental reporting is too thin to back the SOTA claim. read the letter →

arxiv 2009.12028 v1 pith:23HGEPQ7 submitted 2020-09-25 cs.CV cs.LG

Jinyong Hou , Xuejie Ding , Stephen Cranefield , Jeremiah D. Deng This is my paper · ORCID
classification cs.CVcs.LG
keywords unsuperviseddomainadaptationadversariallearningvariationalautoencoderscross-graftedgenerativestackstransitionspacesimage-to-imagetranslationtransferpixel-level
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

Unsupervised domain adaptation normally tries to erase the difference between source and target features. This paper claims that a better route is to build intermediate 'transition' images by swapping halves of two variational autoencoder decoders, one trained on the source and one on the target, and then to align those transitions with adversarial networks. The classifier is trained on transitions produced from labeled source images and tested on transitions produced from unlabeled target images. The paper reports that this cross-grafted architecture outperforms existing methods on digit, fashion, and object-recognition adaptation tasks, and that the setup stays accurate no matter which direction the adaptation runs.

What carries the argument

The cross-grafted generative stack (CGGS) is the central mechanism: each domain's VAE decoder is cut into a high-level stack and a low-level stack, then reassembled into two mixed decoders, $D_{st} \equiv [D_s^h \circ D_t^l]$ and $D_{ts} \equiv [D_t^h \circ D_s^l]$. Feeding a latent code from either domain through these mixed decoders yields transition images $X^{st}_s, X^{st}_t, X^{ts}_s, X^{ts}_t$; the paper's probabilistic analysis shows that such a transition is the original reconstruction with noise added to the weights of one stack, so it sits between the two domains. Two adversarial generator-discriminator pairs ($G_1,D_1$) and ($G_2,D_2$) align the target-side transitions to the source-side ones, an MSE content loss keeps the paired transitions close, and the task classifier is trained on source transitions while tested on the aligned target transitions.

What would settle it

Keep every component of DATL identical but replace the shared high-level encoder layers with two separate encoders trained independently on source and target. If the shared-latent-space assumption is load-bearing, the MNIST→MNIST-M accuracy should fall well below the reported 0.983; if the accuracy holds, then the cross-grafting and adversarial alignment alone, not the shared latent space, carry the transfer.

Watch

Extended reading notes

Core claim

The paper's central claim is that the unsupervised domain adaptation task can be reframed as building a mapping from cross-domain 'transition' images onto the source label space. Transitions are generated by cross-grafted generative stacks (CGGS): the decoder of each domain's variational autoencoder is split into a high-level and a low-level stack, and two mixed decoders are formed by grafting the source's high stack onto the target's low stack and vice versa. A latent code from either domain, passed through such a mixed decoder, produces a transition image that the paper shows to be a probabilistic perturbed version of the same code's within-domain reconstruction. Two GANs then align target-initiated transitions to source-initiated transitions, and the task classifier is trained on source transitions and evaluated on the aligned target transitions. The paper argues from its benchmarks that this is enough to beat state-of-the-art adaptation accuracy, to make performance nearly symmetric across adaptation directions, and to let pretrained components transfer to new tasks.

Load-bearing premise

Source and target images must share a common latent content structure, so that swapping decoder halves still yields transitions that keep the label-relevant content intact.

Editorial extensions

If this is right

  • Adaptation can be done without cycle-consistency losses, simplifying pixel-level domain adaptation training.
  • Performance becomes nearly symmetric across adaptation directions, unlike several competing methods.
  • The learned CGGS and alignment components transfer to new adaptation tasks after only fine-tuning the adversarial alignment, so pretraining for one task reduces retraining cost for another.
  • The exact split of high-level vs low-level layers in CGGS offers a tunable knob: more high-level layers help when domains share content but differ in background, while more low-level layers help when backgrounds match but content differs.
  • Given a small number of labeled target samples, the same framework improves further, giving a semi-supervised extension.

Reading between the lines

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

  • The transition-space idea is not tied to VAEs: any paired generative decoders with a shared latent space could be grafted the same way, so diffusion-based decoders may extend the approach to higher-resolution image adaptation.
  • The pixel-wise MSE content loss in Eq. (13) is the most fragile part of the design; for adaptation tasks with large geometric or viewpoint differences, enforcing pixel-level equality between paired transitions could fight against the adversarial alignment, so a perceptual or contrastive content loss might be a safer regularizer.
  • The reported cross-task transfer suggests that the transition spaces encode something domain-generic; if that holds, the same CGGS could be reused as a fixed front-end for many adaptation tasks, converting domain adaptation from per-task training to fine-tuning.
  • Because the two channels $X^{st}$ and $X^{ts}$ use opposite graftings, ensembling their classifiers—rather than picking one channel—could further reduce the already small directional gap.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes Deep Adversarial Transition Learning (DATL), an unsupervised domain adaptation framework built from two variational autoencoders whose decoder stacks are cross-grafted to generate intermediate 'transition' images, followed by GAN-based alignment of the two transition spaces and a task classifier trained on source transitions. The method is evaluated on digit tasks (MNIST↔MNIST-M, MNIST↔USPS, MNIST↔M-Digits), Fashion↔Fashion-M, CIFAR10↔STL10, and synthetic-to-real LineMod, with ablations, sensitivity analysis of the graft split, t-SNE visualizations, a cross-task generalization study, and a semi-supervised extension. The central claim is that DATL outperforms state-of-the-art unsupervised domain adaptation methods on several benchmarks while being symmetric across adaptation directions and transferable across tasks.

Significance. If the empirical claims were fully supported, DATL would be a useful contribution: the idea of constructing bidirectional pixel-level transition spaces by cross-grafting decoder stacks is a distinctive architectural variant of UNIT-style shared-latent methods, and the paper provides a reasonably broad benchmark suite, ablations separating content similarity from GAN alignment, and an exploratory cross-task generalization study. The visualizations and L2 distance analyses are helpful for understanding the mechanism. However, the paper's value is primarily empirical, and the current evaluation has several load-bearing gaps: no error bars, per-task selection of the graft split without a clear holdout protocol, a semantic-consistency gap in the test-time generator objective, and inconsistent results between the two transition channels. These issues must be addressed before the stated superiority claim can be accepted.

major comments (4)
  1. [Section IV-C, Tables II–IV] No error bars, multiple runs, or significance tests are reported anywhere in the paper. For GAN-based methods, single-run accuracy is highly sensitive to initialization and training stochasticity; this is especially problematic when differences are small, e.g., CIFAR10→STL10 (DATL(Xst)=0.613 vs. DRCN=0.588) and LineMod (DATL=0.998 vs. PixelDA=0.998). The claim that DATL 'outperforms the state-of-the-art' requires mean and standard deviation over several seeds, or at least a clear statement of the number of runs and the variance.
  2. [Section IV-C4 and Table II] The CGGS split (e.g., H5L1 vs. H2L4) is tuned per task based on Figure 8, which plots final test accuracy for different splits. If the numbers in Table II use the best split per task, the evaluation is optimistic and the comparison with baselines is unfair, since the baselines are not given the same per-task test-time selection. The paper must state explicitly which CGGS configuration was used for each Table II entry and how it was selected (e.g., a validation split held out from the test set), or report results for all configurations without cherry-picking.
  3. [Section III-D, Eqs. (10)–(15)] The test-time generators G1 and G2 are trained only by adversarial losses. The content loss Lc compares Xst_s with Xst_t, not G1(Xst_t) with Xst_t, and the classifier loss LT is evaluated only on source transitions. Consequently, nothing in the objective prevents G1/G2 from changing the semantic label of a target transition while matching the source transition marginal distribution; the reported accuracy on generated transitions may reflect distribution matching rather than true label transfer. This is a load-bearing gap in the causal chain from transition alignment to target classification. The paper should add a semantic-consistency objective on the generated transitions (e.g., a content/cycle loss on G1(Xst_t) or a pseudo-label classification loss) or give a concrete argument for why marginal alignment alone is sufficient in this architecture.
  4. [Section IV-C, Tables II and IV] The two transition channels give inconsistent results, which undermines the symmetry claim. On MNIST→MNIST-M, DATL(Xst)=0.890 is below CyCADA (0.921), GtA (0.917), PixelDA (0.982), and UNIT (0.920), while DATL(Xts)=0.983 is much higher. If the method is presented as a two-channel framework, the paper needs a principled aggregation rule or an explicit channel-selection protocol; otherwise reporting both and letting the reader pick the better one is not a fair comparison. Moreover, the abstract's 'outperforms the state-of-the-art' claim is stronger than the evidence: on CIFAR10→STL10 the best DATL accuracy is 0.615, only 2.7 points above DRCN and far below the target-only upper bound of 0.791, and on LineMod DATL ties PixelDA at 0.998 rather than exceeding it.
minor comments (4)
  1. [Section III-C, Eqs. (1)–(5)] The perturbation derivation is circular: epsilon_tl and epsilon_th are defined as the differences between the target and source decoder weights, so the conclusion that a cross-grafted transition is a 'perturbed version' of the reconstruction is true by construction. This should be reframed as a formal identity, not as an explanation of why cross-grafting helps.
  2. [Section IV-C1 and Tables III–IV] The text says the CIFAR10↔STL10 performance is given in Table III, but Table III is the LineMod task and Table IV contains the CIFAR10/STL10 results. The table numbering and cross-references need to be fixed.
  3. [Section III-D, Eq. (13)] The sentence 'For Xts, Lts_s is similarly defined' should presumably refer to Lts_c, the content loss for the Xts channel, not Lts_s. This notational slip makes the loss definition hard to follow.
  4. [Throughout] There are numerous typographical errors that should be cleaned up, including 'perfomance', 'caculated', 'traning', 'NeuIPS' (for NeurIPS), and 'As as result'. The paper also does not mention code release, which would help reproducibility.

Circularity Check

1 steps flagged · score 2.0 of 10

No meaningful circularity; only a minor definitional tautology in the perturbation framing, while the benchmark claim is independent.

  1. self definitional [Section III-C, Eqs. (3)-(5)]
    "with Dt_l being interpreted as its counterpart Ds_l added with perturbations ϵϵϵtl to all the lower layers: ... Therefore transitionPst_s can be seen as a perturbed version of the reconstruction Pss_s."

    The perturbation is introduced by definition: Eq. (3) sets the target lower-stack weights equal to the source lower-stack weights plus ε, and Eq. (4) does the same for the high stack. The subsequent statement that the cross-grafted transition Pst_s (or Pts_s) is a perturbed version of the self-reconstruction Pss_s is therefore true by construction rather than derived from an independent argument. This is the paper's advertised 'theoretical explanation' of transition generation, but it is only a re-labeling of the weight difference; it is not used to predict any benchmark accuracy or to constrain the learned model beyond the already-defined network architecture. Hence it is a self-definitional framing, not a load-bearing circular derivation.

full rationale

This is an empirical architecture paper whose central claim, 'our method outperforms the state-of-the-art on a number of unsupervised domain adaptation benchmarks,' is supported by external benchmark comparisons in Tables II-IV, ablations in Table V, and visualization/L2 analyses in Figures 6-7 and Table VII. The learning objective is fully specified in Eqs. (7)-(15), and no reported number is a fitted parameter renamed as a prediction. The only self-referential element is the Section III-C perturbation reading of cross-grafting: since ε is defined as the weight difference between target and source decoder stacks, calling the resulting transition a perturbed reconstruction is immediate from the definition. This does not enter the loss functions or the benchmark evaluation in a way that would force the empirical outcome, so it does not make the paper circular. The self-citation [15] merely records that this paper extends a conference publication and is not load-bearing. The skeptic's point that adversarial generators G1/G2 are not explicitly trained with a label-preservation objective is a correctness or robustness concern about semantic drift, not a circularity of the derivation.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The method relies on standard UDA assumptions plus a shared latent space assumption inherited from UNIT. The main hand-tuned components are the loss weights and the CGGS layer split, the latter tuned per task in sensitivity analysis. No new physical entities are introduced; the transition space is a model construct with direct visual and distance-based evidence.

free parameters (2)
  • Loss weights lambda_0, lambda_1, lambda_2, lambda_3 = 1, 10, 0.01, 1
    Hand-chosen hyperparameters balancing VAE reconstruction, prior KL, adversarial loss, content similarity, and classifier loss. No tuning procedure or sensitivity analysis is reported for these values.
  • CGGS graft split (high-level vs low-level layers) = H5L1 for some tasks, H2L4 for others, H4L2 for cross-task experiments
    The split is selected per task based on the sensitivity analysis in Figure 8, with no holdout protocol described. This is a form of parameter choice that can overfit the test benchmark.
assumptions (4)
  • domain assumption Source and target domains share a latent space and label space
    Section III-A states 'Inspired by UNIT, we also assume a shared latent space for the joint latent representations.' Section I assumes the same label space. If false, cross-grafted transitions lack common content and the source classifier cannot transfer.
  • domain assumption VAE latent prior is Gaussian N(0, I)
    Section III-B assumes a normal prior for latent encodings, following standard VAE practice. This is a modeling choice rather than a derived fact.
  • standard math The coupling theory guarantees existence of a joint distribution with given marginals
    Section III-A cites coupling theory [43] to motivate constructing a joint distribution. This is a standard existence result, but the paper does not use it to construct a specific joint distribution.
  • standard math GAN minimax optimization minimizes Jensen-Shannon divergence
    Section III-C invokes the standard GAN result from [14] to claim that the adversarial game aligns transition distributions. This is a standard result under ideal discriminator capacity and convergence.
invented entities (1)
  • Transition spaces (cross-grafted transitions Xst and Xts) independent evidence
    purpose: Intermediate pixel-level representations bridging source and target domains for training a classifier without target labels
    The transition images are directly visualized in Figures 5-7 and measured via L2 distances in Table VII, providing empirical handles outside the paper's own accuracy claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Adversarial Transition Learning using Cross-Grafted Generative Stacks." pith.science (2026). https://pith.science/paper/23HGEPQ7

@misc{pith2026200912028,
  author       = {Pith},
  title        = {Pith review of: Deep Adversarial Transition Learning using Cross-Grafted Generative Stacks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/23HGEPQ7}},
  note         = {Machine review of arXiv:2009.12028}
}
read the original abstract

Current deep domain adaptation methods used in computer vision have mainly focused on learning discriminative and domain-invariant features across different domains. In this paper, we present a novel "deep adversarial transition learning" (DATL) framework that bridges the domain gap by projecting the source and target domains into intermediate, transitional spaces through the employment of adjustable, cross-grafted generative network stacks and effective adversarial learning between transitions. Specifically, we construct variational auto-encoders (VAE) for the two domains, and form bidirectional transitions by cross-grafting the VAEs' decoder stacks. Furthermore, generative adversarial networks (GAN) are employed for domain adaptation, mapping the target domain data to the known label space of the source domain. The overall adaptation process hence consists of three phases: feature representation learning by VAEs, transitions generation, and transitions alignment by GANs. Experimental results demonstrate that our method outperforms the state-of-the art on a number of unsupervised domain adaptation benchmarks.

Figures

Figures reproduced from arXiv: 2009.12028 by the authors.

Figure 2
Figure 2. Overview of the the proposed model. Module A: high-level layers of encoders E s h, E t h are shared (indicated by the dashed line). Decoder layers D s h and D t h output high-level representations of the source and target images respectively, whereas D s l , D t l give low-level representations; Module B: Xst s , Xts t , Xts s , Xts t are the transitions reproduced by CGGS (D st ≡ [D s h ◦ D t l ] and D ts ≡ [D t h … view at source ↗
Figure 3
Figure 3. Example images from the datasets used in our experiments. Algorithm 1: Training of the DATL framework Input: Source: Xs,ys, Target:Xt Result: Model Weights of DATL 1 θE, θD, θG, θDis, θT ← initialization 2 for iterations of traning do 3 z ← sample from N (0, I) 4 Xs, Xt ← sample mini-batch 5 zs, zt ← sample from q(zs|Xs) and q(zt|Xt) 6 Generate Xst s , Xst t , Xts s , Xts t by feeding zs, zt through CGGS 7 Generate … view at source ↗
Figure 4
Figure 4. Absolute accuracy difference between different direc [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Visualization of transition generations. For each scenario, the leftmost column is the source and its transition, and the rightmost is for the target. During the experiments, the transitions of source are the ‘real player’. The adversarial generations for target transi…
Figure 6
Figure 6. Figure 6: Visualization of domains and transitions using t-SNE for MNIST→MNIST-M. (a) Xs (blue) vs. Xt (red) (b) Xst s (blue) vs. Xst t (red) (c) Xst s (blue) vs. Xe st t (red) [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Visualization of domains and transitions using t-SNE for MNIST→USPS [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Adaptation performance of different CGGS settings [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 55 canonical work pages

  1. [1]

    A Survey on Transfer Learning,

    S. J. Pan and Q. Yang, “A Survey on Transfer Learning,” IEEE Transactions on Knowledge and Data Engineering (TKDE) , vol. 22, no. 10, pp. 1345–1359, 2010

    work page 2010

  2. [2]

    Cross- Modal Adaptation for RGB-D Detection Eractive Adaptation of Real- Time Object Detectors,

    J. Hoffman, T. Darrell, S. Gupta, J. Leong, and S. Guadarrama, “Cross- Modal Adaptation for RGB-D Detection Eractive Adaptation of Real- Time Object Detectors,” in IEEE Conference on Robotics and Automation (ICRA). IEEE, 2016, pp. 5032–5039

    work page 2016

  3. [3]

    Adapting Deep Visuomotor Representations with Weak Pairwise Constraints,

    E. Tzeng, C. Devin, J. Hoffman, C. Finn, P. Abbeel, S. Levine, K. Saenko, and T. Darrell, “Adapting Deep Visuomotor Representations with Weak Pairwise Constraints,” in Algorithmic Foundations of Robotics XII , 2020, pp. 688–703

    work page 2020

  4. [4]

    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. 1-2, pp. 151–175, 2010

    work page 2010

  5. [5]

    A Probabilistic Covariate Shift Assumption for Domain Adaptation,

    T. Adel and A. Wong, “A Probabilistic Covariate Shift Assumption for Domain Adaptation,” in AAAI Conference on Artificial Intelligence (AAAI), 2015, pp. 2476–2482

    work page 2015

  6. [6]

    Transfer joint matching for unsupervised domain adaptation,

    M. Long, J. Wang, G. Ding, J. Sun, and P. S. Yu, “Transfer joint matching for unsupervised domain adaptation,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2014, pp. 1410–1417

    work page 2014

  7. [7]

    Domain Adaptation for Object Recognition: An Unsupervised Approach,

    R. Gopalan, R. Li, and R. Chellappa, “Domain Adaptation for Object Recognition: An Unsupervised Approach,” in IEEE International Confer- ence on Computer Vision (ICCV) , D. N. Metaxas, L. Quan, A. Sanfeliu, and L. J. V . Gool, Eds., 2011, pp. 999–1006

    work page 2011

  8. [8]

    How transferable are features in deep neural networks?

    J. Yosinski, J. Clune, Y . Bengio, and H. Lipson, “How transferable are features in deep neural networks?” in Advances in Neural Information Processing Systems (NeurIPS) , 2014, pp. 3320–3328

    work page 2014

Show all 55 references
  1. [9]

    Learning Transferrable Representations for Unsupervised Domain Adaptation,

    O. Sener, H. O. Song, A. Saxena, and S. Savarese, “Learning Transferrable Representations for Unsupervised Domain Adaptation,” in Advances in Neural Information Processing Systems (NeurIPS) , 2016, pp. 2110–2118

    2016

  2. [10]

    Learning an Invariant Hilbert Space for Domain Adaptation

    S. Herath, M. T. Harandi, and F. Porikli, “Learning an Invariant Hilbert Space for Domain Adaptation.” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2017, pp. 3956–3965

    2017

  3. [11]

    Nonlinear principal component analysis using autoassociative neural networks,

    M. A. Kramer, “Nonlinear principal component analysis using autoassociative neural networks,” AIChE Journal , vol. 37, no. 2, pp. 233–243, 1991. [Online]. Available: https://aiche.onlinelibrary.wiley.com/ doi/abs/10.1002/aic.690370209

    1991 doi

  4. [12]

    Auto-Encoding Variational Bayes,

    D. P. Kingma and M. Welling, “Auto-Encoding Variational Bayes,” in International Conference on Learning Representations (ICLR) , 2014

    2014

  5. [13]

    Unsupervised Image-to-Image Translation Networks,

    M.-Y . Y . Liu, T. Breuel, and J. Kautz, “Unsupervised Image-to-Image Translation Networks,” in Advances in Neural Information Processing Systems (NeurIPS), 2017, pp. 700–708

    2017

  6. [14]

    Generative Adversarial Nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative Adversarial Nets,” in Advances in neural information processing systems (NeurIPS) , 2014, pp. 2672–2680

    2014

  7. [15]

    Unsupervised Domain Adaptation using Deep Networks with Cross-Grafted Stacks,

    J. Hou, X. DIng, J. D. Deng, and S. Cranefield, “Unsupervised Domain Adaptation using Deep Networks with Cross-Grafted Stacks,” in International Conference on Computer Vision Workshop (ICCVW) , 2019, pp. 3257–3264. 12

    2019

  8. [16]

    Domain Adaptation via Transfer Component Analysis,

    S. J. Pan, I. W. Tsang, J. T. Kwok, and Q. Yang, “Domain Adaptation via Transfer Component Analysis,” IEEE Transactions on Neural Networks (TNN), vol. 22, no. 2, pp. 199–210, 2011

    2011

  9. [17]

    Deep CORAL: Correlation Alignment for Deep Domain Adaptation,

    B. Sun and K. Saenko, “Deep CORAL: Correlation Alignment for Deep Domain Adaptation,” in European Conference on Computer Vision (ECCV), 2016, pp. 443–450

    2016

  10. [18]

    Self-Taught Learning: Transfer Learning from Unlabeled Data,

    R. Raina, A. Battle, H. Lee, B. Packer, and A. Y . Ng, “Self-Taught Learning: Transfer Learning from Unlabeled Data,” in International Conference on Machine Learning (ICML) , 2007, pp. 759–766

    2007

  11. [19]

    Geodesic Flow Kernel for Unsupervised Domain Adaptation,

    B. Gong, Y . Shi, F. Sha, and K. Grauman, “Geodesic Flow Kernel for Unsupervised Domain Adaptation,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2012, pp. 2066–2073

    2012

  12. [20]

    Unsupervised Adaptation Across Domain Shifts by Generating Intermediate Data Representations,

    R. Gopalan, R. Li, and R. Chellappa, “Unsupervised Adaptation Across Domain Shifts by Generating Intermediate Data Representations,” IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) , vol. 36, no. 11, pp. 2288–2302, 2014

    2014

  13. [21]

    Manifold Criterion Guided Transfer Learning via Intermediate Domain Generation,

    L. Zhang, S. Wang, G. B. Huang, W. Zuo, J. Yang, and D. Zhang, “Manifold Criterion Guided Transfer Learning via Intermediate Domain Generation,” IEEE Transactions on Neural Networks and Learning Systems (TNNLS), vol. 30, no. 12, pp. 3759–3773, 2019

    2019

  14. [22]

    Deep Hashing Network for Unsupervised Domain Adaptation,

    H. Venkateswara, J. Eusebio, S. Chakraborty, and S. Panchanathan, “Deep Hashing Network for Unsupervised Domain Adaptation,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2017, pp. 5018–5027

    2017

  15. [23]

    Learning Transferable Features with Deep Adaptation Networks,

    M. Long, Y . Cao, J. Wang, and M. I. Jordan, “Learning Transferable Features with Deep Adaptation Networks,” in International Conference on Machine Learning (ICML) , ser. PMLR, F. R. Bach and D. M. Blei, Eds., vol. 37. JMLR.org, 2015, pp. 97–105

    2015

  16. [24]

    Simultaneous Deep Transfer Across Domains and Tasks,

    E. Tzeng, J. Hoffman, T. Darrell, and K. Saenko, “Simultaneous Deep Transfer Across Domains and Tasks,” in IEEE International Conference on Computer Vision (ICCV) , 2015, pp. 4068–4076

    2015

  17. [25]

    Unsupervised domain adaptation with residual transfer networks,

    M. Long, H. Zhu, J. Wang, and M. I. Jordan, “Unsupervised domain adaptation with residual transfer networks,” in Advances in Neural Information Processing Systems (NeuIPS) , 2016, pp. 136–144

    2016

  18. [26]

    Beyond Sharing Weights for Deep Domain Adaptation,

    A. Rozantsev, M. Salzmann, and P. Fua, “Beyond Sharing Weights for Deep Domain Adaptation,” IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), vol. 41, no. 4, pp. 801–814, 2018

    2018

  19. [27]

    Domain Separation Networks,

    K. Bousmalis, G. Trigeorgis, N. Silberman, D. Krishnan, and D. Erhan, “Domain Separation Networks,” in Advances in Neural Information Processing Systems (NeurIPS) , 2016, pp. 343–351

    2016

  20. [28]

    Deep Reconstruction-Classification Networks for Unsupervised Domain Adap- tation,

    M. Ghifary, W. B. Kleijn, M. Zhang, D. Balduzzi, and W. Li, “Deep Reconstruction-Classification Networks for Unsupervised Domain Adap- tation,” in European Conference on Computer Vision (ECCV) , B. Leibe, J. Matas, N. Sebe, and M. Welling, Eds., 2016, pp. 597–613

    2016

  21. [29]

    DLID: Deep Learning for Domain Adaptation by Inter- polating between Domains,

    S. Chopra, “DLID: Deep Learning for Domain Adaptation by Inter- polating between Domains,” in International Conference on Machine Learning Workshop on Representation (ICMLW) , 2013

    2013

  22. [30]

    Domain Bridge for Unpaired Image-to-Image Translation and Unsupervised Domain Adaptation,

    F. Pizzati, R. de Charette, M. Zaccaria, and P. Cerri, “Domain Bridge for Unpaired Image-to-Image Translation and Unsupervised Domain Adaptation,” in IEEE Winter Conference on Applications of Computer Vision (WACV), 2020, pp. 2990–2998

    2020

  23. [31]

    Adversarial Discrimi- native Domain Adaptation,

    E. Tzeng, J. Hoffman, K. Saenko, and T. Darrell, “Adversarial Discrimi- native Domain Adaptation,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2017, pp. 7167–7176

    2017

  24. [32]

    Deep Do- main Confusion: Maximizing for Domain Invariance,

    E. Tzeng, J. Hoffman, N. Zhang, K. Saenko, and T. Darrell, “Deep Do- main Confusion: Maximizing for Domain Invariance,” arXiv:1412.3474 [cs.CV], 2014

    2014 arXiv

  25. [33]

    Domain-Adversarial Training of Neural Networks,

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. Marchand, and V . S. Lempitsky, “Domain-Adversarial Training of Neural Networks,” Journal of Machine Learning Research (JMLR) , vol. 17, no. 59, pp. 1–35, 2016

    2016

  26. [34]

    Unsupervised Pixel-Level Domain Adaptation with Generative Adver- sarial Networks,

    K. Bousmalis, N. Silberman, D. Dohan, D. Erhan, and D. Krishnan, “Unsupervised Pixel-Level Domain Adaptation with Generative Adver- sarial Networks,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 95–104

    2017

  27. [35]

    Coupled Generative Adversarial Networks,

    M.-Y . Y . Liu and O. Tuzel, “Coupled Generative Adversarial Networks,” in Advances in Neural Information Processing Systems (NeurIPS) , D. D. Lee, M. Sugiyama, U. von Luxburg, I. Guyon, and R. Garnett, Eds., 2016, pp. 469–477

    2016

  28. [36]

    Detach and Adapt: Learning Cross-Domain Disentangled Deep Representation,

    Y . C. Liu, Y . Y . Yeh, T. C. Fu, S. D. Wang, W. C. Chiu, and Y . C. F. Wang, “Detach and Adapt: Learning Cross-Domain Disentangled Deep Representation,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 8867–8876

    2018

  29. [37]

    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,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 8503–8512

    2018

  30. [38]

    Conditional Image Synthesis with Auxiliary Classifier GANs,

    A. Odena, C. Olah, and J. Shlens, “Conditional Image Synthesis with Auxiliary Classifier GANs,” in International Conference on Machine Learning, (ICML), ser. PMLR, D. Precup and Y . W. Teh, Eds., vol. 70, 2017, pp. 2642–2651

    2017

  31. [39]

    CyCADA: Cycle-Consistent Adversarial Domain adaptation,

    J. Hoffman, E. Tzeng, T. Park, J.-Y . Y . Zhu, P. Isola, K. Saenko, A. A. Efros, and T. Darrell, “CyCADA: Cycle-Consistent Adversarial Domain adaptation,” in International Conference on Machine Learning (ICML) , ser. PMLR, J. G. Dy and A. Krause, Eds., vol. 5, 2018, pp. 3162–3174

    2018

  32. [40]

    Unpaired Image-to- Image Translation using Cycle-Consistent Adversarial Networks,

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired Image-to- Image Translation using Cycle-Consistent Adversarial Networks,” in IEEE Conference on Computer Vision (ICCV) , 2017, pp. 2242–2251

    2017

  33. [41]

    Image to Image Translation for Domain Adaptation,

    Z. Murez, S. Kolouri, D. Kriegman, R. Ramamoorthi, and K. Kim, “Image to Image Translation for Domain Adaptation,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2018, pp. 4500–4509

    2018

  34. [42]

    Domain Adaptive Image-to-image Translation,

    Y .-c. Chen and X. Xu, “Domain Adaptive Image-to-image Translation,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2020, pp. 5274–5283

    2020

  35. [43]

    Lindvall, Lectures on the Coupling Method

    T. Lindvall, Lectures on the Coupling Method . Dover, 2002

    2002

  36. [44]

    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 (TPAMI) , vol. 35, no. 8, pp. 1798–1828, 2013

    2013

  37. [45]

    Learning to Discover Cross-Domain Relations with Generative Adversarial Networks,

    T. Kim, M. Cha, H. Kim, J. K. Lee, and J. Kim, “Learning to Discover Cross-Domain Relations with Generative Adversarial Networks,” in International Conference on Machine Learning (ICML) , ser. PMLR, vol. 70, 2017, pp. 1857–1865

    2017

  38. [46]

    Conditional Adversarial Domain Adaptation,

    M. Long, Z. Cao, J. Wang, and M. I. Jordan, “Conditional Adversarial Domain Adaptation,” in Advances in neural information processing systems (NeurIPS), S. Bengio, H. M. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, Eds., 2018, pp. 1640–1650

    2018

  39. [47]

    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, no. 11, pp. 2278–2324, 1998

    1998

  40. [48]

    Handwritten Digit Recognition: Applications of Neural Network Chips and Automatic Learning,

    Y . Le Cun, L. D. Jackel, B. Boser, J. S. Denker, H. P. Graf, I. Guyon, D. Henderson, R. E. Howard, and W. Hubbard, “Handwritten Digit Recognition: Applications of Neural Network Chips and Automatic Learning,” IEEE Communications Magazine , vol. 27, no. 11, pp. 41–46, 1989

    1989

  41. [49]

    Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms,

    H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms,” arXiv:1708.07747 [cs.LG], 2017

    2017 arXiv

  42. [50]

    Gradient Response Maps for Real-Time Detection of Texture- Less Objects,

    S. Hinterstoisser, C. Cagniart, S. Ilic, P. Sturm, N. Navab, P. Fua, and V . Lepetit, “Gradient Response Maps for Real-Time Detection of Texture- Less Objects,” IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), vol. 34, no. 5, pp. 876–888, 2012

    2012

  43. [51]

    Learning Descriptors for Object Recognition and 3D Pose Estimation,

    P. Wohlhart and V . Lepetit, “Learning Descriptors for Object Recognition and 3D Pose Estimation,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 3109–3118

    2015

  44. [52]

    TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems,

    M. Abadi and Others, “TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems,” arXiv:1603.04467 [cs.DC], 2016

    2016 arXiv

  45. [53]

    Adam: A Method for Stochastic Optimiza- tion,

    D. P. Kingma and J. L. Ba, “Adam: A Method for Stochastic Optimiza- tion,” in International Conference on Learning Representations (ICLR) , 2015

    2015

  46. [54]

    Deconvolutional Networks,

    M. D. Zeiler, D. Krishnan, G. W. Taylor, and R. Fergus, “Deconvolutional Networks,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2010, pp. 2528–2535

    2010

  47. [55]

    Visualizing data using t-SNE,

    L. van der Maaten and G. Hinton., “Visualizing data using t-SNE,” Journal of machine learning research (JMLR) , vol. 9, pp. 2579–2605, nov 2008

    2008

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.