pith. sign in

arxiv: 2605.17918 · v1 · pith:5M334RPCnew · submitted 2026-05-18 · 💻 cs.LG · cs.AI· cs.CV

Domain Transfer Becomes Identifiable via a Single Alignment

Sagar Shrestha , Subash Timilsina , Hoang-Son Nguyen , Xiao Fu This is my paper

Pith reviewed 2026-05-20 13:29 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CV
keywords domain transferidentifiabilityJacobian sparsitydistribution matchinganchor samplemeasure-preserving automorphismsimage-to-image translation
0
0 comments X

The pith

Under a fixed sparsity pattern in how inputs affect outputs, matching distributions plus one paired sample identifies the true domain transfer map.

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

Domain transfer seeks to move data from a source distribution to a target distribution without needing many matched pairs, but the problem is usually ill-posed because many different maps can push one distribution onto the other while preserving the marginals. The paper establishes that if the transfer map obeys a structural sparsity condition on the locations of its non-zero partial derivatives, then ordinary distribution matching combined with a single correctly aligned anchor pair is enough to recover the ground-truth map. This approach needs far less supervision than earlier methods that required transferring several matched conditional distributions at once. To make the method work in high dimensions, the authors introduce a randomized masked finite-difference regularizer that enforces the sparsity pattern without ever forming the full Jacobian matrix.

Core claim

Under a structural sparsity condition on the Jacobian support pattern, distribution matching together with a single paired anchor sample suffices to identify the ground-truth transfer.

What carries the argument

Structural sparsity condition on the Jacobian support pattern, which fixes the locations of non-zero partial derivatives and thereby rules out measure-preserving automorphisms once a single anchor pair is supplied.

If this is right

  • Distribution matching alone leaves the transfer non-unique, but the addition of one anchor pair removes the ambiguity when the sparsity pattern holds.
  • The required supervision is limited to a single paired sample rather than multiple matched conditional distributions.
  • The randomized masked finite-difference regularizer allows enforcement of the sparsity condition at scale without explicit Jacobian computation.
  • The theory applies to tasks such as unsupervised image-to-image translation and cross-platform medical imaging under the stated structural assumption.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • In applications where the transfer map is expected to act locally, such as pixel-wise image changes, the sparsity pattern could be checked or learned from data before training.
  • The same single-anchor logic might apply to other distribution-alignment problems in which a Jacobian support pattern can be assumed or estimated.
  • Practitioners could collect one reliable anchor pair as a cheap way to resolve ambiguity in otherwise unsupervised domain-transfer pipelines.

Load-bearing premise

The transfer map must have a fixed and known pattern of which input variables influence which output variables through its partial derivatives.

What would settle it

Observe whether, after enforcing the claimed sparsity pattern and using one anchor pair, any other distinct map still matches the source and target distributions and the anchor correspondence; if such a map exists, the identifiability claim fails.

Figures

Figures reproduced from arXiv: 2605.17918 by Hoang-Son Nguyen, Sagar Shrestha, Subash Timilsina, Xiao Fu.

Figure 1
Figure 1. Figure 1: Domain 1: MNIST digits, Domain 2: rotated MNIST digits. MPA issue causes content-misaligned translation as de￾picted by the change in digit identity after translation by existing methods. Matched Distribution [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of how MPA results in content misalignment. If px were distributed N (µ, σ2 ), m = −x+2µ would be an MPA. Hence g ⋆ ◦m and g ⋆ both are solutions to the distribution transport Problem (3). However, g ⋆ ◦ m results in content misalignment. 2.2. Prior Art: Diversified Distribution Matching To address the non-identifiability challenge brought by MPA, recent works (Shrestha & Fu, 2024; 2025) propo… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of Assumption 3.1 on the structure of g ⋆ . For each input xk (in red circle), one can find a subset of indices Ck (in black circles) corresponding to output coordinates that are influenced by xk, such that xk is the only common influence (illustrated by solid red lines) for these outputs. Assumption 3.1 follows the recent development in nonlin￾ear unmixing via sparse structure (Zheng et al., … view at source ↗
Figure 5
Figure 5. Figure 5: Result of translation from MNIST digits to rotated MNIST digits for all methods. layers: standard instance normalization couples all spatial locations within a channel and thus precludes a sparse Ja￾cobian. We therefore replace instance normalization with channel-only normalization in our image experiments. We did not observe any degradation from this replacement, and most other architectural choices remai… view at source ↗
Figure 6
Figure 6. Figure 6: Result of translation from Edges to rotated shoes for all methods [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: K-NN accuracy of RNA-seq to ATAC-seq translation on the A549 single-cell dataset (Cao et al., 2018), comparing the proposed method to CM-AE (Yang et al., 2021). The proposed method clearly benefits from anchor matching combined with sparsity regularization. labels or attributes) for all samples. We propose an alter￾native route based on structural sparsity: under a Jacobian support sparsity condition, a si… view at source ↗
Figure 8
Figure 8. Figure 8: Ablation study on MNIST → rotated MNIST. Case I: only distribution matching, Case II: sparsity regularization + distribution matching, Case III: anchor alignment + distribution matching, Case IV: anchor alignment + sparsity regularization + distribution matching [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Variance (a) and bias (b) of D S ∥Jz∥0 as a function of S. Note that S = 1 corresponds to the column-wise finite-difference based regularizer in (17). Source Target S=100 S=1 [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: MNIST to Rotated MNIST using finite-masked difference with S = 1 and S = 100. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
read the original abstract

Domain transfer (DT) maps source to target distributions and supports tasks such as unsupervised image-to-image translation, single-cell analysis, and cross-platform medical imaging. However, DT is fundamentally ill-posed: push-forward mappings are generally non-identifiable, as measure-preserving automorphisms (MPAs) preserve marginals while altering cross-domain correspondences, leading to content-misaligned translation. Recent work shows that MPAs can be eliminated by jointly transferring multiple corresponding source/target conditional distributions, but supervision signals labeling such conditionals are not always available in practice. We develop an alternative route to DT identifiability. Under a structural sparsity condition on the Jacobian support pattern, we show that distribution matching together with a single paired anchor sample suffices to identify the ground-truth transfer -- requiring substantially less supervision than prior approaches. To enable practical high-dimensional learning, we further propose an efficient Jacobian sparsity regularizer based on randomized masked finite differences, yielding a scalable surrogate without explicit Jacobian evaluation. Empirical results on synthetic and real-world DT tasks validate the theory.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript claims that under a structural sparsity condition on the support pattern of the Jacobian of the domain transfer mapping, distribution matching combined with a single paired anchor sample suffices to identify the ground-truth transfer mapping, eliminating measure-preserving automorphisms. To operationalize this in high dimensions, the authors introduce a randomized masked finite-difference regularizer as a scalable surrogate for enforcing the required Jacobian sparsity without explicit Jacobian evaluation. The theory is supported by a proof sketch and validated empirically on synthetic data and real-world tasks such as image-to-image translation.

Significance. If the central identifiability result holds and the surrogate regularizer provably recovers the exact support pattern required by the theorem, the work meaningfully reduces the supervision needed for identifiable domain transfer relative to prior approaches that rely on multiple corresponding conditional distributions. This could have practical impact in unsupervised translation, single-cell analysis, and medical imaging. The proposal of an efficient sparsity-inducing regularizer is a constructive contribution, though its theoretical linkage to the identifiability guarantee is the key point requiring verification.

major comments (2)
  1. [§3] §3 (Identifiability Theorem): The result states that distribution matching plus one anchor identifies the ground-truth mapping once the Jacobian support pattern is fixed and known. However, the manuscript does not explicitly address whether this pattern must be provided as prior knowledge or can be recovered from data; if the pattern is only approximately recovered, the uniqueness argument among MPAs may not go through. A concrete statement of the assumption (e.g., whether the support mask is an input or an output of the procedure) is needed.
  2. [§4] §4 (Surrogate Regularizer): The randomized masked finite-difference regularizer penalizes non-sparsity only in expectation over masks and finite-difference steps. The manuscript provides no proof that any minimizer of this surrogate objective necessarily satisfies the exact, deterministic support condition used in the §3 theorem. Without such a guarantee, the practical algorithm may admit solutions whose effective support differs from the assumed pattern, leaving residual MPAs that the single anchor cannot disambiguate. A counter-example or a lemma showing implication from surrogate to exact support would resolve this.
minor comments (2)
  1. [§4] Notation: The definition of the masking probability and finite-difference step size (listed as free parameters) should be moved to a dedicated paragraph or table so that readers can immediately see the hyper-parameters that are not part of the identifiability claim.
  2. [§5] Experiments: The synthetic validation would benefit from an explicit ablation that varies the mismatch between the assumed support pattern and the pattern recovered by the regularizer, quantifying how often the single-anchor selection still recovers the ground truth.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for their constructive comments on our work. We address the major comments point by point below and outline the revisions we plan to make.

read point-by-point responses

  1. Referee: [§3] §3 (Identifiability Theorem): The result states that distribution matching plus one anchor identifies the ground-truth mapping once the Jacobian support pattern is fixed and known. However, the manuscript does not explicitly address whether this pattern must be provided as prior knowledge or can be recovered from data; if the pattern is only approximately recovered, the uniqueness argument among MPAs may not go through. A concrete statement of the assumption (e.g., whether the support mask is an input or an output of the procedure) is needed.

    Authors: We agree that the manuscript would benefit from a clearer statement on this point. The identifiability result in Section 3 treats the Jacobian support pattern as a known structural assumption on the transfer mapping. It is provided as prior knowledge in the theorem statement, not recovered as an output of the procedure. The randomized masked finite-difference regularizer is introduced in Section 4 as a practical mechanism to encourage the learned mapping to satisfy a sparse support pattern consistent with this assumption. We will revise the text in Section 3 to explicitly clarify that the support pattern is an assumed input to the identifiability guarantee, and discuss how the regularizer approximates this condition in the empirical setting. revision: yes

  2. Referee: [§4] §4 (Surrogate Regularizer): The randomized masked finite-difference regularizer penalizes non-sparsity only in expectation over masks and finite-difference steps. The manuscript provides no proof that any minimizer of this surrogate objective necessarily satisfies the exact, deterministic support condition used in the §3 theorem. Without such a guarantee, the practical algorithm may admit solutions whose effective support differs from the assumed pattern, leaving residual MPAs that the single anchor cannot disambiguate. A counter-example or a lemma showing implication from surrogate to exact support would resolve this.

    Authors: This is a valid observation. The current manuscript does not include a formal proof establishing that minimizers of the surrogate regularizer exactly satisfy the deterministic Jacobian support condition required by the theorem. We acknowledge this as a theoretical gap between the surrogate objective and the identifiability assumption. In the revised manuscript, we will add a remark in Section 4 discussing this limitation and provide additional empirical analysis demonstrating that the regularizer reliably induces the desired sparsity pattern on both synthetic and real data. A complete lemma connecting the expectation-based surrogate to the exact support condition is beyond the scope of the current work but represents an interesting direction for future research. revision: partial

standing simulated objections not resolved
  • Formal proof that the surrogate regularizer's minimizers satisfy the exact Jacobian support pattern assumed in the identifiability theorem.

Circularity Check

0 steps flagged

No circularity: identifiability theorem rests on explicit external assumption

full rationale

The paper states an identifiability result under a structural sparsity condition on the Jacobian support pattern as a premise, then shows that distribution matching plus one anchor suffices given that premise. This is a standard conditional theorem, not a reduction of the conclusion to the inputs by construction. The randomized masked finite-difference regularizer is introduced separately as a practical surrogate for enforcing the assumption in high dimensions; the theoretical claim does not rely on the regularizer equaling the exact support pattern. No self-citations, fitted inputs renamed as predictions, or ansatzes smuggled via prior work appear in the derivation chain. The result is therefore self-contained against the stated assumption.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the structural sparsity assumption for the Jacobian support pattern and on the existence of at least one paired anchor sample. No free parameters are explicitly fitted in the abstract statement of the theory, though the practical regularizer introduces hyperparameters for the masking and finite-difference steps.

free parameters (1)
  • masking probability and finite-difference step size
    Hyperparameters controlling the randomized masked finite-difference approximation to the Jacobian sparsity regularizer.
axioms (1)
  • domain assumption The ground-truth transfer mapping has a fixed, sparse support pattern in its Jacobian.
    Invoked to eliminate measure-preserving automorphisms and achieve identifiability with one anchor.

pith-pipeline@v0.9.0 · 5718 in / 1279 out tokens · 32470 ms · 2026-05-20T13:29:39.624469+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

135 extracted references · 135 canonical work pages · 2 internal anchors

  1. [1]

    Langley , title =

    P. Langley , title =. Proc. International Conference on Machine Learning (ICML) , address =. 2000 , pages =

    work page 2000

  2. [2]

    T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

    work page 1980

  3. [3]

    M. J. Kearns , title =

    work page

  4. [4]

    Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

    work page 1983

  5. [5]

    R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

    work page 2000

  6. [6]

    Suppressed for Anonymity , author=

    work page

  7. [7]

    Newell and P

    A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

    work page 1981

  8. [8]

    A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

    work page 1959

  9. [9]

    Neural Information Processing Systems (NeurIPS) , year =

    Generative Adversarial Networks , author =. Neural Information Processing Systems (NeurIPS) , year =

    work page

  10. [10]

    Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks , author =. Proc. IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    work page

  11. [11]

    Kernel of

    Moriakov, Nikita and Adler, Jonas and Teuwen, Jonas , booktitle =. Kernel of

    work page

  12. [12]

    The Role of Minimal Complexity Functions in Unsupervised Learning of Semantic Mappings , author =. Proc. International Conference on Learning Representations (ICLR) , year =

    work page

  13. [13]

    Towards Identifiable Unsupervised Domain Translation: A Diversified Distribution Matching Approach , author =. Proc. International Conference on Learning Representations (ICLR) , year =

    work page

  14. [14]

    Content-Style Learning from Unaligned Domains: Identifiability under Unknown Latent Dimensions , author =. Proc. International Conference on Learning Representations (ICLR) , year =

    work page

  15. [15]

    Proceedings of the IEEE , volume=

    A comprehensive survey on transfer learning , author=. Proceedings of the IEEE , volume=. 2020 , publisher=

    work page 2020

  16. [16]

    Journal of machine learning research , volume=

    Domain-adversarial training of neural networks , author=. Journal of machine learning research , volume=

    work page

  17. [17]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Optimal transport for domain adaptation , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2016 , publisher=

    work page 2016

  18. [18]

    Unsupervised cross-domain image generation , author=. Proc. International Conference on Learning Representations (ICLR) , year=

    work page

  19. [19]

    Learning to discover cross-domain relations with generative adversarial networks , author=. Proc. International Conference on Machine Learning (ICML) , pages=

    work page

  20. [20]

    Image-to-image translation with conditional adversarial networks , author=. Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    work page

  21. [21]

    Biometrika , year =

    Harold Hotelling , title =. Biometrika , year =

    work page

  22. [22]

    IEEE Signal Processing Magazine , volume=

    Nonnegative Matrix Factorization for Signal and Data Analytics: Identifiability, Algorithms, and Applications , author=. IEEE Signal Processing Magazine , volume=. 2019 , publisher=

    work page 2019

  23. [23]

    Identifiability results for multimodal contrastive learning , author=. Proc. International Conference on Learning Representations (ICLR) , year=

    work page

  24. [24]

    Deep canonical correlation analysis , author=. Proc. International Conference on Machine Learning (ICML) , pages=

    work page

  25. [25]

    IEEE Transactions on Multimedia , volume=

    Image-to-image translation: Methods and applications , author=. IEEE Transactions on Multimedia , volume=. 2021 , publisher=

    work page 2021

  26. [26]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Multimodal machine learning: A survey and taxonomy , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2018 , publisher=

    work page 2018

  27. [27]

    Understanding latent correlation-based multiview learning and self-supervision: An identifiability perspective , author=. Proc. International Conference on Learning Representations (ICLR) , year=

    work page

  28. [28]

    Neural Information Processing Systems (NeurIPS) , volume=

    Self-supervised learning with data augmentations provably isolates content from style , author=. Neural Information Processing Systems (NeurIPS) , volume=

    work page

  29. [29]

    Word translation without parallel data , author=. Proc. International Conference on Learning Representations (ICLR) , year=

    work page

  30. [30]

    Multimodal unsupervised image-to-image translation , author=. Proc. European Conference on Computer Vision (ECCV) , pages=

    work page

  31. [31]

    Identifiability Conditions for Domain Adaptation , author=. Proc. International Conference on Machine Learning (ICML) , pages=

    work page

  32. [32]

    Partial disentanglement for domain adaptation , author=. Proc. International Conference on Machine Learning (ICML) , pages=

    work page

  33. [33]

    Identifying Latent Causal Content for Multi-Source Domain Adaptation , author=. Proc. International Conference on Learning Representations (ICLR) , year=

    work page

  34. [34]

    Neural Information Processing Systems (NeurIPS) , volume=

    Unsupervised image-to-image translation networks , author=. Neural Information Processing Systems (NeurIPS) , volume=

    work page

  35. [35]

    Maximum Spatial Perturbation Consistency for Unpaired Image-to-Image Translation , author=. Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    work page

  36. [36]

    Kim, Junho and Kim, Minjae and Kang, Hyeonwoo and Lee, Kwanghee , booktitle=

    work page

  37. [37]

    Diverse image-to-image translation via disentangled representations , author=. Proc. European Conference on Computer Vision (ECCV) , pages=

    work page

  38. [38]

    2020 , publisher=

    Lee, Hsin-Ying and Tseng, Hung-Yu and Mao, Qi and Huang, Jia-Bin and Lu, Yu-Ding and Singh, Maneesh and Yang, Ming-Hsuan , journal=. 2020 , publisher=

    work page 2020

  39. [39]

    Neural Information Processing Systems (NeurIPS) , volume=

    Toward multimodal image-to-image translation , author=. Neural Information Processing Systems (NeurIPS) , volume=

    work page

  40. [40]

    Choi, Yunjey and Choi, Minje and Kim, Munyoung and Ha, Jung-Woo and Kim, Sunghun and Choo, Jaegul , booktitle=

    work page

  41. [41]

    Choi, Yunjey and Uh, Youngjung and Yoo, Jaejun and Ha, Jung-Woo , booktitle=

    work page

  42. [42]

    Dual diffusion implicit bridges for image-to-image translation , author=. Proc. International Conference on Learning Representations (ICLR) , year=

    work page

  43. [43]

    Sasaki, Hiroshi and Willcocks, Chris G and Breckon, Toby P , journal=

    work page

  44. [44]

    GLIDE: Towards Photorealistic Image Generation and Editing with Text- Guided Diffusion Models

    Pretraining is all you need for image-to-image translation , author=. arXiv preprint arXiv:2205.12952 , year=

    work page arXiv

  45. [45]

    Contrastive learning for unpaired image-to-image translation , author=. Proc. European Conference on Computer Vision (ECCV) , pages=

    work page

  46. [46]

    2020 , howpublished =

    Mostafa Mozafari , title =. 2020 , howpublished =

    work page 2020

  47. [47]

    Wu, Wayne and Cao, Kaidi and Li, Cheng and Qian, Chen and Loy, Chen Change , booktitle=

    work page

  48. [48]

    Improving shape deformation in unsupervised image-to-image translation , author=. Proc. European Conference on Computer Vision (ECCV) , pages=

    work page

  49. [49]

    Geometry-consistent generative adversarial networks for one-sided unsupervised domain mapping , author=. Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    work page

  50. [50]

    Layer Normalization

    Layer normalization , author=. arXiv preprint arXiv:1607.06450 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  51. [51]

    Learning transferable visual models from natural language supervision , author=. Proc. International Conference on Machine Learning (ICML) , pages=

    work page

  52. [52]

    Discriminative region proposal adversarial networks for high-quality image-to-image translation , author=. Proc. European Conference on Computer Vision (ECCV) , pages=

    work page

  53. [53]

    High-resolution image synthesis and semantic manipulation with conditional

    Wang, Ting-Chun and Liu, Ming-Yu and Zhu, Jun-Yan and Tao, Andrew and Kautz, Jan and Catanzaro, Bryan , booktitle=. High-resolution image synthesis and semantic manipulation with conditional

    work page

  54. [54]

    Least squares generative adversarial networks , author=. Proc. IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    work page

  55. [55]

    Deep Learning Face Attributes in the Wild , author =. Proc. IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    work page

  56. [56]

    Nowozin, Sebastian and Cseke, Botond and Tomioka, Ryota , booktitle=

    work page

  57. [57]

    Yi, Zili and Zhang, Hao and Tan, Ping and Gong, Minglun , booktitle=

    work page

  58. [58]

    Amodio, Matthew and Krishnaswamy, Smita , booktitle=

    work page

  59. [59]

    Unpaired image-to-image translation using adversarial consistency loss , author=. Proc. European Conference on Computer Vision (ECCV) , pages=

    work page

  60. [60]

    Torbunov, Dmitrii and Huang, Yi and Yu, Haiwang and Huang, Jin and Yoo, Shinjae and Lin, Meifeng and Viren, Brett and Ren, Yihui , booktitle=

    work page

  61. [61]

    Neural Information Processing Systems (NeurIPS) , volume=

    One-sided unsupervised domain mapping , author=. Neural Information Processing Systems (NeurIPS) , volume=

    work page

  62. [62]

    Chen, Xinyuan and Xu, Chang and Yang, Xiaokang and Tao, Dacheng , booktitle=

    work page

  63. [63]

    Mode seeking generative adversarial networks for diverse image synthesis , author=. Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    work page

  64. [64]

    Yu, Lu and van de Weijer, Joost and others , booktitle=

    work page

  65. [65]

    Neural Information Processing Systems (NeurIPS) , volume=

    Breaking the dilemma of medical image-to-image translation , author=. Neural Information Processing Systems (NeurIPS) , volume=

    work page

  66. [66]

    Neural Information Processing Systems (NeurIPS) , volume=

    Unsupervised attention-guided image-to-image translation , author=. Neural Information Processing Systems (NeurIPS) , volume=

    work page

  67. [67]

    1976 , publisher=

    Principles of mathematical analysis , author=. 1976 , publisher=

    work page 1976

  68. [68]

    2000 , publisher=

    Real analysis , author=. 2000 , publisher=

    work page 2000

  69. [69]

    Analyse des liaisons de probabilit

    Darmois, George , booktitle=. Analyse des liaisons de probabilit

    work page

  70. [70]

    Neural networks , volume=

    Nonlinear independent component analysis: Existence and uniqueness results , author=. Neural networks , volume=

    work page

  71. [71]

    Nonlinear

    Hyvarinen, Aapo and Morioka, Hiroshi , booktitle=. Nonlinear

    work page

  72. [72]

    Contrastive learning inverts the data generating process , author=. Proc. International Conference on Machine Learning (ICML) , pages=

    work page

  73. [73]

    Kingma and Jimmy Ba , title =

    Diederik P. Kingma and Jimmy Ba , title =. Proc. International Conference on Learning Representations (ICLR) , year =

    work page

  74. [74]

    Rectifier nonlinearities improve neural network acoustic models , author=. Proc. International Conference on Machine Learning (ICML) , volume=

    work page

  75. [75]

    Deep sparse rectifier neural networks , author=. Proc. International Conference on Artificial Intelligence and Statistics (AISTATS) , pages=

    work page

  76. [76]

    LeCun, Yann and Cortes, Corinna and Burges, CJ , journal=

    work page

  77. [77]

    2000 , publisher=

    An introduction to ergodic theory , author=. 2000 , publisher=

    work page 2000

  78. [78]

    Neural Information Processing Systems (NeurIPS) , volume=

    Joint distribution optimal transportation for domain adaptation , author=. Neural Information Processing Systems (NeurIPS) , volume=

    work page

  79. [79]

    Cross-domain image retrieval with a dual attribute-aware ranking network , author=. Proc. IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    work page

  80. [80]

    The role of minimal complexity functions in unsupervised learning of semantic mappings , author=. Proc. International Conference on Learning Representations (ICLR) , year=

    work page

Showing first 80 references.