Pith. sign in

REVIEW 5 major objections 6 minor 24 references

Domain-Adaptive Diagnosis of Lewy Body Disease with Transferability Aware Transformer

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper claims that a transferability-aware transformer adapts Alzheimer's disease data to diagnose Lewy body disease, reaching 66.7% normal-cognition and 88.9% MCI accuracy while flagging LBD as a separate class.

desk verdict The central claim is contradicted by the paper's own Table 1: TAT's LBD accuracy is 14.5%, below the 33% chance level, and the threshold that makes it work is hand-picked. read the letter →

arxiv 2507.08839 v1 pith:IPGALX6O submitted 2025-07-07 cs.LG cs.AIeess.IV

classification cs.LGcs.AIeess.IV
keywords Lewybodydiseasedomainadaptationtransferability-awareattentionstructuralconnectivityopen-setrecognitionvisiontransformerAlzheimer'sdementiadiagnosis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that scarce Lewy body disease (LBD) data can be diagnosed by transferring knowledge from the much larger Alzheimer's disease (AD) data, despite site-induced differences in acquisition. The proposed Transferability Aware Transformer (TAT) learns which local patches of structural connectivity carry disease-transferable features and weights attention toward those patches while suppressing domain-specific ones. On an ADNI-to-LBD adaptation setup, TAT reports 66.7% accuracy on normal cognition, 88.9% on mild cognitive impairment, and 14.5% on LBD, and it is the only tested method that attempts LBD as an open-set category. A sympathetic reader would care because, if correct, the method offers a route to deep-learning diagnosis for rare dementias where collecting large labeled datasets is impractical.

What carries the argument

The load-bearing object is the transferability matrix. For each structural-connectivity patch, a local discriminator $D_l$ outputs the probability that the patch comes from the AD source; the patch's transferability score is the entropy of that output, $c(\tilde{z}_{ip}) = H(D_l(\tilde{z}_{ip})) \in [0,1]$, so an uncertain discriminator near 0.5 yields a score near 1. These scores form a learned transferability-aware graph with adjacency matrix $A$, which is multiplied into the self-attention logits via $\mathrm{TAG\text{-}SA}(Q,K,V,A) = \mathrm{softmax}((QK^\top \odot A)/\sqrt{d_k})V$, and the final layer's class-token attention is reweighted as $TAS(q_{cls},K,V) = \mathrm{softmax}(q_{cls}K^\top/\sqrt{d}) \odot [1; C(K_{\mathrm{patch}})]V$. A global discriminator aligns whole-image class-token representations, and an entropy threshold $\tau$ on the CN/MCI classifier output flags LBD as an unknown class.

What would settle it

One decisive check is to randomize the transferability matrix while keeping all other components fixed; if the reported accuracy does not drop materially, the transferability mechanism is not doing the work. A complementary external check is to compare the high-transferability structural-connectivity patches against known neuroanatomical markers of Lewy body disease, such as posterior cortical involvement; if high-entropy patches are not enriched for disease-relevant shared structure, the paper's interpretation of equation (2) is not supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that a transferability-aware transformer reduces cross-disease domain shift by explicitly learning, from a local discriminator, which image patches are transferable between AD and LBD. The authors report that TAT outperforms closed-set Transformer baselines on CN and MCI categories and is the only one able to flag LBD subjects at all, with the LBD accuracy limited to 14.5 ± 4.3% but the overall framework presented as a first demonstration. The paper further claims that both the patch-level local discriminator and the whole-image global discriminator contribute to this result, since removing either one lowers accuracy in the ablation study.

Load-bearing premise

The method assumes that a patch whose domain discriminator is maximally uncertain (output near 0.5) really is a disease-transferable patch; if high-entropy patches instead reflect acquisition noise or unresolved domain overlap, the transferability matrix amplifies the wrong features, and the hand-set threshold $\tau$ becomes a tuned heuristic.

Editorial extensions

If this is right

  • If correct, TAT makes deep-learning diagnosis feasible for LBD and other rare dementias by borrowing abundant AD data, sidestepping the need for large target-site cohorts.
  • The learned transferability matrix provides a patch-level map of which structural-connectivity regions are shared between AD and LBD, giving a concrete neuroanatomical hypothesis for follow-up studies.
  • The open-set design means a model trained only to separate CN from MCI can still flag LBD as an unknown class, which matters because medical label sets differ across diseases and sites.
  • Ablation results imply that both patch-level and whole-image domain alignment are needed, since removing either the local or global discriminator degrades accuracy.
  • The framework extends in principle to any source/target disease pair with overlapping clinical stages and site-induced domain shift, a direction the paper proposes for rare-disease diagnosis.

Reading between the lines

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

  • An implicit extension of the paper's mechanism is to use the transferability scores as a hypothesis generator for shared disease mechanisms, then validate the top-scoring structural-connectivity patches against autopsy-confirmed Lewy body pathology.
  • Because the open-set threshold $\tau$ is hand-set and the ablation shows it trades LBD recall against CN/MCI accuracy, an automatic threshold-selection rule would be needed before the method could run without per-dataset tuning.
  • The comparison set is restricted to closed-set Transformer baselines, so benchmarking against open-set or partial-set domain adaptation methods would clarify whether the gain comes from the transferability mechanism or from the open-set formulation itself.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes TAT, a ViT-based domain adaptation method that uses a local discriminator to compute patch-level transferability scores and a global discriminator to align source (AD) and target (LBD) domains, plus a threshold-based open-set head to detect LBD as an unknown class. The method is evaluated on structural connectivity data from ADNI and a private LBD dataset. The reported results (Table 1) show CN 66.7%, MCI 88.9%, and LBD 14.5% accuracy for TAT.

Significance. The problem is relevant: data scarcity in LBD diagnosis is a real obstacle, and a working domain adaptation framework would be valuable. The paper is purely empirical, with no code or formal guarantees, and the central experimental claim is not supported. The main metric (LBD accuracy) is below chance, and the open-set detection is threshold-dominated. The idea of weighting features by discriminator entropy is interesting but unvalidated. The authors do not provide reproducible code or machine-checked proofs, which further limits the contribution.

major comments (5)
  1. [§3.2, Table 1] The central claim of the abstract is that TAT improves LBD diagnostic accuracy. Table 1 reports LBD accuracy of 14.5% ± 4.3 on 77 LBD subjects, which corresponds to about 11 correct classifications. This is below the 33% accuracy of uniform random three-class guessing (binomial p < 0.001). A trivial majority-class classifier that always predicts LBD would achieve 72.6% overall accuracy, far above TAT's roughly 30% overall accuracy. The reported results therefore contradict the paper's headline claim.
  2. [§3.3, Table 2] The ablation in Table 2 shows that the LBD accuracy is dominated by the hand-set threshold τ. Changing τ from 0.8 to 0.7 raises LBD accuracy from 14.5% to 54.1% while degrading CN from 66.7% to 47.8% and MCI from 88.9% to 77.7%. No principled selection rule for τ is given, so the reported operating point appears to be a tuned heuristic rather than a robust property of the method. This undermines the claim that TAT reliably detects LBD.
  3. [§2.5, Eq. (13)] The open-set detection mechanism is a threshold on the entropy of a binary CN-vs-MCI classifier's output. Because the classifier is trained only on CN and MCI, the assignment of high-entropy samples to LBD is an unvalidated assumption. In particular, with only 6 MCI test subjects (Table 1), the entropy distribution for MCI is extremely noisy; a threshold that separates CN/MCI from LBD on 6 subjects cannot be established reliably. The paper provides no calibration or sensitivity analysis beyond the two τ values in Table 2.
  4. [§2.2, Eq. (2)] The transferability score c(z_ip) = H(D_l(z_ip)) is the core of the method: high entropy of the local discriminator output is equated with high cross-disease transferability. This equivalence is asserted without external validation. If patches with discriminator output near 0.5 are merely noisy or represent unresolved domain overlap, the transferability matrix will amplify irrelevant features. The paper lacks an ablation that replaces the entropy-based score with a baseline such as random weighting or an anatomical prior, so the mechanism's benefit is not established.
  5. [§3.2, Table 1] The comparison with baselines is not meaningful for the LBD class. ViT, TVT, and SSRT are closed-set methods and report N/A for LBD; only TAT reports an LBD number, and that number is below chance. Thus Table 1 does not demonstrate that TAT outperforms state-of-the-art domain adaptation methods; it demonstrates that the baselines cannot address the open-set problem, while TAT's solution performs worse than random guessing on the target disease.
minor comments (6)
  1. [§2.2, Eq. (3)] The formula sums over H heads although C_i does not depend on the head index; the head dimension in the sum is spurious. Additionally, the '×' at the end of Eq. (3) appears to be a dangling artifact.
  2. [§2.2, Eq. (4) and Fig. 1] The abbreviation for Transferability Aware Self-Attention is inconsistent: the text uses TAS, while Figure 1 and Eq. (4) caption use TSA.
  3. [§2.3, Eq. (7)] It is unclear whether Q, K, V include the class token; if they include all tokens, the adjacency matrix A defined in Eq. (3) over P patches has incompatible shape, whereas if they include only patches, the role of the class token in the graph-guided layers is unstated.
  4. [§2.2, Eq. (2) and §2.5, Eq. (12)] The entropy function H is not specified with a base; stating H ∈ [0,1] implies base 2, but Eq. (12) uses the same notation without specifying the base, which matters for interpreting the threshold τ = 0.8.
  5. [§3.1] The MCI target set has only 6 subjects; the reported MCI accuracy of 88.9% ± 7.9 corresponds to about 5 correct subjects, and the standard deviation is therefore not meaningful. The paper should state this limitation explicitly.
  6. [Throughout] Section headings contain stray spaces (e.g., '2.2 T ransferability A ware T ransformer Layer'), and equations use inconsistent capitalization (LCE vs Lce).

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the transferability mechanism is an explicit modeling assumption, and the experimental claim is tested against external baselines.

full rationale

The derivation chain is not circular. The transferability score c = H(D_l(z_ip)) (Eq. 2) is defined as the entropy of a local discriminator output, but this is an explicit assumption about what makes patches transferable, not a result derived from the target disease labels. The local discriminator is trained only on macro domain labels (AD vs. LBD), and the transferability matrix in Eq. 3 is explicitly detached from backpropagation, so the attention reweighting is not fitted to LBD diagnostic labels. The classifier is trained on ADNI CN/MCI labels, and LBD detection is a thresholded entropy rule (Eq. 13) with a hand-set threshold tau; this is a tuned heuristic rather than a prediction forced by construction. The central experimental claim is supported by comparisons with external baselines (ViT, TVT, SSRT) and by ablations, so the result is not equivalent to its inputs. The paper does reuse the authors' prior TAT framework (Refs. [11], [14]) and cites many of the authors' own works, but these self-citations are not load-bearing: the method is described in the paper itself, and the LBD application is independently evaluated. The reported LBD accuracy of 14.5% in Table 1, below the 33% chance level, is a serious correctness/robustness problem, but it is not circularity; the threshold sensitivity shown in Table 2 is an honest ablation, not a fitted parameter renamed as prediction. No equation reduces to another by construction, and no load-bearing claim rests solely on a self-citation chain.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The method's effectiveness rests on three hand-set hyperparameters (alpha, beta, tau) and the unvalidated assumption that discriminator entropy equals transferability. The sensitivity of LBD accuracy to tau (Table 2) shows the central result is not stable under parameter choice.

free parameters (3)
  • alpha (global discriminator weight) = 1.0 (default; 0.1 in ablation)
    Hand-set in Eq. (11) despite being called 'trainable'; ablation shows sensitivity.
  • beta (local discriminator weight) = 0.01 (default; 0.1 in ablation)
    Hand-set; fixed after initialization.
  • tau (entropy threshold for LBD detection) = 0.8 (default; 0.7 in ablation)
    Hand-set; directly controls the tradeoff between LBD recall and CN/MCI accuracy (Table 2).
assumptions (4)
  • domain assumption Structural connectivity derived from structural MRI is a valid input representation for cross-disease transfer between AD and LBD.
    Invoked in Section 2.1 and dataset section 3.1; if SC does not carry disease-discriminative signals transferable across sites, the whole method has no substrate.
  • domain assumption LBD and AD share transferable neuropathological features.
    Stated in Figure 1 caption ('share several common neuropathology'); motivates the transferability concept.
  • ad hoc to paper The entropy of a binary domain discriminator's output is a valid measure of patch transferability.
    This is Eq. (2), the core mechanism; there is no external validation that high-entropy patches are disease-transferable rather than noisy.
  • domain assumption Unlabeled LBD samples during training are sufficient to align domains without label information.
    Standard unsupervised domain adaptation assumption, used throughout the method.
invented entities (1)
  • Transferability score c(z_ip)
    purpose: Weights self-attention to favor patches the discriminator cannot distinguish between domains
    Defined via Eq. (2) with no external benchmark; it is an internal model construct, not a measured quantity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Domain-Adaptive Diagnosis of Lewy Body Disease with Transferability Aware Transformer." pith.science (2026). https://pith.science/paper/IPGALX6O

@misc{pith2026250708839,
  author       = {Pith},
  title        = {Pith review of: Domain-Adaptive Diagnosis of Lewy Body Disease with Transferability Aware Transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IPGALX6O}},
  note         = {Machine review of arXiv:2507.08839}
}
read the original abstract

Lewy Body Disease (LBD) is a common yet understudied form of dementia that imposes a significant burden on public health. It shares clinical similarities with Alzheimer's disease (AD), as both progress through stages of normal cognition, mild cognitive impairment, and dementia. A major obstacle in LBD diagnosis is data scarcity, which limits the effectiveness of deep learning. In contrast, AD datasets are more abundant, offering potential for knowledge transfer. However, LBD and AD data are typically collected from different sites using different machines and protocols, resulting in a distinct domain shift. To effectively leverage AD data while mitigating domain shift, we propose a Transferability Aware Transformer (TAT) that adapts knowledge from AD to enhance LBD diagnosis. Our method utilizes structural connectivity (SC) derived from structural MRI as training data. Built on the attention mechanism, TAT adaptively assigns greater weights to disease-transferable features while suppressing domain-specific ones, thereby reducing domain shift and improving diagnostic accuracy with limited LBD data. The experimental results demonstrate the effectiveness of TAT. To the best of our knowledge, this is the first study to explore domain adaptation from AD to LBD under conditions of data scarcity and domain shift, providing a promising framework for domain-adaptive diagnosis of rare diseases.

Figures

Figures reproduced from arXiv: 2507.08839 by the authors.

Figure 1
Figure 1. Illustration of proposed framework. The TAT layer employs a local dis [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The adjacency matrix of the learned TAG. where H(P) quantifies the uncertainty of the prediction. If the entropy exceeds a predefined threshold τ , the sample is considered to belong to the unknown LBD category: P redicted Label =  LBD, if H(P) > τ arg max(P), otherwise (13) 3 Experiments 3.1 Datasets and Training Details We use the structural connectivity (SC) constructed from MRI images, and the pre-processing fo… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 19 canonical work pages

  1. [1]

    L., Boeve, B

    Walker, Z., Possin, K. L., Boeve, B. F. and Aarsland, D.: Lewy body dementias. The Lancet 386(10004), 1683-1697 (2015)

  2. [2]

    and Zhang, L.: Classification of alzheimer’s disease via vision transformer

    Lyu, Y., Yu, X., Zhu, D. and Zhang, L.: Classification of alzheimer’s disease via vision transformer. In Proceedings of the 15th international conference on PErvasive technologies related to assistive environments, pp. 463-468. (2022)

  3. [3]

    and Zhu, D.: Mild Cognitive Impairment Classification Using A Novel Finer-Scale Brain Connectome

    Lyu, Y., Zhang, L., Yu, X., Cao, C., Liu, T. and Zhu, D.: Mild Cognitive Impairment Classification Using A Novel Finer-Scale Brain Connectome. In IEEE International Symposium on Biomedical Imaging (ISBI), pp. 1-5. (2024)

  4. [4]

    D., Farlow, M

    Capouch, S. D., Farlow, M. R. and Brosch, J. R.: A review of dementia with Lewy bodies’ impact, diagnostic criteria and treatment. Neurology and Therapy7, 249- 263 (2018)

  5. [5]

    and Zhu, D.: Classification of Mild Cognitive Impairment Based on Dy- namic Functional Connectivity Using Spatio-Temporal Transformer

    Zhang, J., Lyu, Y., Yu, X., Zhang, L., Cao, C., Chen, T., Chen, M., Zhuang, Y., Liu, T. and Zhu, D.: Classification of Mild Cognitive Impairment Based on Dy- namic Functional Connectivity Using Spatio-Temporal Transformer. arXiv preprint arXiv:2501.16409 (2025)

  6. [6]

    and Zhu, D.: Brain-Adapter: Enhancing Neurological Disorder Anal- ysis with Adapter-Tuning Multimodal Large Language Models

    Zhang, J., Yu, X., Lyu, Y., Zhang, L., Chen, T., Cao, C., Zhuang, Y., Chen, M., Liu, T. and Zhu, D.: Brain-Adapter: Enhancing Neurological Disorder Anal- ysis with Adapter-Tuning Multimodal Large Language Models. arXiv preprint arXiv:2501.16282 (2025)

  7. [7]

    Zhu, D.C, Zhang, R

    Yu, X., Scheel, N., Zhang, L. Zhu, D.C, Zhang, R. and Zhu, D.: Free water in T2 FLAIR white matter hyperintensity lesions. Alzheimer’s & Dementia, e057398 (2021)

  8. [8]

    BrainNet-MoE: Brain-Inspired Mixture-of-Experts Learning for Neurological Disease Identification

    Zhang, J., Yu, X., Chen, T., Cao, C., Chen, M., Zhuang, Y., Lyu, Y., Zhang, L., Su, L., Liu, T. and Zhu, D.: BrainNet-MoE: Brain-Inspired Mixture-of-Experts Learning for Neurological Disease Identification. arXiv preprint arXiv:2503.07640 (2025)

Show all 24 references
  1. [9]

    and Ma, P.: Knowledge Distillation and Dataset Distillation of Large Language Models: Emerging Trends, Challenges, and Future Directions

    Fang, L., Yu, X., Cai, J., Chen, Y., Wu, S., Liu, Z., Yang, Z., Lu, X., Gong, X., Liu, Y., Ma, T., Ruan, W., Abbasi, A., Zhang, J., Wang, T., Latif, E., Liu, W., Zhang, W., Kolouri, S., Zhai, X., Zhu, D., Zhong, W., Liu, T. and Ma, P.: Knowledge Distillation and Dataset Distil...

  2. [10]

    and Zhu, D.: Gp-gpt: Large language model for gene-phenotype mapping

    Lyu, Y., Wu, Z., Zhang, L., Zhang, J., Li, Y., Ruan, W., Liu, Z., Yu, X., Cao, C., Chen, T., Chen, M., Zhuang, Y., Li, X., Liu, R., Huang, C., Li, W., Liu, T. and Zhu, D.: Gp-gpt: Large language model for gene-phenotype mapping. arXiv preprint arXiv:2409.09825 (2024)

  3. [11]

    and Zhang, Z.: Feature fusion transferability aware transformer for unsupervised domain adaptation

    Yu, X., Huang, Z. and Zhang, Z.: Feature fusion transferability aware transformer for unsupervised domain adaptation. In Proceedings of the IEEE/CVF winter con- ference on applications of computer vision (WACV), pp. 6752-6761. (2025) 10 X. Yu et al

  4. [12]

    and Zhu, D.: Classification of mild cognitive impair- ment by fusing neuroimaging and gene expression data

    Lyu, Y., Yu, X., Zhang, L. and Zhu, D.: Classification of mild cognitive impair- ment by fusing neuroimaging and gene expression data. In Proceedings of the 14th PErvasive Technologies Related to Assistive Environments Conference, pp. 26-32. (2021)

  5. [13]

    and Zhu, D.: Core-Periphery Principle Guided State Space Model for Functional Connectome Classification

    Chen, M., Yu, X., Zhang, J., Chen, T., Cao, C., Zhuang, Y., Lyu, Y., Zhang, L., Liu, T. and Zhu, D.: Core-Periphery Principle Guided State Space Model for Functional Connectome Classification. arXiv preprint arXiv:2503.14655 (2025)

  6. [14]

    and Liu, T.: Robust core-periphery constrained transformer for domain adaptation

    Yu, X., Zhu, D. and Liu, T.: Robust core-periphery constrained transformer for domain adaptation. arXiv preprint arXiv:2308.13515 (2023)

  7. [15]

    and Zhu, D.: Core-Periphery Multi-Modality Feature Alignment for Zero-Shot Medical Image Analysis

    Yu, X., Zhang, L., Wu, Z. and Zhu, D.: Core-Periphery Multi-Modality Feature Alignment for Zero-Shot Medical Image Analysis. IEEE Transactions on Medical Imaging (2024)

  8. [16]

    and Zhu, D.: Cp-clip: Core-periphery feature alignment clip for zero-shot medical image analysis

    Yu, X., Wu, Z., Zhang, L., Zhang, J., Lyu, Y. and Zhu, D.: Cp-clip: Core-periphery feature alignment clip for zero-shot medical image analysis. In International Confer- ence on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pp. 88-97. (2024)

  9. [17]

    and Zhu, D.: Representative Functional Con- nectivity Learning for Multiple Clinical groups in Alzheimer’s Disease

    Zhang, L., Yu, X., Lyu, Y., Liu, T. and Zhu, D.: Representative Functional Con- nectivity Learning for Multiple Clinical groups in Alzheimer’s Disease. In IEEE International Symposium on Biomedical Imaging (ISBI), pp. 1-5. (2023)

  10. [18]

    and Zhu, D.: Supervised deep tree in alzheimer’s disease

    Yu, X., Zhang, L., Lyu, Y., Liu, T. and Zhu, D.: Supervised deep tree in alzheimer’s disease. In IEEE International Symposium on Biomedical Imaging (ISBI), pp. 1-5. (2023)

  11. [19]

    and Hughes, M.C.: Interlude: Interactions between labeled and unlabeled data to enhance semi-supervised learning

    Huang, Z., Yu, X., Zhu, D. and Hughes, M.C.: Interlude: Interactions between labeled and unlabeled data to enhance semi-supervised learning. arXiv preprint arXiv:2403.10658 (2024)

  12. [20]

    and Niu, W.: Real-time core-periphery guided ViT with smart data layout selection on mo- bile devices

    Shu, Z., Yu, X., Wu, Z., Jia, W., Shi, Y., Yin, M., Liu, T., Zhu, D. and Niu, W.: Real-time core-periphery guided ViT with smart data layout selection on mo- bile devices. Advances in Neural Information Processing Systems, pp. 95744-95763. (2024)

  13. [21]

    and Zhu, D.: Noisynn: Exploring the influence of information entropy change in learning systems

    Yu, X., Huang, Z., Xue, Y., Zhang, L., Wang, L., Liu, T. and Zhu, D.: Noisynn: Exploring the influence of information entropy change in learning systems. arXiv preprint arXiv:2309.10625 (2023)

  14. [22]

    and Zhu, D.: Core-periphery principle guided redesign of self-attention in transformers

    Yu, X., Zhang, L., Dai, H., Lyu, Y., Zhao, L., Wu, Z., Liu, D., Liu, T. and Zhu, D.: Core-periphery principle guided redesign of self-attention in transformers. arXiv preprint arXiv:2303.15569 (2023)

  15. [23]

    and Huang, J.: Tvt: Transferable vision transformer for un- supervised domain adaptation

    Yang, J., Liu, J., Xu, N. and Huang, J.: Tvt: Transferable vision transformer for un- supervised domain adaptation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision (WACV), pp. 520-530. (2023)

  16. [24]

    and Ling, H.: Safe self-refinement for transformer-based domain adaptation

    Sun, T., Lu, C., Zhang, T. and Ling, H.: Safe self-refinement for transformer-based domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pp. 7191-7200. (2022)

Pith tools

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