Pith. sign in

REVIEW 4 major objections 4 minor 58 references

AME: Aligned Manifold Entropy for Robust Vision-Language Distillation

T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Adding a shared-manifold entropy term to vision-language distillation improves few-shot generalization and tightens the theoretical generalization bound.

desk verdict A broadly tested entropy-minimization module for VLM distillation that likely helps on average, but the paper's theoretical claim of a tighter bound is not proved — the DPI step in the appendix is simply wrong. read the letter →

arxiv 2508.08644 v1 pith:BBG5NRRP submitted 2025-08-12 cs.CV

Guiming Cao , Yuming Ou This is my paper · ORCID
classification cs.CV
keywords vision-languagemodelsknowledgedistillationentropyminimizationfew-shotgeneralizationsharedmanifoldcross-modalrepresentationpromptlearningbound
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

The paper argues that vision-language knowledge distillation fails under limited data because ambiguous or boundary-adjacent samples carry high predictive uncertainty, and that this uncertainty can be compressed by minimizing entropy over a shared image-text manifold. It introduces AME, a plug-and-play module that projects both modalities into a common space and adds an entropy penalty to the usual distillation loss, with no change to the backbone. The central empirical claim is that this module improves the PromptKD distillation baseline by 2.79% in average harmonic-mean accuracy under 16-shot training across 11 datasets, and by 1.92% in cross-dataset transfer. The central theoretical claim is that the combined distillation-plus-entropy objective satisfies a tighter generalization error bound than distillation alone, giving a principled reason why the gains appear.

What carries the argument

The central object is the Reconfigured Shared Manifold (RSM), a lightweight training-time module. It forms a latent space Z by row-wise concatenation of projected text and image embeddings, averages along the feature dimension to obtain scores s_i, applies softmax to obtain probabilities p_i, and defines the manifold entropy as H(Z) = -sum_i p_i log p_i. The total loss is L_total = L_KD + lambda H(Z) with lambda = 50, where L_KD is the teacher-student KL divergence. The load-bearing identity is the paper's identification of H(Z) with the conditional joint entropy H(T,I|S) of the two modalities given the sample; this identification lets entropy minimization appear as an increase in conditional mutual information I(T;I|S) (Corollary 3.2) and, through the data processing inequality, as a reduction in the generalization gap (Corollary 3.3). The projection functions are essential to the mechanism: the paper reports that without them the entropy gradient stays nearly orthogonal to the distillation gradient, whereas with them the angle drops below 90 degrees.

What would settle it

On a small labelled dataset where the true conditional joint entropy H(T,I|S) can be estimated, or computed exactly for synthetic features, compare it with the AME manifold entropy H(Z) across training; if the two diverge or move in opposite directions, the mutual-information mechanism in Corollaries 3.2 and 3.3 is not what explains the accuracy gains.

Watch

Extended reading notes

Core claim

The paper claims that entropy minimization performs structural compression of cross-modal representations when it is applied to a deliberately reconfigured shared manifold rather than to each modality's feature space separately. Given a teacher-student distillation setup with CLIP-style encoders, the proposed Reconfigured Shared Manifold (RSM) module learns a convolution-based projection for image features and an MLP-based projection for text features, concatenates them row-wise, reduces them to scalar scores, and minimizes the entropy of the resulting softmax distribution as a regularizer. The authors report that this module, added to PromptKD, raises average harmonic-mean accuracy from 77.40% to 80.19% in 16-shot base-to-new generalization and improves cross-dataset average accuracy by 1.92%. They also claim a theoretical mechanism: the KL distillation loss induces a conditional joint distribution over the shared manifold, so the manifold entropy approximates the conditional joint entropy H(T,I|S), and minimizing it increases the conditional mutual information between modalities, which yields a tighter generalization bound of order O(sqrt((epsilon+eta)/n)).

Load-bearing premise

The argument collapses if the entropy computed over the pooled projected image and text features does not actually measure the combined uncertainty of the two modalities for a given sample, because that identification is asserted rather than derived.

Editorial extensions

If this is right

  • Under 16-shot base-to-new training, adding RSM to PromptKD raises average harmonic-mean accuracy from 77.40% to 80.19%, an improvement of 2.79%, and also surpasses the best prompt-tuning baselines on the New-class and harmonic-mean metrics.
  • Under full-shot training the same module still improves PromptKD, but by a smaller margin of 0.90% in harmonic mean, which the paper attributes to the 1/sqrt(n) rate of the generalization bound.
  • In cross-dataset generalization with 16-shot training, RSM improves the average accuracy of PromptKD by 1.92%, with the largest single-dataset gains on OxfordPets, DTD, and EuroSAT.
  • The module transfers to other prompt-tuning frameworks: on the DTD dataset it improves MaPLe by 1.66% and PromptSRC by 1.31% in harmonic mean.
  • The learnable projections are necessary for the effect: without them the entropy loss and distillation loss have near-orthogonal gradients, while with them the gradient angle falls below 90 degrees and performance improves.

Reading between the lines

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

  • If the entropy term truly approximates the conditional joint entropy of the two modalities, the same module should transfer to other paired-modality distillation settings, such as audio-text or video-text, whenever a teacher-student setup exists; the paper does not test this extension.
  • One can isolate the source of the gain by comparing RSM against the same entropy penalty applied separately to image and text feature spaces without the shared projection; the paper states that separate entropy compression does not reliably work but does not report a full 11-dataset ablation of that variant.
  • The theoretical identification of H(Z) with H(T,I|S) could be tested on synthetic data where the true conditional joint entropy is known, which would distinguish between a mechanism and an analogy.
  • The bound predicts that the benefit of entropy regularization shrinks as training shots grow, so fitting the accuracy-gain-versus-shot curve to the predicted 1/sqrt(n) scaling would provide a quantitative check beyond the paper's reported curves.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 Aligned Manifold Entropy (AME), implemented as a plug-and-play Reconfigured Shared Manifold compression (RSM) module, which projects image and text embeddings into a shared manifold and adds an entropy-minimization regularizer to a vision-language distillation loss. The authors report improved base-to-new and cross-dataset generalization when RSM is appended to PromptKD and to a simple distillation baseline, and they provide an information-theoretic analysis intended to show that entropy minimization tightens a generalization bound. The empirical evaluation covers 11 datasets, 16-shot and full-shot settings, compatibility with MaPLe and PromptSRC, and ablations of the learnable projection.

Significance. If the empirical findings hold, the contribution is practically useful: the module is lightweight, requires no backbone changes, and the reported gains (e.g., +2.79 HM under 16-shot over PromptKD and +1.92 cross-dataset average) are consistent with the stated claims. The empirical evaluation is extensive and appears internally consistent, and the ablations strengthen the practical case. However, the theoretical analysis in Section 3.3 and Appendix A contains multiple unproved or incorrect steps, and the promised tightening of the generalization bound is not established as written. The paper's core advertised novelty includes this theoretical claim, so the unsupported theory materially affects the assessment.

major comments (4)
  1. [Section 3.3, before Corollary 3.2] The claim that H(Z) 'can be interpreted as an approximation' of the conditional joint entropy H(T,I|S) is asserted without derivation. Equation (5) defines entropy over softmax probabilities of averaged projected scalar scores, while Equation (11) concerns the conditional joint entropy of the text and image embeddings. No argument connects these two objects. Theorem 3.1, which is invoked to support this identification, is stated without proof; the Appendix proves only Corollaries 3.2 and 3.3. Consequently, Corollaries 3.2 and 3.3 rest on an unverified identification.
  2. [Appendix A, Eq. (13)] The data-processing inequality is misapplied. DPI gives I(θ;S)≤I(Z;S) only if S→Z→θ is a Markov chain, i.e., if θ is conditionally independent of S given Z. Here θ is trained on S, and Z is computed from θ and S, so no such conditional independence holds; I(θ;S) can be much larger than I(Z;S). The inequality in Eq. (13) is therefore not established.
  3. [Appendix A, Eq. (14)] The chain-rule expansion is incorrect. The correct identity is I(θ;S)=I(Z;S)+I(θ;S|Z)−I(Z;S|θ). The omitted term I(Z;S|θ) is nonnegative and is exactly the information about the training data that is carried by the entropy-minimizing projections beyond Z itself; minimizing H(Z) can increase it. Moreover, with D defined as I(θ;S)−I(Z;S), the displayed inequality I(θ;S)≤H(Z)+D reduces to the unconditional bound I(Z;S)≤H(Z) and says nothing about whether H(Z)+D decreases when H(Z) is minimized. Corollary 3.3's claim of a tighter bound than the KL-only setting is therefore unsupported. No theorem is provided connecting E_S[I(W;D)−I(W;S)] to the expected generalization error, which is another gap in the same derivation.
  4. [Section 3.3, Eq. (7)] The gradient of the softmax entropy with respect to the score s_j is ∂H/∂s_j = −p_j(log p_j + H), not −(1+log p_i). The expression in Eq. (7) is the derivative of H with respect to p_i and omits the softmax Jacobian and the entropy term. The subsequent claim that this gradient encourages all scores to converge toward the dominant score s* does not follow from the correct expression.
minor comments (4)
  1. [Figure 1 caption] The caption is incomplete or garbled: 'shows the performance fluctuations' lacks a subject, and the parenthetical in part (a) is broken. Please rewrite it for clarity.
  2. [Section 4.2] The listed 'relative improvements' (e.g., 4.82% on OxfordPets) are percentage-point differences in HM, not relative improvements in the usual sense. Please either label them as percentage-point gains or compute relative percentages.
  3. [Figure 4 caption] There is a typo: 'infomation entropy' should be 'information entropy'.
  4. [Section 3.3 and Appendix A] The notation for the conditional joint entropy is inconsistent: H(T,I|S), H(T,I|S), and H'(T,I|S) are used interchangeably. Please standardize the notation.

Circularity Check

2 steps flagged · score 6.0 of 10

The theoretical tighter-bound claim reduces to definitions and an assumed Markov property; empirical gains are independent.

  1. self definitional [Section 3.3, paragraph after Theorem 3.1 and Corollary 3.2 (Eq. 11–12)]
    "The theorem demonstrates that the KL divergence loss ... induces a conditional joint distribution P(T,I|S) ... As a result, the information entropy H(Z) can be interpreted as an approximation of the conditional joint entropy H′(T,I|S)."

    Corollary 3.2's conclusion is the definition of conditional mutual information with the asserted approximation inserted: I(T;I|S)=H(T|S)+H(I|S)-H(T,I|S). If H(Z) only 'approximates' H(T,I|S), minimizing H(Z) need not decrease H(T,I|S); if it is treated as equal, the corollary restates the input assumption. No independent derivation of the identification is given, so the claimed mechanism (entropy minimization increases cross-modal mutual information) is equivalent to the paper's definition of Z rather than a derived result.

  2. other [Appendix A, Proof of Corollary 3.3, Eqs. (13)–(14); main text after Corollary 3.3]
    "Since θ is a deterministic function of (Z,S), the data processing inequality gives: I(θ;S) ≤ I(Z) (13). By applying the chain rule of mutual information, we obtain: I(θ;S) = I(Z;S)+ I(θ;S|Z) ≤ I(Z)+R ≤ H+R (14), where R = I(θ;S)−I(Z;S) quantifies the residual dependency not captured by Z."

    Eq. (13) invokes DPI, which requires the Markov chain S→Z→θ, i.e., θ conditionally independent of S given Z. That is precisely what the theorem must prove, since θ is trained on S and Z is computed from θ and S. Eq. (14) then defines R=I(θ;S)−I(Z;S), so I(θ;S)≤H(Z)+R is true by construction (it is just I(Z;S)≤H(Z)). The paper's claim that 'H+R is strictly smaller than the original mutual information I(θ;S)' would require H(Z)<I(Z;S), contradicting I(Z;S)≤H(Z). The bound is therefore a restatement of the loss plus a defined residual, not a derived consequence of entropy minimization.

full rationale

The empirical contribution is self-contained and benchmarked against external baselines (PromptKD, PromptSRC, etc.), so the reported gains (e.g., +2.79 HM under 16-shot) are not circular. The circularity is confined to the theoretical narrative. Corollary 3.2 inserts an asserted approximation H(Z)≈H(T,I|S) into the definition of conditional mutual information, making its conclusion a restatement of the assumption. Corollary 3.3's generalization bound is built by defining the residual R=I(θ;S)−I(Z;S), so Eq. (14) is tautological, while Eq. (13)'s DPI use assumes the Markov screening property that entropy minimization is supposed to establish. The claimed 'tighter bound' (H+R<I(θ;S)) is not only unproven but contradicts I(Z;S)≤H(Z). Because the central theoretical claim reduces to definitions and assumptions, while the experiments stand independently, the appropriate score is 6.

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

The central method introduces one fitted hyperparameter, the entropy weight lambda, and several unproved modeling assumptions in the theoretical section. The most consequential assumptions are that the computed entropy approximates the true conditional joint entropy and that the residual dependency in the generalization bound vanishes. These assumptions are load-bearing for the theoretical contribution but not for the empirical benchmark results.

free parameters (1)
  • lambda (entropy weight) = 50
    Coefficient of the entropy loss in Eq. (6), set to 50 with no sensitivity analysis; the magnitude of the reported gains could depend on this hand-chosen value.
assumptions (5)
  • domain assumption High-dimensional image and text data reside on low-dimensional manifolds (Manifold Hypothesis).
    Invoked in Section 3.2 to justify projecting both modalities into a shared manifold.
  • ad hoc to paper H(Z), the entropy over concatenated projected features, approximates the conditional joint entropy H(T,I|S) of text and image embeddings.
    Stated in Section 3.3 before Corollary 3.2; no proof or empirical check is provided.
  • ad hoc to paper Conditional entropies H(T|S) and H(I|S) remain approximately constant under the total loss.
    Assumed in the Appendix proof of Corollary 3.2 to conclude that reducing H(T,I|S) increases conditional mutual information.
  • ad hoc to paper The expression E_S[I(W;D) - I(W;S)] bounds generalization error, with no theorem connecting it to expected loss.
    Corollary 3.3 and Appendix A use this quantity as a generalization bound without deriving it from a stated mutual-information generalization theorem.
  • ad hoc to paper The residual dependency D in Corollary 3.3 is implicitly suppressed by the learnable projection.
    Asserted in the text following Eq. (10); no bound, estimate, or proof is given for this suppression.
invented entities (1)
  • Reconfigured Shared Manifold compression (RSM) module
    purpose: Projects image and text features into a shared space and adds an entropy-minimization loss to compress ambiguous cross-modal representations.
    The only evidence is the benchmark results reported in this paper; no code, external validation, or falsifiable prediction outside the paper is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AME: Aligned Manifold Entropy for Robust Vision-Language Distillation." pith.science (2026). https://pith.science/paper/BBG5NRRP

@misc{pith2026250808644,
  author       = {Pith},
  title        = {Pith review of: AME: Aligned Manifold Entropy for Robust Vision-Language Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BBG5NRRP}},
  note         = {Machine review of arXiv:2508.08644}
}
read the original abstract

Knowledge distillation is a long-established technique for knowledge transfer, and has regained attention in the context of the recent emergence of large vision-language models (VLMs). However, vision-language knowledge distillation often requires sufficient training data to achieve robust generalization on amples with ambiguous or boundary-adjacent representations, which are associated with high predictive uncertainty. Critically, collecting such large-scale, task-specific data for training is often impractical in real-world scenarios. To address this major challenge arising from the entanglement of uncertainty and cross-modal feature representation, we propose Aligned Manifold Entropy for Robust Vision-Language Distillation (AME), aiming to achieve robust generalization under real-world conditions. AME applies entropy minimization over a reconfigured shared manifold, where multi-modal data (i.e., image and text) are bridged through a pair of projection functions, conducive to structural compression for cross-modal feature representations. This enables robust knowledge distillation under low-data regimes, while requiring no architectural modifications to the backbone. As a result, it can serve as a plug-and-play module compatible with a wide range of vision-language distillation frameworks. Notably, our theoretical analysis reveals that integrating knowledge distillation with entropy minimization over the shared manifold leads to a tighter generalization error bound. Extensive experiments across diverse distillation architectures and training settings demonstrate that AME consistently facilitates robust knowledge distillation, resulting in superior generalization performance across a wide spectrum of downstream tasks.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 19 canonical work pages

  1. [1]

    Yoshua Bengio, Aaron Courville, and Pascal Vincent. 2013. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence 35, 8 (2013), 1798–1828

    2013

  2. [2]

    Lucas Beyer, Xiaohua Zhai, Amélie Royer, Larisa Markeeva, Rohan Anil, and Alexander Kolesnikov. 2022. Knowledge distillation: A good teacher is patient and consistent. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10925–10934

    2022

  3. [3]

    Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. 2014. Food-101 – Min- ing Discriminative Components with Random Forests. In Computer Vision – ECCV 2014, David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars (Eds.). Springer International Publishing, Cham, 446–461

    2014

  4. [4]

    Wei-Chun Chen, Chia-Che Chang, and Che-Rung Lee. 2019. Knowledge distilla- tion with feature maps for image classification. In Computer Vision–ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14 . Springer, 200–215

    work page 2019

  5. [5]

    Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. 2020. UNITER: UNiversal Image-TExt Representation Learning. arXiv:1909.11740 [cs.CV]

    arXiv 2020

  6. [6]

    Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and An- drea Vedaldi. 2013. Describing Textures in the Wild. arXiv:1311.3618 [cs.CV]

    arXiv 2013

  7. [7]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Ima- geNet: A large-scale hierarchical image database. In2009 IEEE Conference on Com- puter Vision and Pattern Recognition . 248–255. doi:10.1109/CVPR.2009.5206848

    arXiv 2009

  8. [8]

    Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. 2022. Decoupling Zero-Shot Semantic Segmentation. arXiv:2112.07910 [cs.CV]

    work page Pith review arXiv 2022

Show all 58 references
  1. [9]

    Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Shuyang Gu, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. 2022. Clip itself is a strong fine-tuner: Achieving 85.7% and 88.0% top-1 accuracy with vit-b and vit-l on imagenet. arXiv preprint arXiv:2212.06138 (2022)

    2022 arXiv

  2. [10]

    Qi Dou, Quande Liu, Pheng Ann Heng, and Ben Glocker. 2020. Unpaired multi- modal segmentation via knowledge distillation. IEEE transactions on medical imaging 39, 7 (2020), 2415–2425

    2020

  3. [11]

    Fergus, and P

    Li Fei-Fei, R. Fergus, and P. Perona. 2004. Learning Generative Visual Models from Few Training Examples: An Incremental Bayesian Approach Tested on 101 Object Categories. In 2004 Conference on Computer Vision and Pattern Recognition Workshop. 178–178. doi:10.1109/CVPR.2004.383

    2004 doi

  4. [12]

    Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. 2022. PromptDet: Towards Open-vocabulary Detec- tion using Uncurated Images. arXiv:2203.16513 [cs.CV]

    2022 arXiv

  5. [13]

    Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. 2021. CLIP-Adapter: Better Vision-Language Models with Feature Adapters. arXiv:2110.04544 [cs.CV]

    2021 arXiv

  6. [14]

    Peng Gao, Jiasen Lu, Hongsheng Li, Roozbeh Mottaghi, and Aniruddha Kembhavi

  7. [15]

    Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. 2021. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921 (2021)

    2021 arXiv

  8. [16]

    Tong He, Chunhua Shen, Zhi Tian, Dong Gong, Changming Sun, and Youliang Yan. 2019. Knowledge adaptation for efficient semantic segmentation. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition . 578–587

    2019

  9. [17]

    Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. 2019. EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification. arXiv:1709.00029 [cs.CV]

    2019 arXiv

  10. [18]

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)

    2015 arXiv

  11. [19]

    Le, Yunhsuan Sung, Zhen Li, and Tom Duerig

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. 2021. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. arXiv:2102.05918 [cs.CV]

    2021 arXiv

  12. [20]

    Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. 2023. Maple: Multi-modal prompt learning. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 19113–19122

    2023

  13. [21]

    Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. 2023. Self-regulating Prompts: Foun- dational Model Adaptation without Forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)...

    2023

  14. [22]

    Jin-Hwa Kim, Jaehyun Jun, and Byoung-Tak Zhang. 2018. Bilinear Attention Networks. arXiv:1805.07932 [cs.CV]

    2018 arXiv

  15. [23]

    Konwoo Kim, Michael Laskin, Igor Mordatch, and Deepak Pathak. 2022. How to Adapt Your Large-Scale Vision-and-Language Model. https://openreview.net/ forum?id=EhwEUb2ynIa

    2022

  16. [24]

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 2013. 3D Object Repre- sentations for Fine-Grained Categorization. In 2013 IEEE International Conference on Computer Vision Workshops. 554–561. doi:10.1109/ICCVW.2013.77

    2013 doi

  17. [25]

    Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl

    Boyi Li, Kilian Q. Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl

  18. [26]

    Zheng Li, Xiang Li, Xinyi Fu, Xin Zhang, Weiqiang Wang, Shuo Chen, and Jian Yang. 2024. Promptkd: Unsupervised prompt distillation for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 26617–26626

    2024

  19. [27]

    Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. 2019. ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks. arXiv:1908.02265 [cs.CV]

    2019 arXiv

  20. [28]

    Timo Lüddecke and Alexander Ecker. 2022. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 7086–7096

    2022

  21. [29]

    Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. 2022. Class-agnostic Object Detection with Multi-modal Transformer. arXiv:2111.11430 [cs.CV]

    2022 arXiv

  22. [30]

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi

  23. [31]

    Marco Mistretta, Alberto Baldrati, Marco Bertini, and Andrew D Bagdanov. 2024. Improving zero-shot generalization of learned prompts via unsupervised knowl- edge distillation. In European Conference on Computer Vision . Springer, 459–477

    2024

  24. [32]

    Maria-Elena Nilsback and Andrew Zisserman. 2008. Automated Flower Classifica- tion over a Large Number of Classes. In2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing . 722–729. doi:10.1109/ICVGIP.2008.47

    2008 doi

  25. [33]

    Parkhi, Andrea Vedaldi, Andrew Zisserman, and C

    Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. 2012. Cats and dogs. 2012 IEEE Conference on Computer Vision and Pattern Recognition (2012), 3498–3505. https://api.semanticscholar.org/CorpusID:383200

    2012

  26. [34]

    Mary Phuong and Christoph Lampert. 2019. Towards understanding knowledge distillation. In International conference on machine learning . PMLR, 5142–5151

    2019

  27. [35]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. arXiv:2103.000...

    2021 arXiv

  28. [36]

    Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. 2022. DenseCLIP: Language-Guided Dense Prediction with Context-Aware Prompting. arXiv:2112.01518 [cs.CV]

    2022 arXiv

  29. [37]

    Sam T Roweis and Lawrence K Saul. 2000. Nonlinear dimensionality reduction by locally linear embedding. science 290, 5500 (2000), 2323–2326

    2000

  30. [38]

    Marvin K Simon and Mohamed-Slim Alouini. 2004. Digital communication over fading channels. John Wiley & Sons

    2004

  31. [39]

    Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. 2012. UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild. arXiv:1212.0402 [cs.CV]

    2012 arXiv

  32. [40]

    Hao Tan and Mohit Bansal. 2019. LXMERT: Learning Cross-Modality Encoder Representations from Transformers. arXiv:1908.07490 [cs.CL]

    2019 arXiv

  33. [41]

    Joshua B Tenenbaum, Vin de Silva, and John C Langford. 2000. A global geometric framework for nonlinear dimensionality reduction.science 290, 5500 (2000), 2319– 2323

    2000

  34. [42]

    Xing, and Zachary C

    Haohan Wang, Songwei Ge, Eric P. Xing, and Zachary C. Lipton. 2019. Learn- ing Robust Global Representations by Penalizing Local Predictive Power. arXiv:1905.13549 [cs.CV]

    2019 arXiv

  35. [43]

    Zhengbo Wang, Jian Liang, Ran He, Nan Xu, Zilei Wang, and Tieniu Tan. 2023. Improving zero-shot generalization for clip with synthesized prompts. arXiv preprint arXiv:2307.07397 (2023)

    2023 arXiv

  36. [44]

    Ehinger, Aude Oliva, and Antonio Torralba

    Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba

  37. [45]

    Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. 2019. Deep Modular Co-Attention Networks for Visual Question Answering. arXiv:1906.10770 [cs.CV]

    2019 arXiv

  38. [46]

    Li Yuan, Francis EH Tay, Guilin Li, Tao Wang, and Jiashi Feng. 2020. Revisiting knowledge distillation via label smoothing regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 3903–3911

    2020

  39. [47]

    Sukmin Yun, Jongjin Park, Kimin Lee, and Jinwoo Shin. 2020. Regularizing class- wise predictions via self-knowledge distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 13876–13885

    2020

  40. [48]

    Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. 2022. Open-Vocabulary DETR with Conditional Matching . Springer Nature Switzerland, 106–122

    2022

  41. [49]

    Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexan- der Kolesnikov, and Lucas Beyer. 2022. Lit: Zero-shot transfer with locked-image ���� ������� �������� ������� ��� ������ ��������������� ������������ ���������� ���� ���� �������� ������ ����� ��...

    2022

  42. [50]

    Renrui Zhang, Rongyao Fang, Wei Zhang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. 2021. Tip-Adapter: Training-free CLIP-Adapter for Better Vision-Language Modeling. arXiv:2111.03930 [cs.CV]

    2021 arXiv

  43. [51]

    Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. 2018. Deep mutual learning. In Proceedings of the IEEE conference on computer vision and pattern recognition. 4320–4328

    2018

  44. [52]

    Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. 2022. Decoupled knowledge distillation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition . 11953–11962

    2022

  45. [53]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Conditional Prompt Learning for Vision-Language Models. arXiv:2203.05557 [cs.CV]

    2022 arXiv

  46. [54]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Learning to Prompt for Vision-Language Models. International Journal of Computer Vision 130, 9 (July 2022), 2337–2348. doi:10.1007/s11263-022-01653-1 Received 18 May 2025; revised 18 May 2025; accepted 18 May 2025

    2022 doi

  47. [2010]

    In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition

    SUN database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition . 3485–3492. doi:10.1109/CVPR.2010.5539970

    2010

  48. [2013]

    arXiv:1306.5151 [cs.CV]

    Fine-Grained Visual Classification of Aircraft. arXiv:1306.5151 [cs.CV]

    arXiv

  49. [2021]

    arXiv:2106.01401 [cs.CV]

    Container: Context Aggregation Network. arXiv:2106.01401 [cs.CV]

    arXiv

  50. [2022]

    arXiv:2201.03546 [cs.CV]

    Language-driven Semantic Segmentation. arXiv:2201.03546 [cs.CV]

    arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.