Pith. sign in

REVIEW 5 major objections 4 minor 57 references

Interpretable Zero-Shot Learning with Locally-Aligned Vision-Language Model

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper claims that zero-shot classification improves and becomes interpretable when local image regions are aligned with LLM-generated attributes through an optimal transport plan, with no extra training.

desk verdict LaZSL is a plausible training-free recipe for interpretable ZSL, but the paper under-specifies the OT solver and never shows the transport plan is actually non-uniform, leaving open that the gains are just multi-crop ensembling. read the letter →

arxiv 2506.23822 v1 pith:QUUU6NTV submitted 2025-06-30 cs.CV

classification cs.CV
keywords zero-shotlearninginterpretabilityvision-languagemodelsoptimaltransportlocalvisual-semanticalignmentCLIPlargelanguagedomaingeneralization
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 proposes LaZSL, a training-free way to make vision-language models like CLIP both more accurate and more interpretable in zero-shot classification. Instead of comparing a whole image to class names or to global attribute descriptions, LaZSL crops the image into many regions, generates a set of visual attributes per class with a large language model, and uses optimal transport to match each region to the attributes it most plausibly supports. The match is combined with the global image-text similarity for the final class score. The paper shows on nine datasets that this locally aligned scoring beats prior interpretable baselines and improves robustness to domain shift. If correct, LaZSL implies that fine-grained visual-semantic correspondence, not just holistic similarity, is a practical source of zero-shot performance without additional training.

What carries the argument

The load-bearing mechanism is the optimal transport plan $T$ computed with the Sinkhorn algorithm between two discrete distributions: the visual set formed by random multi-scale crops of the query image and the semantic set formed by LLM-generated attributes of a candidate class. Three design choices carry the argument: a vision-selection step that zeroes out the probability mass of crops whose cosine similarity to the global image embedding falls below the average (Eqs. 6-9), a hybrid cost matrix that linearly blends region-attribute similarity with global image-attribute similarity (Eq. 10), and the Frobenius-inner-product scoring of the final plan against the same hybrid similarity (Eqs. 15-16). These components together let the model align fine-grained visual evidence with specific attributes while retaining the global context of the frozen CLIP encoder.

What would settle it

A concrete test: replace the random-crop visual set with fixed grid patches, or with crops taken from a region known to contain the attribute, and re-run LaZSL. If accuracy does not drop when the crops are decorrelated from attribute locations, then the OT alignment is not actually using local evidence; alternatively, if randomly shuffling the rows of the cost matrix destroys the gain, the specific alignment is doing the work.

Watch

Extended reading notes

Core claim

The paper's central claim is that local visual-semantic alignment, realized as an optimal transport plan between a set of random multi-scale image crops and a set of LLM-generated attribute descriptions, yields better zero-shot classification than global image-text matching while also making predictions explainable. For each class, LaZSL builds a semantic set of attributes and a visual set of cropped regions, filters out regions that look unrelated to the global image, and solves a Sinkhorn-regularized optimal transport problem whose cost matrix mixes region-attribute similarity with global image-attribute similarity. The resulting transport plan provides both a per-region-to-attribute correspondence, which serves as the interpretable evidence, and a category score via a Frobenius inner product with the hybrid similarity. The paper reports consistent gains over DCLIP, CuPL, and WaffleCLIP across three CLIP backbones, and improved average accuracy on ImageNet domain-shift benchmarks, including a large gain on ImageNet-A.

Load-bearing premise

The method assumes that random multi-scale crops, after being filtered by similarity to the whole image, contain the local visual evidence described by the LLM-generated attributes, so that the optimal transport plan reflects a true visual-semantic correspondence rather than noise.

Editorial extensions

If this is right

  • LaZSL should allow CLIP-based zero-shot classifiers to show which attribute of a predicted class is supported by which image region, giving a per-prediction explanation without any training.
  • Because the method applies to any frozen VLM with patch-level visual features and an attribute generator, interpretability gains can transfer to newer backbones without retraining.
  • The domain-generalization improvement, especially on ImageNet-A, suggests that attribute-region alignment is more robust to natural distribution shift than global prompt similarity.
  • Since no training is involved, LaZSL can be deployed on new classes simply by generating attributes with an LLM, keeping the same cost profile as description-based baselines plus a small optimal-transport overhead.

Reading between the lines

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

  • The paper does not directly verify that the regions selected by the optimal transport plan actually contain the object parts named by the attributes; a dedicated attribution study comparing OT-selected regions with human-annotated part locations would test whether the interpretability is causal or merely correlational.
  • The Sinkhorn regularization and random-crop sampling introduce stochasticity; averaging over multiple crops or seeds could further stabilize accuracy and offers a low-cost variance-reduction extension.
  • The hybrid coefficient $\theta = 0.8$ is fixed across all datasets; an adaptive $\theta$ per dataset or per class might yield further gains, especially on fine-grained datasets where local information dominates.
  • The method's dependence on LLM attribute quality is acknowledged as a limitation; a filtering mechanism that scores attributes by their transport contribution could remove noisy descriptions and potentially improve both accuracy and interpretability.
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

5 major / 4 minor

Summary. The paper proposes LaZSL, a training-free interpretable zero-shot learning method built on CLIP. It constructs a semantic set of LLM-generated attributes per class and a visual set of random multi-scale crops per image, filters crops by similarity to the global image embedding, computes a Sinkhorn optimal-transport plan between filtered local visual features and attribute features, and scores each class with a hybrid local-global Frobenius inner product. The authors report accuracy gains over DCLIP, WaffleCLIP, CuPL, and several prompt-learning baselines on five classification benchmarks and four ImageNet variants, plus qualitative examples intended to show attribute-based interpretability.

Significance. If the reported results are taken at face value, the paper offers a useful training-free way to obtain attribute-level explanations from CLIP while improving average accuracy, and it ships code and a relatively broad evaluation across nine datasets. The core derivation is straightforward and the component ablations are directionally consistent. However, the evaluation currently has load-bearing weaknesses: hyperparameters are selected on the same test benchmarks used in the headline tables, the entropic OT parameters are never reported, the ablations do not isolate local semantic alignment from multi-crop ensembling, and there is no direct verification that selected crops contain the evidence described by the attributes. These issues prevent the paper from fully establishing its central claim that locally-aligned OT, rather than ensembling or test-set tuning, drives the improvements.

major comments (5)
  1. [Section 4.5, Tables 1–2] Hyperparameter selection is performed on the same benchmarks used for the headline results. Figure 5 chooses α and θ by evaluating accuracy on CUB and ImageNet, and those same two datasets appear in Tables 1 and 2 with the final accuracy reported. This means the gains over DCLIP (e.g., +3.8 on CUB in Table 1) are partly a function of test-set tuning, so the central accuracy claim is not evaluated cleanly. Please fix α, θ, N, and β to predetermined values, tune on held-out validation splits, or explicitly state which datasets were used for tuning; if tuning was done on the test benchmarks, the comparisons should be re-run with settings selected without access to those test labels.
  2. [Section 3.2, Eq. (14)] The entropic regularization λ and the number of Sinkhorn iterations k are never reported. Equation (14) defines M = exp(−C*/λ), so if λ is large the OT plan T collapses toward the product of the uniform marginals and Eq. (15) becomes a multi-crop average of attribute similarities. In that regime the method is indistinguishable from random-crop ensembling with global hybrid features. The qualitative evidence in Fig. 3 is consistent with this worry: the displayed attribute scores are nearly identical (about 0.067). Please report λ and k, measure the sharpness of T (e.g., mean maximum entry or the effective number of matches), and add an ablation that replaces the OT plan with simple averaging over the same filtered crops.
  3. [Table 3, Ablation Study] The ablation study only adds components cumulatively relative to DCLIP, so it does not establish the causal role of local visual-semantic alignment. There is no control that keeps the OT pipeline but replaces LLM attributes with class-name embeddings, and no control that replaces OT with average pooling over the same filtered crops. Without these controls, the gains in Table 3 could come from multi-crop ensembling or the hybrid global term rather than from aligning local regions with attributes. Please add the missing controls, including a uniform-plan variant of Eq. (15).
  4. [Section 3.1, Eqs. (6)–(8)] The method assumes that random multi-scale crops, after filtering by similarity to the global image embedding, contain the local evidence described by the LLM attributes. This assumption is not directly verified. Please provide quantitative evidence that selected regions correspond to the attributes used for scoring, for example by measuring region-attribute retrieval precision on a subset with part annotations or by visualizing which crops receive high OT mass for a given attribute. Without such verification, the OT plan may be adding noise rather than signal, which would undermine the interpretability and alignment claims.
  5. [Tables 1–2] Tables 1 and 2 report single accuracy numbers with no error bars or repeated trials, while several claimed improvements are small (e.g., +0.2 and +0.5 points on Food101 and Oxford Pets in Table 1) and the largest domain-generalization gain is concentrated in ImageNet-A (+6.0 over CuPL in Table 2). Please report mean ± std over at least three random-crop or random-seed runs. In addition, the provenance of the SaLS baseline in Table 2 should be clarified: reference [32] in the bibliography is a duplicate of reference [31] and does not correspond to the SaLS method cited in the table.
minor comments (4)
  1. [Section 3.2, Eqs. (10) and (16)] The notation for C*_i and sim*_i mixes vector and matrix expressions; please clarify the dimensions and broadcasting of the global term pg⊤Qy so the hybrid cost and similarity are unambiguous.
  2. [Figure 1] There are typos in the figure: 'LAZSL' should be 'LaZSL' and 'bely' should be 'belly'.
  3. [Table 2, footnote] The footnote describing the training and attribute markers is incomplete: the sentence 'Methods take the source data for training marked with' cuts off, and the meaning of the quoted and percent symbols should be stated explicitly.
  4. [Section 5, Limitations] The Limitations paragraph correctly notes dependence on LLM-generated attribute quality, but the paper does not quantify robustness to attribute noise. Please add an analysis that perturbs or subsamples attributes to show how sensitive the reported accuracy and interpretability are to attribute quality.

Circularity Check

1 steps flagged · score 4.0 of 10

Core LaZSL scoring is not circular by construction, but the headline results are partly tuned because α and θ are selected on the same CUB and ImageNet test benchmarks in Sec. 4.5.

  1. fitted input called prediction [Section 4.5 (Hyper-Parameter Analysis), Fig. 5; implementation setting θ=0.8 and α=0.6; reported in Tables 1-2.]
    "As can be seen from Fig. 5(a)(b), LaZSL is not sensitive to α, and it achieves best performances on all datasets when α is set to 0.6... Accordingly, we experimentally set the α = 0.6 for all datasets. ... Based on our experiments, we set θ = 0.8 to fuse key local visual features with a few global features in LaZSL."

    The method's two free hyperparameters are chosen by maximizing accuracy on the same CUB and ImageNet test sets used to report its headline results (Tables 1-2). The phrase 'achieves best performances' refers to the selection criterion itself, so the reported accuracy is not an independent evaluation of LaZSL: it is tuned to those benchmarks. This is a partial fit-to-evaluation rather than a collapse of Eq. 15 into the baseline, so it raises the circularity score without making the whole derivation circular.

full rationale

The central scoring chain is internally derived, not defined in terms of its own output: attributes are generated externally by GPT-3 following DCLIP (Sec. 3.1), visual regions are random crops (Eq. 2), and the category score ψ_y = <T, sim*>_F (Eq. 15) is a Frobenius inner product of an OT plan with a hybrid similarity matrix; none of these quantities by construction equals the ground-truth label or the DCLIP baseline. The OT machinery and vision selection are therefore not circular in the self-definitional sense. The main circularity concern is test-set hyperparameter selection in Sec. 4.5, where α=0.6 and θ=0.8 are chosen by looking at accuracy on CUB and ImageNet and then the same numbers are reported as results. That is a fitted-input issue and warrants a moderate score. Two further concerns are verification gaps rather than circularity: the Sinkhorn entropy λ and iteration count k in Eqs. 11-14 are never specified, so the sharpness of T is uncontrolled, and Fig. 3 shows nearly flat attribute scores (about 0.067), consistent with a dispersive plan that degenerates toward multi-crop averaging. These gaps weaken the causal claim that OT alignment drives the gains, but they do not make the derivation equivalent to its inputs. The paper's own Limitations section acknowledges dependence on LLM attribute quality but does not introduce a circular step.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a mix of standard OT math and two strong domain assumptions: that random crops capture attribute evidence and that LLM attributes are reliable. The paper also introduces several hyperparameters whose values are either chosen on the test data or left unreported, which limits the strength of the experimental support.

free parameters (6)
  • α (crop scale lower bound) = 0.6
    Selected via hyperparameter analysis on CUB and ImageNet test sets (Fig. 5a-b, Sec. 4.5); controls the minimum size of random crops and affects all downstream alignment.
  • β (crop scale upper bound) = not reported
    The paper defines γ_i ~ U(α, β) in Eq. (2) but never reports β; it only says 'cropping scale to 0.6'.
  • θ (hybrid coefficient) = 0.8
    Selected via hyperparameter analysis on CUB and ImageNet test sets (Fig. 5c-d, Sec. 4.5); blends global and local similarity in cost and scoring (Eqs. 10, 16).
  • N (number of crops) = 60-90
    Randomly chosen per image in [60, 90] (Sec. 4.1); the exact value is stochastic and not specified further.
  • λ (Sinkhorn entropic regularization) = not reported
    Used in Eq. (14) M = exp(-C*/λ), but the value of λ is never given.
  • Sinkhorn iterations k = not reported
    Iteration count in Eqs. (12)-(13) is not specified; affects approximation accuracy.
assumptions (5)
  • domain assumption CLIP embeddings of image crops and attribute text are comparable in cosine space
    The entire cost matrix (Eq. 6, Sec. 3.2) rests on cosine similarity between region features and attribute text embeddings; if these spaces are not well-aligned, OT will match unrelated items.
  • domain assumption LLM-generated attributes are relevant and descriptive for each class
    The semantic set Sy = h(prompt(y)) in Eq. (1) is taken from GPT-3; the paper's own Limitations section acknowledges that the LLM 'may occasionally produce irrelevant attribute descriptions'.
  • domain assumption Random multi-scale crops filtered by global-similarity threshold contain the attribute evidence
    Vision set construction (Eq. 2) and vision selection (Eqs. 6-9) assume that crops similar to the global image are the relevant ones; this is not tested against attribute localization.
  • standard math Sinkhorn algorithm yields a stable and meaningful OT plan
    Equations (11)-(14) use Sinkhorn's entropic OT; its convergence and the effect of λ are taken as standard.
  • ad hoc to paper Uniform marginal on attributes is a sensible prior
    The semantic probability vector c is initialized as uniform (Sec. 3.2), weighting every attribute equally; no justification is given for why attributes should be equally important.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interpretable Zero-Shot Learning with Locally-Aligned Vision-Language Model." pith.science (2026). https://pith.science/paper/QUUU6NTV

@misc{pith2026250623822,
  author       = {Pith},
  title        = {Pith review of: Interpretable Zero-Shot Learning with Locally-Aligned Vision-Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QUUU6NTV}},
  note         = {Machine review of arXiv:2506.23822}
}
read the original abstract

Large-scale vision-language models (VLMs), such as CLIP, have achieved remarkable success in zero-shot learning (ZSL) by leveraging large-scale visual-text pair datasets. However, these methods often lack interpretability, as they compute the similarity between an entire query image and the embedded category words, making it difficult to explain their predictions. One approach to address this issue is to develop interpretable models by integrating language, where classifiers are built using discrete attributes, similar to human perception. This introduces a new challenge: how to effectively align local visual features with corresponding attributes based on pre-trained VLMs. To tackle this, we propose LaZSL, a locally-aligned vision-language model for interpretable ZSL. LaZSL employs local visual-semantic alignment via optimal transport to perform interaction between visual regions and their associated attributes, facilitating effective alignment and providing interpretable similarity without the need for additional training. Extensive experiments demonstrate that our method offers several advantages, including enhanced interpretability, improved accuracy, and strong domain generalization. Codes available at: https://github.com/shiming-chen/LaZSL.

Figures

Figures reproduced from arXiv: 2506.23822 by the authors.

Figure 1
Figure 1. Comparison between the similarity of CLIP [ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The pipeline of our LaZSL. LaZSL consists of three main module components, including the construction of semantic and vision [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of interpretable classification of LaZSL. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Classification comparison between LaZSL and DCLIP [ [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Hyper-parameters analysis. We show the ZSL performance variations on CUB and ImageNet by adjusting the value of cropping [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 55 canonical work pages

  1. [32]

    Robust calibration of large vision- language adapters

    Balamurali Murugesan, Julio Silva-Rodr ´ıguez, Ismail Ben Ayed, and Jose Dolz. Robust calibration of large vision- language adapters. In ECCV, pages 147–165, 2024. 7

  2. [31]

    Robust calibration of large vision- language adapters

    Balamurali Murugesan, Julio Silva-Rodr ´ıguez, Ismail Ben Ayed, and Jos ´e Dolz. Robust calibration of large vision- language adapters. In ECCV, 2024. 1, 3

  3. [1]

    Label-embedding for image classification

    Zeynep Akata, Florent Perronnin, Za ¨ıd Harchaoui, and Cordelia Schmid. Label-embedding for image classification. IEEE Trans. Pattern Anal. Mach. Intell. , 38(7):1425–1438,

  4. [2]

    Wasserstein generative adversarial networks

    Mart ´ın Arjovsky, Soumith Chintala, and L ´eon Bottou. Wasserstein generative adversarial networks. InICML, pages 214–223, 2017. 3

  5. [3]

    Food-101 - mining discriminative components with random forests

    Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, pages 446–461, 2014. 5

  6. [4]

    PLOT: prompt learning with optimal transport for vision-language models

    Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. PLOT: prompt learning with optimal transport for vision-language models. In ICLR,

  7. [5]

    Hsva: Hi- erarchical semantic-visual adaptation for zero-shot learning

    Shiming Chen, Guo-Sen Xie, Yang Yang Liu, Qinmu Peng, Baigui Sun, Hao Li, Xinge You, and Ling Shao. Hsva: Hi- erarchical semantic-visual adaptation for zero-shot learning. In NeurIPS, 2021. 2

  8. [6]

    MSDN: mutually semantic distillation network for zero-shot learn- ing

    Shiming Chen, Ziming Hong, Guo-Sen Xie, Wenhan Yang, Qinmu Peng, Kai Wang, Jian Zhao, and Xinge You. MSDN: mutually semantic distillation network for zero-shot learn- ing. In CVPR, pages 7602–7611, 2022. 2

Show all 57 references
  1. [7]

    Transzero++: Cross attribute-guided transformer for zero-shot learning

    Shiming Chen, Ziming Hong, Wenjin Hou, Guo-Sen Xie, Yibing Song, Jian Zhao, Xinge You, Shuicheng Yan, and Ling Shao. Transzero++: Cross attribute-guided transformer for zero-shot learning. IEEE Trans. Pattern Anal. Mach. In- tell., 45(11):12844–12861, 2023. 2

  2. [8]

    Evolving semantic prototype improves generative zero-shot learning

    Shiming Chen, Wen Qing Hou, Ziming Hong, Xiaohan Ding, Yibing Song, Xinge You, Tongliang Liu, and Kun Zhang. Evolving semantic prototype improves generative zero-shot learning. In ICML, 2023. 2

  3. [9]

    Khan, and Fa- had Shahbaz Khan

    Shiming Chen, Wenjin Hou, Salman H. Khan, and Fa- had Shahbaz Khan. Progressive semantic-guided vision transformer for zero-shot learning. In CVPR, pages 23964– 23974, 2024. 2

  4. [10]

    Khan, and Fa- had Shahbaz Khan

    Shiming Chen, Dingjie Fu, Salman H. Khan, and Fa- had Shahbaz Khan. Genzsl: Generative zero-shot learning via inductive variational autoencoder. 2025. 2

  5. [11]

    Semantics-conditioned generative zero-shot learning via fea- ture refinement

    Shiming Chen, Ziming Hong, Xinge You, and Ling Shao. Semantics-conditioned generative zero-shot learning via fea- ture refinement. International Journal of Computer Vision ,

  6. [12]

    Evolving interpretable visual classifiers with large language models

    Mia Chiquier, Utkarsh Mall, and Carl V ondrick. Evolving interpretable visual classifiers with large language models. In ECCV, 2024. 1, 3

  7. [13]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS, pages 2292–2300, 2013. 5

  8. [14]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 5

  9. [15]

    Image2sentence based asymmetrical zero-shot composed image retrieval

    Yongchao Du, Min Wang, Wen gang Zhou, Shuping Hui, and Houqiang Li. Image2sentence based asymmetrical zero-shot composed image retrieval. In ICLR, 2024. 1

  10. [16]

    Improving clip training with language rewrites

    Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. Improving clip training with language rewrites. In NeurIPS, 2023. 1, 3

  11. [17]

    Diverse data augmentation with diffusions for effective test-time prompt tuning

    Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In ICCV, pages 2704– 2714, 2023. 1

  12. [18]

    Clip-adapter: Better vision-language models with feature adapters

    Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Jiao Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision , 2023. 1

  13. [19]

    The many faces of robustness: A criti- cal analysis of out-of-distribution generalization

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kada- vath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A criti- cal analysis of out-of-distribution generalization. I...

  14. [20]

    Natural adversarial examples

    Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Stein- hardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262–15271, 2021. 5

  15. [21]

    Fine-grained generalized zero-shot learning via dense attribute-based attention

    Dat Huynh and Ehsan Elhamifar. Fine-grained generalized zero-shot learning via dense attribute-based attention. In CVPR, pages 4482–4492, 2020. 2

  16. [22]

    Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V . Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904– 4916, 2021. 2

  17. [23]

    Multi- modal classifiers for open-vocabulary object detection

    Prannay Kaul, Weidi Xie, and Andrew Zisserman. Multi- modal classifiers for open-vocabulary object detection. In ICML, 2023. 1

  18. [24]

    Khan, and Fahad Shahbaz Khan

    Muhammad Uzair Khattak, Hanoona Abdul Rasheed, Muhammad Maaz, Salman H. Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. InCVPR, pages 19113–19122, 2023. 1, 3, 5, 6, 7

  19. [25]

    Kolkin, Jason Salavon, and Gregory Shakhnarovich

    Nicholas I. Kolkin, Jason Salavon, and Gregory Shakhnarovich. Style transfer by relaxed optimal transport and self-similarity. In CVPR, pages 10051–10060, 2019. 3

  20. [26]

    Co-clustering through optimal transport

    Charlotte Laclau, Ievgen Redko, Basarab Matei, Youn `es Bennani, and Vincent Brault. Co-clustering through optimal transport. In ICML, pages 1955–1964, 2017. 3

  21. [27]

    Visual-text cross alignment: Refining the similarity score in vision-language models

    Jinhao Li, Haopeng Li, Sarah Monazam Erfani, Lei Feng, James Bailey, and Feng Liu. Visual-text cross alignment: Refining the similarity score in vision-language models. In ICML, 2024. 3

  22. [28]

    Patchct: Align- ing patch set and label set with conditional transport for multi-label image classification

    Miaoge Li, Dongsheng Wang, Xinyang Liu, Zequn Zeng, Ruiying Lu, Bo Chen, and Mingyuan Zhou. Patchct: Align- ing patch set and label set with conditional transport for multi-label image classification. In ICCV, pages 15302– 15312, 2023. 2, 3

  23. [29]

    Progressive semantic-visual mutual adaption for generalized zero-shot learning

    Man Liu, Feng Li, Chunjie Zhang, Yunchao Wei, Huihui Bai, and Yao Zhao. Progressive semantic-visual mutual adaption for generalized zero-shot learning. In CVPR, pages 15337– 15346, 2023. 2

  24. [30]

    Visual classification via description from large language models

    Sachit Menon and Carl V ondrick. Visual classification via description from large language models. In ICLR, 2023. 1, 2, 3, 4, 5, 6, 7, 8, 9

  25. [33]

    I2dformer+: Learning image to document summary attention for zero-shot image classifi- cation

    Muhammad Ferjad Naeem, Yongqin Xian, Luc Van Gool, and Federico Tombari. I2dformer+: Learning image to document summary attention for zero-shot image classifi- cation. Internal Journal of Computer Vision , 132(9):3806– 3822, 2024. 2

  26. [34]

    Pomerleau, Geoffrey E

    Mark Palatucci, D. Pomerleau, Geoffrey E. Hinton, and Tom Michael Mitchell. Zero-shot learning with semantic output codes. In NeurIPS, pages 1410–1418, 2009. 1, 2

  27. [35]

    Parkhi, Andrea Vedaldi, Andrew Zisserman, and C

    Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V . Jawahar. Cats and dogs. In CVPR, pages 3498–3505,

  28. [36]

    Computational optimal transport

    Gabriel Peyr ´e and Marco Cuturi. Computational optimal transport. Found. Trends Mach. Learn. , 11(5-6):355–607,

  29. [37]

    What does a platypus look like? generating customized prompts for zero- shot image classification

    Sarah Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero- shot image classification. In ICCV, pages 15645–15655,

  30. [38]

    Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi

    Sarah M. Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In ICCV, pages 15645–15655, 2023. 7, 9

  31. [39]

    Proapo: Pro- gressively automatic prompt optimization for visual classifi- cation

    Xiangyan Qu, Gaopeng Gou, Jiamin Zhuang, Jing Yu, Kun Song, Qihao Wang, Yili Li, and Gang Xiong. Proapo: Pro- gressively automatic prompt optimization for visual classifi- cation. In CVPR, 2025. 6

  32. [40]

    Learning transferable visual models from natural language supervision

    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. Learning transferable visual models from natural language supervision. In ICML, pages 8748–...

  33. [41]

    Do imagenet classifiers generalize to im- agenet? In ICML, pages 5389–5400, 2019

    Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to im- agenet? In ICML, pages 5389–5400, 2019. 5

  34. [42]

    Sophia Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata

    Karsten Roth, Jae-Myung Kim, A. Sophia Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, pages 15700–15711,

  35. [43]

    Generalized zero- and few- shot learning via aligned variational autoencoders

    Edgar Sch ¨onfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. Generalized zero- and few- shot learning via aligned variational autoencoders. In CVPR, pages 8247–8255, 2019. 2

  36. [44]

    Test- time prompt tuning for zero-shot generalization in vision- language models

    Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test- time prompt tuning for zero-shot generalization in vision- language models. In NeurIPS, 2022. 1, 5, 6, 7

  37. [45]

    Hunting attributes: Con- text prototype-aware learning for weakly supervised seman- tic segmentation

    Feilong Tang, Zhongxing Xu, Zhaojun Qu, Wei Feng, Xingjian Jiang, and Zongyuan Ge. Hunting attributes: Con- text prototype-aware learning for weakly supervised seman- tic segmentation. In CVPR, pages 3324–3334, 2024. 1

  38. [46]

    Argue: Attribute-guided prompt tuning for vision-language models

    Xinyu Tian, Shu Zou, Zhaoyuan Yang, and Jing Zhang. Argue: Attribute-guided prompt tuning for vision-language models. In CVPR, pages 28578–28587, 2024. 1, 3, 5, 6, 7

  39. [47]

    Tuning multi-mode token- level prompt alignment across modalities

    Dongsheng Wang, Miaoge Li, Xinyang Liu, Mingsheng Xu, Bo Chen, and Hanwang Zhang. Tuning multi-mode token- level prompt alignment across modalities. In NeurIPS, 2023. 2, 3

  40. [48]

    Lipton, and Eric P

    Haohan Wang, Songwei Ge, Zachary C. Lipton, and Eric P. Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, pages 10506–10518,

  41. [49]

    Zero-shot visual recognition via bidirectional latent embedding

    Qian Wang and Ke Chen. Zero-shot visual recognition via bidirectional latent embedding. Internal Journal of Com- pution Vision, 124(3):356–383, 2017. 2

  42. [50]

    Welinder, S

    P. Welinder, S. Branson, T. Mita, C. Wah, Florian Schroff, Serge J. Belongie, and P. Perona. Caltech-ucsd birds 200. Technical Report CNS-TR-2010-001, Caltech,, 2010. 5

  43. [51]

    Schiele, and Zeynep Akata

    Yongqin Xian, B. Schiele, and Zeynep Akata. Zero-shot learning — the good, the bad and the ugly. CVPR, pages 3077–3086, 2017. 1, 2

  44. [52]

    Lorenz, B

    Yongqin Xian, T. Lorenz, B. Schiele, and Zeynep Akata. Feature generating networks for zero-shot learning. In CVPR, pages 5542–5551, 2018. 2

  45. [53]

    Khan, Zhiqiang Shen, Muzammal Naseer, Guangyi Chen, and Fahad Shahbaz Khan

    Sheng Zhang, Salman H. Khan, Zhiqiang Shen, Muzammal Naseer, Guangyi Chen, and Fahad Shahbaz Khan. Prompt- cal: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery. InCVPR, pages 3479– 3488, 2023. 3

  46. [54]

    Places: A 10 million image database for scene recognition

    Bolei Zhou, `Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Trans. Pattern Anal. Mach. In- tell., 40(6):1452–1464, 2018. 5

  47. [55]

    Conditional prompt learning for vision-language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zi- wei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16795–16804, 2022. 1, 3, 5, 6, 7

  48. [56]

    Learning to prompt for vision-language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. Int. J. Comput. Vis., 130(9):2337–2348, 2022. 1, 5, 6, 7

  49. [57]

    Prompt-aligned gradient for prompt tuning

    Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. InICCV, pages 15613–15623, 2023. 5, 6, 7

Pith tools

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