Spotlighter: Revisiting Prompt Tuning from a Representative Mining View

Chuang Zhu; Guoshun Nan; Lijuan Sun; Maoyuan Shao; Xinyang Huang; Xuan Liu; Yutong Gao; Yu Weng

REVIEW 4 major objections 6 minor 12 references

Spotlighter: Revisiting Prompt Tuning from a Representative Mining View

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Spotlighter claims that keeping only the most activated visual tokens, refined by learned class prototypes, improves both accuracy and speed in CLIP few-shot classification—with just 21 extra parameters.

desk verdict The token-selection idea is reasonable but the paper's central efficiency claim ('21 extra parameters', impossible FPS) is contradicted by its own architecture and protocol, so the paper needs major correction before it can be taken seriously. read the letter →

arxiv 2509.00905 v2 pith:A4ZK3QM5 submitted 2025-08-31 cs.CV cs.AI

Yutong Gao , Maoyuan Shao , Xinyang Huang , Chuang Zhu , Lijuan Sun , Yu Weng , Xuan Liu , Guoshun Nan This is my paper

classification cs.CVcs.AI

keywords prompttuningfew-shotclassificationCLIPtokenselectionsemanticprototypesrepresentativefeatureminingefficientinferencevision-languagemodels

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

Spotlighter aims to show that prompt tuning for CLIP can be improved by treating most visual tokens as noise. The paper proposes scoring every image token twice—once by its alignment with the text embedding, once by its match to class prototypes in a semantic memory bank—and keeping only the top-scoring tokens for prediction. A two-level ranking then fuses those tokens with the prototypes through tiny mapping modules, so a compact token set carries the classification signal. On 11 few-shot benchmarks the method is claimed to beat CLIP by up to 11.19% in harmonic mean accuracy and to add up to 0.8K frames per second, using only 21 trainable parameters, while also lifting existing methods such as CoOp, MaPLe, CLIPFit, and PromptKD when plugged in. If correct, this would make token-level selection a near-free accuracy and efficiency lever for vision-language prompt tuning.

What carries the argument

The load-bearing mechanism is the activation score computed per visual token: a sample-wise term from image-text similarity plus a semantic-wise term from matching the token against class prototypes in a Semantic Memory Bank. A fixed top-k by summed score selects the only tokens that survive to prediction. The Semantic Memory Bank, seeded from class-name text embeddings and updated by momentum, supplies semantic context that compensates for discarded tokens. A two-level ranking stratifies selected tokens into two tiers, and small Image and Text Representative Mapping modules fuse each tier with its prototype, so a compact, semantically weighted token set replaces the full feature map.

What would settle it

Take a fine-grained dataset with heavily entangled classes, train Spotlighter with the paper's protocol, and sweep k from small to full retention: the central claim fails if keeping all tokens beats the top-k selection, or if a control with the same losses but randomly chosen tokens matches Spotlighter's accuracy. A separate end-to-end timing measurement that includes the IRM/TRM modules would settle whether the claimed 0.8K FPS gain is real.

Watch

Extended reading notes

Core claim

At the core, Spotlighter claims that in CLIP-based few-shot image classification, most visual tokens are redundant or noisy, and that selective pruning guided by semantics improves both accuracy and speed. Each visual token receives an activation score combining sample-wise similarity to the text feature and semantic-wise similarity to class prototypes stored in a Semantic Memory Bank. Only the top-k tokens are kept; prototypes are updated by momentum and used to compensate for discarded information. The selected tokens are split into two tiers and fused with their matched prototypes through Image and Text Representative Mapping modules, and only these representative tokens are used for the

Load-bearing premise

The method's accuracy gain depends on the activation score reliably ranking the truly discriminative visual tokens at the top, so that discarding the rest removes noise rather than signal; the authors themselves note this fails in ultra-fine-grained or highly entangled classes.

Editorial extensions

If this is right

Spotlighter can be attached to existing prompt-tuning and adapter methods, improving both their seen-class and unseen-class accuracy while keeping inference fast.
With only 21 trainable parameters, the method makes token-level pruning a practical accuracy-efficiency lever for CLIP, not a costly architectural change.
Because prediction uses a compact set of representative tokens, the method reduces the scale of cross-modal feature interactions at test time, which is the source of the claimed FPS gain.
The semantic memory bank can compensate for discarded visual regions by injecting prototype knowledge, which is why aggressive pruning does not simply throw away semantics.
The method's gains persist across base, novel, few-shot, and cross-dataset settings, so the selection mechanism generalizes beyond the training classes.

Reading between the lines

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

My inference: this activation-ranking recipe should be portable to other contrastive vision-language models, since token redundancy is a general property of patch-token encoders; a quick check on a second model family would show whether the gains are CLIP-specific.
My inference: because prototypes are seeded from class-name text embeddings, the semantic-wise score inherits the quality of CLIP's text descriptions; enriching that seed with attributes or definitions could raise the ceiling on fine-grained classes.
My inference: the token count k is fixed; the paper's own limitation on entangled classes suggests an adaptive k, chosen per sample or per class from the activation distribution, is the next natural variable to test.
My inference: to show the accuracy gain is really due to selection rather than to the extra losses, an equal-cost control that trains the same losses on randomly chosen tokens would isolate the ranking mechanism; the paper's bottom-k experiments partially cover this, but not with an equal training budget.

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.

Desk Editor's Note

read the letter

The core efficiency claim is not credible. The paper says “only 21 extra parameters” while §3.2 describes multi-head attention, an FFN, a linear layer, and a trainable memory bank. Appendix C says each category keeps five 512-dim prototypes, which alone is over 2.5M parameters for ImageNet. That is a load-bearing contradiction, not a typo. The FPS numbers are also physically implausible: 131K FPS on ImageNet with ViT-B/16 on a single 4090 is not achievable, and the pruning happens after the full CLIP encoder, so the described mechanism cannot speed up the backbone forward pass. The stress-test note is correct on both counts.

What the paper does well: the idea of scoring visual tokens from both sample-wise and semantic-wise perspectives, with a prototype memory bank and two-level ranking, is a sensible extension of prompt tuning. The paper runs a broad set of experiments across 11 benchmarks and includes ablations for losses, prototype initialization, and token counts. That is genuinely useful work for the few-shot VL subfield.

The soft spots are serious and proportionate. The accuracy improvements are weakened by Appendix C's unequal retraining: CoOp gets 15 epochs, CLIPFit and PromptKD get 30, while Spotlighter presumably trains longer. No error bars appear in the tables despite the claim of three runs. The paper's own Limitations section concedes performance degrades in ultra-fine-grained settings, which is exactly where the token-selection assumption is most fragile. None of this is fatal to the idea, but it is fatal to the paper as written.

Who gets value: someone working on prompt tuning might want to read this to extract the token-selection and prototype-memory idea and to see what ablations look like. But no one should cite the efficiency numbers or parameter count.

Recommendation: this deserves a serious referee only if the authors are required to correct the parameter count and provide plausible, reproducible efficiency measurements. As is, I would not accept it, but I would not desk-reject it either—it needs a major revision. Send it to review with the expectation that the efficiency claims be withdrawn or properly substantiated.

Referee Report

4 major / 6 minor

Summary. The paper proposes Spotlighter, a plug-and-play token-selection module for CLIP few-shot classification. Spotlighter computes per-token activation scores from sample-wise and semantic-wise similarities, selects top-k tokens, refines them using a learned class-prototype semantic memory bank and IRM/TRM modules, and trains with multiple auxiliary losses. The headline claims are that it outperforms CLIP by up to 11.19% in harmonic mean accuracy, gains up to 0.8K FPS, and adds only 21 extra parameters. The paper reports experiments on 11 few-shot benchmarks, base-to-novel generalization, cross-dataset transfer, ablations, and efficiency comparisons.

Significance. If the efficiency and parameter-count claims were correct, Spotlighter could be a useful lightweight baseline for prompt tuning. The central conceptual idea, selecting representative tokens after CLIP encoding, is not inherently circular. However, the paper's own equations and appendix contradict the '21 extra parameters' claim, and the reported FPS values are internally inconsistent and physically implausible. The accuracy comparisons are also weakened by non-standard training protocols and missing error bars. Consequently, the stated contribution, as presented, cannot be supported.

major comments (4)

[§3.2, Eq. (8)-(11); Appendix C] The '21 extra parameters' claim is contradicted by the architecture described in the manuscript. Eq. (8) uses a multi-head attention block (Q/K/V/output projections plus LayerNorms), Eq. (9) an FFN, Eq. (11) a linear layer, and the semantic memory bank U is updated as a trainable quantity in Eq. (6). Appendix C further states that each category maintains five 512-dimensional prototypes, which alone gives 5×512×C parameters (2.56 million for ImageNet's 1000 classes). No derivation of the '21' count is provided. This is load-bearing because the paper's lightweight positioning and Table IV depend on it.
[Table IV vs Table VI; §3.2] The efficiency results are internally inconsistent and physically implausible. Table IV reports PromptKD at 12943.34 FPS and 'PromptKD w/ Spotlighter +1813.52', while Table VI reports 131.25K FPS for the toklev_1+2 configuration. Even ignoring the internal mismatch, 131.25K images per second on ImageNet with ViT-B/16 on a single GPU is not credible. Moreover, token selection occurs after the full image encoder has produced FI (Eq. (1)), so pruning tokens after the backbone cannot reduce backbone FLOPs. The claimed FPS gains are therefore unexplained.
[Appendix C] The comparison is not controlled. Appendix C states that CoOp is reduced to 15 training epochs and CLIPFit/PromptKD to 30 epochs, but it does not state how many epochs Spotlighter is trained. With these protocol differences, the 'up to 11.19%' improvement over CLIP is not clearly separable from retraining effects. Despite claiming results are averaged over three seeds, the paper reports no variance or error bars in any table, so statistical significance cannot be assessed.
[Appendix A; §4.2] The paper's own Limitations section concedes that performance degrades when discriminative tokens are sparse or class boundaries are highly entangled, particularly in ultra-fine-grained settings. This is precisely the regime where the method's core selection assumption—that top-k activated tokens are the most informative—could fail. Yet the paper claims consistent improvements across 11 datasets, including several fine-grained ones, without providing any quantitative analysis of ranking quality or performance in the acknowledged failure regime. This weakens the generality claim.

minor comments (6)

[§4.1] Typo: 'Apeendix C' should be 'Appendix C'. Also, Table II's caption uses 'PrompKD'.
[Table III] The caption contains 'CoOpOp' and the row labels are confusingly repeated ('CoOp' appears twice). Please fix.
[Table X] The StanfordCars prompt is listed as 'a photo of a [CLASS], a type of flowers.'; it should be 'a type of car.'
[Eq. (16)] The symbol 'tokrep_r' is not defined; presumably it should be 'tokrep_v'.
[§3.2] The notation 'toklevi' and 'toklev_i' is used inconsistently; please define the tier indexing explicitly.
[Introduction] Grammar: 'Our main contributions are lies in' should be 'Our main contributions lie in'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the empirical pipeline is self-contained and falsifiable.

full rationale

The claimed result is an empirical accuracy/efficiency comparison, not a derivation from first principles. The pipeline extracts CLIP visual/text features (Eq. 1), computes activation scores from those features and from a learned prototype bank (Eqs. 3-6), selects top-k tokens, refines them through IRM/TRM (Eqs. 8-11), and trains with contrastive, local, regularization, and KL losses (Eqs. 12-15). None of these steps uses the reported HM/FPS numbers as an input. The prototype bank is initialized from text embeddings and updated by momentum; although the same prototypes are used both to select tokens and to supervise them via Llocal, this is a standard learned-feature feedback loop, not a case of a fitted parameter being renamed as a prediction. The accuracy numbers are measured on held-out base/novel splits, with the paper's own Appendix A conceding settings where the method degrades, so the claim is falsifiable rather than tautological. The only notable issue raised by a skeptical reading — the '21 extra parameters' claim vs. the multi-head attention/FFN/linear-layer architecture in §3.2 — is an accounting or correctness inconsistency, not a circularity: it does not reduce a derived result to its own inputs. No load-bearing self-citations were found; prior-work citations are standard external references.

Assumptions & free parameters 7 free parameters · 4 assumptions · 1 invented entities

The method's empirical claims rest on several domain assumptions about CLIP feature distributions and on a set of hand-chosen hyperparameters. The most serious burden is the unexplained '21 parameters' count, which contradicts the described architecture.

free parameters (7)

IRM FFN weights = unknown (claimed 21 total params)
Trainable weights in Eq. 9; count not specified.
TRM linear weights = unknown (claimed 21 total params)
Trainable weights in Eq. 11.
Prototype matrix U = k x c (5 prototypes per class)
Learned in the semantic memory bank (Eq. 6).
alpha (fusion coefficient) = 0.2
Chosen by hand (Appendix C).
beta (momentum coefficient) = 0.8
Controls prototype update (Appendix C).
lambda_1, lambda_2, lambda_3 = 0.02, 20, 0.1
Loss weights fixed across datasets (Appendix C).
k (number of activated tokens) = not stated in main text
Ablated in Fig. 7; exact value missing.

assumptions (4)

domain assumption Top-k tokens with highest activation scores contain the most discriminative information.
Central premise of the selection; validated only via ablation on ImageNet (Table XIV).
domain assumption Prototypes initialized from text-embedding seeds are better than random.
Shown in Table IX across 11 datasets.
domain assumption Pre-trained transformer layers in Eq. 10 can be used without destroying performance.
The paper does not specify whether these are frozen; if trainable, the 21-parameter claim fails.
standard math Cosine similarity in the shared CLIP space is a reliable proxy for cross-modal token relevance.
Used in Eqs. 2-6.

invented entities (1)

Semantic Memory Bank (SMB)
purpose: Store class-specific prototypes to guide token selection and recovery of discarded information.
The SMB is an internal model component; its effectiveness is shown only by in-paper ablations (Table IX, Table VIII), no external falsifiable handle.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Spotlighter: Revisiting Prompt Tuning from a Representative Mining View." pith.science (2026). https://pith.science/paper/A4ZK3QM5

@misc{pith2026250900905,
  author       = {Pith},
  title        = {Pith review of: Spotlighter: Revisiting Prompt Tuning from a Representative Mining View},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A4ZK3QM5}},
  note         = {Machine review of arXiv:2509.00905}
}

read the original abstract

CLIP's success has demonstrated that prompt tuning can achieve robust cross-modal semantic alignment for tasks ranging from open-domain recognition to fine-grained classification. However, redundant or weakly relevant feature components introduce noise and incur unnecessary computational costs. In this work, we propose Spotlighter, a lightweight token-selection framework that simultaneously enhances accuracy and efficiency in prompt tuning. Spotlighter evaluates each visual token's activation from both sample-wise and semantic-wise perspectives and retains only the top-scoring tokens for downstream prediction. A class-specific semantic memory bank of learned prototypes refines this selection, ensuring semantic representativeness and compensating for discarded features. To further prioritize informative signals, we introduce a two-level ranking mechanism that dynamically weights token--prototype interactions. Across 11 few-shot benchmarks, Spotlighter outperforms CLIP by up to 11.19\% in harmonic mean accuracy and achieves up to 0.8K additional FPS, with only 21 extra parameters. These results establish Spotlighter as an effective and scalable baseline for prompt tuning. Code for our method will be available at https://github.com/greatest-gourmet/Spotlighter.

Figures

Figures reproduced from arXiv: 2509.00905 by the authors.

**Figure 2.** Overview of SpotLighter. The visual and textual features first compute sample-wise activations via a [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Visualization of the effect of the activated [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: The impact of different aggregation coefficient [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

**Figure 5.** Figure 5: The effect of different loss balance parameters [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

**Figure 6.** Figure 6: Performance of few-shot learning across 11 datasets compared with CLIPFit ( [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

**Figure 7.** Figure 7: The impact of different activated and repre [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 7 canonical work pages

[12]

a photo of a [CLASS]

for texture classification; and SUN397 (Xiao et al., 2010) for scene understanding.In distribu- tion shift experiments,we also introduce ImageNet- V2 (Recht et al., 2019), ImageNet-Sketch (Wang et al., 2019), ImageNet-A (Hendrycks et al., 2021b) and ImageNet-R (Hendrycks et al., 2021a).These datasets are all to improve ImageNet test reliabil- ity. This co...

work page 2010
[512]

Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhu- oshu Li, Hao Yang, et al

Springer. Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhu- oshu Li, Hao Yang, et al. 2024. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525. Ziqian Lu, Mushui Liu, Yunlong Yu, Zhao Wang, Xi Li, and Jungong Han. 2025. Variational adapter: Im- proving clip in data-...

arXiv 2024
[2009]

for object classification; Caltech (Fei-Fei et al., 2007) for natural object recognition; Ox- fordPets (Parkhi et al., 2012) for fine-grained pet classification; StanfordCars (Krause et al., 2013) for vehicle categorization; Flowers (Nilsback and Zisserman, 2008) for flower species identification; Food101 (Bossard et al., 2014) for food classifi- cation; ...

work page 2007
[2012]

for action recognition; DTD (Cimpoi et al.,

work page
[2014]

Food-101 - mining discriminative components with random forests. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzer- land, September 6-12, 2014, Proceedings, Part VI , volume 8694 of Lecture Notes in Computer Science, pages 446–461. Springer. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, ...

arXiv 2014
[2020]

In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24–25

Score-cam: Score-weighted visual explana- tions for convolutional neural networks. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24–25. Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. 2019. Learning robust global representations by penalizing local predictive power. Advances in neural inf...

work page 2019
[2021]

arXiv preprint arXiv:2104.13921

Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921. 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. IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens., 12(7):2217–2226. Dan H...

arXiv 2019
[2022]

In International conference on machine learning, pages 11162–11172

Vit-net: Interpretable vision transformers with neural tree decoder. In International conference on machine learning, pages 11162–11172. PMLR. Sungyeon Kim, Boseung Jeong, Donghyun Kim, and Suha Kwak. 2024. Efficient and versatile robust fine- tuning of zero-shot models. In European Conference on Computer Vision, pages 440–458. Springer. Jonathan Krause, ...

arXiv 2024

Show all 12 references

[2023]

In The IEEE/CVF Conference on Computer Vision and Pat- tern Recognition

Maple: Multi-modal prompt learning. In The IEEE/CVF Conference on Computer Vision and Pat- tern Recognition. Muhammad Uzair Khattak, Syed Talal Wasim, Muza- mmal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. 2023a. Self-regulating prompts: Foundational model ad...
[2024]

In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 28578–28587

Argue: Attribute-guided prompt tuning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 28578–28587. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Ba...

2023 arXiv
[2025]

arXiv preprint arXiv:2503.17782

Goal: Global-local object alignment learning. arXiv preprint arXiv:2503.17782. Table IX: Different ways for prototype initialization. Method Base Novel HM Random Initialization 82.54 77.46 79.92 Text Embedding Seeds 85.65 80.46 82.89 Mircea Cimpoi, Subhransu Maji, Iasonas Kokk...

2014 arXiv
[3492]

Jingjing Xie, Yuxin Zhang, Jun Peng, Zhaohong Huang, and Liujuan Cao

IEEE Computer Society. Jingjing Xie, Yuxin Zhang, Jun Peng, Zhaohong Huang, and Liujuan Cao. 2024. Textrefiner: Internal visual feature as efficient refiner for vision-language models prompt tuning. arXiv preprint arXiv:2412.08176. Chen Xu, Yuhan Zhu, Haocheng Shen, Boheng Che...

2024 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [12]

a photo of a [CLASS]

for texture classification; and SUN397 (Xiao et al., 2010) for scene understanding.In distribu- tion shift experiments,we also introduce ImageNet- V2 (Recht et al., 2019), ImageNet-Sketch (Wang et al., 2019), ImageNet-A (Hendrycks et al., 2021b) and ImageNet-R (Hendrycks et al., 2021a).These datasets are all to improve ImageNet test reliabil- ity. This co...

work page 2010

[2] [512]

Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhu- oshu Li, Hao Yang, et al

Springer. Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhu- oshu Li, Hao Yang, et al. 2024. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525. Ziqian Lu, Mushui Liu, Yunlong Yu, Zhao Wang, Xi Li, and Jungong Han. 2025. Variational adapter: Im- proving clip in data-...

arXiv 2024

[3] [2009]

for object classification; Caltech (Fei-Fei et al., 2007) for natural object recognition; Ox- fordPets (Parkhi et al., 2012) for fine-grained pet classification; StanfordCars (Krause et al., 2013) for vehicle categorization; Flowers (Nilsback and Zisserman, 2008) for flower species identification; Food101 (Bossard et al., 2014) for food classifi- cation; ...

work page 2007

[4] [2012]

for action recognition; DTD (Cimpoi et al.,

work page

[5] [2014]

Food-101 - mining discriminative components with random forests. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzer- land, September 6-12, 2014, Proceedings, Part VI , volume 8694 of Lecture Notes in Computer Science, pages 446–461. Springer. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, ...

arXiv 2014

[6] [2020]

In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24–25

Score-cam: Score-weighted visual explana- tions for convolutional neural networks. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24–25. Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. 2019. Learning robust global representations by penalizing local predictive power. Advances in neural inf...

work page 2019

[7] [2021]

arXiv preprint arXiv:2104.13921

Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921. 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. IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens., 12(7):2217–2226. Dan H...

arXiv 2019

[8] [2022]

In International conference on machine learning, pages 11162–11172

Vit-net: Interpretable vision transformers with neural tree decoder. In International conference on machine learning, pages 11162–11172. PMLR. Sungyeon Kim, Boseung Jeong, Donghyun Kim, and Suha Kwak. 2024. Efficient and versatile robust fine- tuning of zero-shot models. In European Conference on Computer Vision, pages 440–458. Springer. Jonathan Krause, ...

arXiv 2024