S-CLIP: Semi-supervised Vision-Language Learning using Few Specialist Captions

Vision-language models, such as contrastive language-image pre-training (CLIP), have demonstrated impressive results in natural image domains. However, these models often struggle when applied to specialized domains like remote sensing, and adapting to such domains is challenging due to the limited number of image-text pairs available for training. To address this, we propose S-CLIP, a semi-supervised learning method for training CLIP that utilizes additional unpaired images. S-CLIP employs two pseudo-labeling strategies specifically designed for contrastive learning and the language modality. The caption-level pseudo-label is given by a combination of captions of paired images, obtained by solving an optimal transport problem between unpaired and paired images. The keyword-level pseudo-label is given by a keyword in the caption of the nearest paired image, trained through partial label learning that assumes a candidate set of labels for supervision instead of the exact one. By combining these objectives, S-CLIP significantly enhances the training of CLIP using only a few image-text pairs, as demonstrated in various specialist domains, including remote sensing, fashion, scientific figures, and comics. For instance, S-CLIP improves CLIP by 10% for zero-shot classification and 4% for image-text retrieval on the remote sensing benchmark, matching the performance of supervised CLIP while using three times fewer image-text pairs.

Paper

Similar papers

Peer review

Reviewer Q2NM5/10 · confidence 4/52023-07-06

Summary

The paper presents S-CLIP, a novel semi-supervised learning method for training contrastive language-image pre-training (CLIP) models in specialized domains where limited image-text pairs are available. Vision-language models like CLIP have achieved impressive results in natural image domains, but they struggle when applied to specialized domains like remote sensing due to the scarcity of paired data. To address this, S-CLIP leverages additional unpaired images and introduces two pseudo-labeling strategies: caption-level and keyword-level pseudo-labeling. Caption-level pseudo-labeling assigns pseudo-labels to unlabeled images based on the captions of paired images. This is achieved by solving an optimal transport problem between the unlabeled and labeled images, ensuring that the pseudo-labels represent a probability distribution over the labeled images. On the other hand, keyword-level pseudo-labeling assigns pseudo-labels based on the keywords in the captions of visually similar paired images. It formulates the training as a partial label learning problem, considering a candidate set of target keywords instead of a single exact one. The experiments conducted in various specialist domains, including remote sensing, fashion, scientific figures, and comics, demonstrate the effectiveness of S-CLIP. It outperforms CLIP fine-tuning and other semi-supervised learning competitors in zero-shot classification and image-text retrieval tasks. For example, in the remote sensing domain, S-CLIP improves zero-shot accuracy by 10.4% and image-to-text retrieval R@5 by 4.4% compared to CLIP fine-tuning. Moreover, S-CLIP remains robust even when the unlabeled images are from a different dataset. Overall, S-CLIP shows promise in enhancing CLIP training using only a few image-text pairs.

Strengths

- Effective Pseudo-Labeling Strategies: The paper proposes two pseudo-labeling strategies, caption-level and keyword-level pseudo-labeling, which effectively guide the training process. The caption-level pseudo-labeling leverages optimal transport to assign pseudo-labels based on the captions of paired images, ensuring robust training even with distribution shifts. The keyword-level pseudo-labeling assigns pseudo-labels based on keywords in visually similar paired images, capturing local components of unlabeled images. The combination of these strategies results in improved zero-shot classification and image-text retrieval tasks. - Experimental Demonstrations: The paper provides extensive experimental evaluations in various specialist domains, including remote sensing, fashion, scientific figures, and comics. The results consistently show that S-CLIP outperforms CLIP fine-tuning and other semi-supervised learning competitors, achieving significant improvements in zero-shot classification and image-text retrieval tasks. The robustness of S-CLIP, even when unlabeled images are from a different dataset, further strengthens its effectiveness. - Relevance and Applicability: The paper addresses an important and practical problem in the field of vision-language models. The proposed S-CLIP method has the potential to enhance the training of CLIP models in specialized domains with limited paired data. The findings are relevant not only for researchers working on vision-language models but also for practitioners in various domains where image-text understanding is crucial. - Clarity and Presentation: The paper is well-written, organized, and presents the proposed method and experimental results clearly. The figures and illustrations aid in understanding the concepts and methodologies. The strengths and weaknesses of the proposed method are discussed in a concise and informative manner.

Weaknesses

- Baselines with other semi supervised approaches are severely limited with the added baselines being severely outdated. Several much newer few shot learning approaches are present and should be compared against. A summary could be found in the paper: A Survey on Deep Semi-supervised Learning by Yang et al. for example: SEMI-MAE: MASKED AUTOENCODERS FOR SEMI-SUPERVISED VISION TRANSFORMERS by Yu et al or Semi-supervised Multimodal Representation Learning through a Global Workspace by Devillers et al - A lot of the core content like limitations, experimental setup etc has been moved to the appendix which I feel is key to the understanding of the paper and unfairly grants the author extra writing space - The approach combines keyword and caption level pseudo labelling, both of which have individually been well studied in literature. The combination of both is certainly a novel effort, but the contribution of the same doesnt seem like a strong enough novelty for a venue liek Neurips - The choice of using OT as the distance measure between 2 images is interesting, it would be nice to see how other image distance measures compare to it

Questions

- Could you clarify why the soft-PL, hard-PL were the only semi supervised baselines chosen for comparison despite the presence of a vast body of literature for semi supervised learning?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

No ethical limitations as such

Reviewer XrtM6/10 · confidence 4/52023-07-06

Summary

The paper studies semi-supervised learning for contrastive vision-language pre-training, during which unlabelled/unpaired data is also accessible. The authors formulate the pseudo-labeling as an optimal transport (OT) problem and propose to use the Sinkhore algorithm to produce pseudo labels, which is validated to be superior to vanilla had/soft PL solution. Comparisons with other works in literature utilizing OT are properly discussed. Furthermore, the paper proposes to use the key-level pseudo label to account for partial label learning. Experiments on several data domains are conducted to validate its effectiveness.

Strengths

- The paper is technically sound and easy to follow. - The evaluation is self-sufficient with proper baselines, which helps to corroborate the effectiveness of the proposed loss. - The ablative study is sufficient and showcases the impact of different composing parts of the method. - Consistent performance gain over a variety of downstream datasets.

Weaknesses

- For keyword-level pseudo-label, the compared baseline is somewhat not fair, since one could also leverage the keyword information for labeled/paired data in the vanilla fine-tuning setting, i.e., fine-tuning pre-trained CLIP with CLIP loss and multi-class classification loss. Additional experiments should be conducted to verify the gain derives from the semi-supervised part of it instead of simply leveraging the label information. - It would also be nice to showcase results in general data (e.g., the ones evaluated in the CLIP paper, Flicker, etc.). Manual data split may be necessary but it helps to corroborate the generalizability of the proposed method.

Questions

- The title is somewhat confusing. The paper studies semi-supervised fine-tuning on pre-trained CLIP weights, which is not suitable to phrase as "semi-supervised pre-training", where you are dealing with an unlabelled general domain. - When keywords are obtained from class names, does one labeled image normally only contain just one keyword? If yes, does this reduce to a classification loss (since the target is one-hot)?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

All limitations are properly addressed.

Reviewer 7J1f6/10 · confidence 3/52023-07-07

Summary

The paper introduces an additional training method for vision-language models like contrastive language-image pre-training (CLIP). The proposed method consists of semi-supervised learning with two different pseudo labeling including caption- and keyword-level pserudo-label. These two pseudo labelings are complementarily understand image through different types of text representation. The experimental section describes the effectiveness of the proposed method in terms of zero-shot image recognition and image-text retrieval in several representative datasets. A relatively small image-text paired dataset can improve the zero-shot recognition and image-text retrieval with the proposed pseudo labelings.

Strengths

+ The paper serves a method for semi-supervised learning on image-text paired datasets. Usually, it is quite difficult to construct the kind of dataset, however, the paper proposed that a simple yet effective idea to solve a problem in vision language pre-training like CLIP. + This paper also points out that limited focused areas such as remote sensing and fashion cannot be effectively used a vision-language models like CLIP as is. Therefore, the paper insists that an easily adaptive labeling approach and semi-supervised fine-tuning from pseudo labels are important for CLIP to adapt a specific image dataset. + In the experimental section, the proposed method S-CLIP is much better than that of used the original CLIP and fine-tuned CLIP (CLIP-FT). It is reasonable to assign soft labels rather than hard labels (Hard-PL) in this problem. + The paper is well-written and effectively communicates the main ideas in the paper, making it accessible to a broad audience. Moreover, there contains various visualizations, e.g., Figures 1 and 2, to confirm that the learnt representations of vision-language pre-training are meaningful.

Weaknesses

- CLIP is pre-trained on image-text paired dataset widely extracted from the major space on the Internet. Therefore, it is understandable that the dataset collection does not work well in situations where only a small amount of paired data is collected, or where the paired data is not steadily available on the web. On the other hand, does the proposed dataset collection method only work on remote sensing, fashion, scientific figures, and comics? If you could show the vision-language application based on the proposed dataset collection in more situations, the paper must be more effective to broader readers. Probably, the approach could be shown in more generic situations, however, there are only 4 types of situations. Though the authors do not necessarilly add the experimental results, how about discussing the effectiveness for general purpose dataset collection?

Questions

Please see above-mentioned comment in 'weaknesses'.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

There are no negative limitations and societal impacts.

Reviewer 7J1f2023-08-21

Response to author rebuttal

Thank you so much for the authors. The rebuttal has addressed my question and concern. I will keep my paper rating.

Reviewer CjR36/10 · confidence 4/52023-07-08

Summary

Fine-tuning CLIP in specialized domains suffer the problem of limited available image-text pairs. This paper propose to address this challenge via semi-supervised learning methods with the additional unpaired images. Two pseudo-labeling strategies including the caption-level pseudo labeling and keyword-level pseudo labeling are proposed to construct valuable information about the unpaired images. The problem addressed in the paper is meaningful and the method proposed in the article is demonstrated effective through comprehensive experiments.

Strengths

(+) This article is well written, and the description of the method pipeline is clear and easy to understand. (+) Experiments are sufficient. (+) The problem it solves is meaningful.

Weaknesses

(-) "Baseline" in table 5 is a little confusing. See question (1). (-) The effectiveness of OT-based "Caption-level pseudo-label" compared with soft-PL should be further verified.

Questions

(1) "Baseline" in table 5 is a little confusing. Is that refers to the "CLIP (fine-tune)" method or the "hard-PL" or "soft-PL" baselines ? (2) It seems the role of the proposed OT-based "Caption-level pseudo-label" is similar as the soft-PL baseline. I suggest the authors provide a comparison of the effect of using the OT-based method and the soft-PL method on top of CLIP fine-tuned; or replace the OT-based method with the soft-PL, ablation on S-CLIP to show the effectiveness. (3) I suggest the accuracy of training on 100% labels (full data) should be reported, as the upper bound performance to be reffered.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

No

Reviewer CjR32023-08-17

I thank the authors for their response. My concerns have been adequately addressed. My opinion about the paper is still positive.

Reviewer EbWV6/10 · confidence 3/52023-07-09

Summary

This paper proposes a method S-CLIP, to enhance CLIP in a semi-supervised way by leveraging captions (via optimal transport) and keywords (via image-keyword similarity) from other unpaired images in the same batch. This is beneficial for specialized domains like remote sensing, where image-text data is usually limited. Results show better performance than the traditional pseudo-labelling methods, especially in OOD domains.

Strengths

Addressing low-data regimes for contrastive learning. Such a problem is common in specialized domains like remote sensing. Novel and simple method, but showing promising results compared to the standard pseudo-labelling methods, especially in OOD domains. Experiments and ablations are comprehensive. Well written and easy to read.

Weaknesses

In section 5 experiments, missing experiment details for reproducing, such as model size, epoch, training schedule (lr, weight decay etc.)... Besides the traditional pseudo-labelling methods, it would be nicer if we could compare with other SOTA pseudo-labelling methods for vision-language training – if there were no such SOTA methods, we could claim that in the paper.

Questions

Given this method is sensitive to the batch data (e.g. statement in 5.2), it would be more solid if we could consider some data preprocessing strategies, e.g., to batch similar images in the same batch as much as possible? Or at least we should study how batch size affects the results in ablations? In line 237, “we employ 32 image-caption pairs and 32 unpaired images”: this seems contrary with the problem to solve, i.e., N << M? Could we give some qualitative examples for caption-level and keyword-level selection? E.g., given a few images, which are the selected text (weighted captions/keywords) look like in the same batch? minor comment1: in Figure 1b, Hard-PL is “-2.7” instead of “-1.7”? minor comment2: the entire section 3 background could be removed to save more space for ablations/analysis, which is more valuable to readers (given the background intro is too basic)? minor comment3: explain what “YAKE-100/200/300” mean in Table 5? minor comment4: “(f) Partial” to be consistent with other items.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

No specific concern.

Reviewer EbWV2023-08-17

Thanks for the clear and detailed response, which has generally addressed all my concerns. I would like to keep my original positive rating.

Reviewer Q2NM2023-08-17

I appreciate the authors feedback. I would like to keep my original positive rating :)

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC