CAPro: Webly Supervised Learning with Cross-Modality Aligned Prototypes

Webly supervised learning has attracted increasing attention for its effectiveness in exploring publicly accessible data at scale without manual annotation. However, most existing methods of learning with web datasets are faced with challenges from label noise, and they have limited assumptions on clean samples under various noise. For instance, web images retrieved with queries of tiger cat (a cat species) and drumstick (a musical instrument) are almost dominated by images of tigers and chickens, which exacerbates the challenge of fine-grained visual concept learning. In this case, exploiting both web images and their associated texts is a requisite solution to combat real-world noise. In this paper, we propose Cross-modality Aligned Prototypes (CAPro), a unified prototypical contrastive learning framework to learn visual representations with correct semantics. For one thing, we leverage textual prototypes, which stem from the distinct concept definition of classes, to select clean images by text matching and thus disambiguate the formation of visual prototypes. For another, to handle missing and mismatched noisy texts, we resort to the visual feature space to complete and enhance individual texts and thereafter improve text matching. Such semantically aligned visual prototypes are further polished up with high-quality samples, and engaged in both cluster regularization and noise removal. Besides, we propose collective bootstrapping to encourage smoother and wiser label reference from appearance-similar instances in a manner of dictionary look-up. Extensive experiments on WebVision1k and NUS-WIDE (Web) demonstrate that CAPro well handles realistic noise under both single-label and multi-label scenarios. CAPro achieves new state-of-the-art performance and exhibits robustness to open-set recognition. Codes are available at https://github.com/yuleiqin/capro.

Paper

References (100)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer QoQL5/10 · confidence 5/52023-06-21

Summary

This paper proposes a unified prototypical contrastive learning framework, named Cross-modality Aligned Prototypes (CAPro), to learn visual representations with correct semantics. CAPro exploits web data across modalities to formulate semantically-correct textual and visual prototypes. The authors propose text matching to leverage textual prototypes to establish their noise-robust estimation. They also bring in text enhancement with visual guidance from image neighbors to mitigate the effect of noisy texts on clean sample selection. Further, the authors propose collective bootstrapping (CB) to provide smoother label reference by extending bootstrapping with collective knowledge. Experimental results on WebVision1k and NUS-WIDE (Web) are provided to validate the effectiveness of the proposed method.

Strengths

- The idea of bridging visual and texture prototypes to cope with webly supervised learning is interesting and promising. - The motivation is clear, and the paper is written well. - The code is released. The ablation study seems extensive.

Weaknesses

- Eq.(5), (6), and (7) seem very similar to those in [28]. It is expected to discuss the difference between the proposed one with [28]. Why is the proposed one better? - The total objective loss function shown in Ln 242 contains 4 loss weight hyper-parameters. However, only \lambda_{bts} is discussed in the ablation study. What about other hyper-parameters? Moreover, it is also a concern that the proposed method involves too many hyper-parameters, including these four loss weights, \gamma, update frequency, etc. How to decide the proper value of these hyper-parameters? What about the robustness? - In Table 1, the result of NCR[58] (i.e., 73.9) seems wrong. The result shown in the NCR paper is 75.7 for NCR and 76.8 for NCR+Mixup+DA, both surpassing the reported result of the proposed method. This raises a concern about the performance of the proposed method. - Some minor issues: - In Ln 62, a period is missing between "selection" and "We". - In Ln 69, the sentence uses the past tense, while the former and latter sentences basically use the present tense. It is recommended to change this sentence to present tense to make them consistent.

Questions

N/A

Rating

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

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors have briefly discussed the limitations and broader impact.

Reviewer SPGi5/10 · confidence 4/52023-07-03

Summary

This paper dives into the study of webly-supervised learning and aims to utilize the neglected alt-text of web images to enhance the learning process. To this end, the authors propose the approach called Cross-modality Aligned Prototypes (CAPro). CAPro is adopted with two modules, namely, text matching&enhacement, and collective bootstrapping. The former aims to assign text to the corresponding prototypes by resorting to the LLMs and cross-modal nearest neighbor mechanism. The latter provides smoother label reference by extending bootstrapping with collective knowledge. Extensive experiments have been conducted on several benchmarks to verify the effectiveness of the proposed CAPro.

Strengths

Most existing webly-supervised learning works mainly focus on the web images and corresponding (noisy) labels while omitting the potential alt-texts (captions). This paper provides a new perspective that uses the alt-text to complement the webly-supervised learning. From this point, the motivation and idea of this paper are novel and interesting.

Weaknesses

1. My major concern is the differences between this work and the existing problem or techniques including the noisy correspondence, noise-robust learning from NNs, and noise removal. First, this paper claims that the existing webly-supervised learning works mainly address certain types of noise including label-flipping noise and out-of-distribution (OOD), while neglecting the semantic noise, namely, the misalignment between image contents and the associated texts. The claim might be correct for the webly-supervised learning community. However, the so-called semantic noise is very similar to the definition of noisy correspondence [61,62, 65]. The authors should provide more discussion to clarify the differences between the so-called noisy correspondence and semantic noise. If the two problems are somewhat similar, I think it would be better to give more discussion on the related works like 'Noisy Correspondence Learning with Meta Similarity Correction'. Second, the differences between the used KNN-graph mechanism and the works in 'Noise-Robust Learning from Neighbors' are encouraged to be further discussed. Third, the proposed noise removal strategy seems to be similar to the sophisticated WSL method (MoPro), which is an important baseline for WSL learning. 2. The performance improvement is limited (See Table 1). This paper adopts a relatively complex pipeline and additional resort to the existing LLMs. However, the performance is marginal compared to the sophisticated WSL baseline (MoPro, ICLR 2020). 2. There are some typos and unclear statements. For example: i) the definition of 'concept definition texts (Line 56-57)' is lacking ii) 'with visual guidance from image neighbors (Line 61)' is unclear. iii) 'selection We' (Line 62)

Questions

My major concern is the differences between this work and the existing problem and techniques as eloborated on in Weaknesses. Thanks

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

3 good

Contribution

3 good

Limitations

The authors have discussed the limitations.

Reviewer XBs56/10 · confidence 4/52023-07-04

Summary

To handle the label noise problems especially the semantic noise in webly supervised learning, the authors propose a unified prototypical contrastive learning framework named as Cross-modality Aligned Prototypes (CAPro). It exploits web data across modalities to formulate semantically-correct textual and visual prototypes. Besides, collective bootstrapping is proposed to encourage smoother and wiser label reference from appearance-similar instances in a manner of dictionary look-up. Extensive experiments on WebVision1K and NUS-WIDE(Web) demonstrate that the proposed CAPro could handle realistic noise under different scenarios. It also achieves new state-of-the-art performance on open-set recognition.

Strengths

1. The authors propose a new cross-modality prototypical learning framework named as CAPro, which aims to handle various noises, especially semantic noise. 2. The proposed CAPro framework is carefully designed and fully explores the cross-modality intervention to filter out noise. 3. Extensive experiments have been conducted to prove the effectiveness of the proposed method. On the other hand, the code is available and enables reproducing this work.

Weaknesses

1. The framework figure, i.e., fig. 2, is too complex to be understood. The arrows are not in the same directions, and the symbols are not introduced in caption or texts. There is also lack of explicit module division in the figure. These aspects increase the difficulty to understanding the whole framework. 2. In experiments, it seems that different methods obtain different “vanilla” performance even with the same backbone, such as R50. Why is the performance different among different methods? Does it mean unfair comparison? 3. The whole framework is complex since it introduces many modules to handle noise. The whole training process should be organized as explicit algorithm to facilitate the understanding. 4. There are a lot of hyper-parameters in the total objective. How to choose them during experiments? 5. There are some typos in the manuscript.

Questions

Please try to address the weaknesses above.

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

2 fair

Contribution

3 good

Limitations

n/a

Reviewer HQEo6/10 · confidence 4/52023-07-06

Summary

The authors propose a prototypically-aligned contrastive learning framework for vision and language in order to enable better web-scraping of fine-grained, rare concepts that are easily confused or mapped to other more common concepts when either vision or language is considered in isolation (what they term “semantic noise”). Their goal is to improve webly-supervised learning, which is often plagued by label noise, particularly for fine-grained categories. They call their method CAPro, for “Cross-modality Aligned Prototypes.” Their method is built from a few simple ideas: 1) use text prototypes, well-defined in the literature, to scrape a cleaner set of visual prototypes for fine-grained classes, 2) Use visual features to fill gaps or correct errors within text prototypes for fine-grained concepts, generating “semantically-aligned visual prototypes,” 3) an additional cleaning step to further reduce noise between the visual and textual prototypes that uses cluster regularization, 4) “collective bootstrapping” to further smooth concepts or “label references” via essentially an adapted/aggregated/bootstrapped dictionary lookup over the entire dynamic concept dictionary that reduces the effect of examples divergent from the average when making predictions, particularly helpful for visually similar classes. Their method is highly similar to MoPro, but with the addition of the above described semantic noise correction and regularization across modalities. After reading the authors rebuttal and discussing with them, as well as reading the comments and discussion with other reviewers, I will increase my score to a 6.

Strengths

The authors show that their method performs well for single-label and multi-labe, on WebVision1k and Imagenet1k, and shows some open-set generalizability. In particular, they define 235 categories from these datasets as exhibiting “polysemy concepts”, and show that their method is particularly effective for these concepts. The gains shown are nice, and the more detailed performance breakdown of the 235 class versions of both datasets is interesting (though it would be nice to also show the performance breakdown for the non-polysemy concepts to help build reader intuition of the impact of polysemy). The method does seem to clearly have an advantage in multi-label challenges, and performs better than CoTeach or VSGraph on open-set concepts. The qualitative examples in table 3 are quite compelling, though they seem potentially highly cherry-picked.

Weaknesses

Notably, their system is quite complex, and feels a bit like a bag of tricks (albeit an effective one), and the method both performs quite similarly to MoPro and often underperforms VSGraph on Top1 (the authors point out that these may not be directly comparable, as they use different backbones, but this makes it difficult to understand or interpret the table. Perhaps they could somehow point out which methods are comparable along which axes in the table? As it is, it’s not very easy to determine the methods value vs other methods). I would also have liked to see more thorough ablations to better understand how impactful their suggested components were vs the additional computational complexity added, to better understand how “worth it” each component would be to use or implement. I find table 4 to be a bit difficult to interpret. I would also like a more thorough analysis of why top-1 suffers, but top-5 benefits (this may be related to over-regularization somehow?). Nits: In the abstract: “exacerbates fine-grained visual concept learning.” should perhaps be “exacerbates the challenge of fine-grained visual concept learning.” Line 26: “The large scale” -> “Large scale” Line 305: “study” -> “studies”

Questions

I’m curious how this method would handle visual dimorphism in fine-grained classes, for instance dark-eyed juncos which have multiple distinct color morphs, some more common than others. Since the method explicitly seeks prototypical examples, is this at the expense of recognition of anomalies or rarities within-concept? Have the authors seen any interesting failure modes for their method, where simpler methods succeed?

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

The authors have not addressed this at all, which I see as a significant weakness. What are the potential benefits or harms when building prototypical definitions of concepts? What might this mean for instances which do not fit the prototypical visual appearance of a given concept on the internet? How might biases in internet data limit our prototypes?

Reviewer SPGi2023-08-18

Replying to the authors' rebuttal

Thanks for the detailed rebuttal. In the rebuttal, the author have clarified the difference between some related works. I would like to maintain my positive rating.

Reviewer HQEo2023-08-19

Thank you for the clarification and analysis in your rebuttal

I thank the authors for their rebuttal, and I believe that with the additions they have proposed in the general rebuttal across all reviews the paper will be significantly clarified and strengthened which is a win for the peer review process and will lead to a notably improved manuscript. If all 1-10 of the promised improvements are incorporated thoroughly and well into the final manuscript I believe that would justify an increased score of 6.

Reviewer XBs52023-08-21

I appreciate the authors for addressing most of my concerns and questions. Therefore, I will keep my rating.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC