Vision-Language Models are Strong Noisy Label Detectors

Recent research on fine-tuning vision-language models has demonstrated impressive performance in various downstream tasks. However, the challenge of obtaining accurately labeled data in real-world applications poses a significant obstacle during the fine-tuning process. To address this challenge, this paper presents a Denoising Fine-Tuning framework, called DeFT, for adapting vision-language models. DeFT utilizes the robust alignment of textual and visual features pre-trained on millions of auxiliary image-text pairs to sieve out noisy labels. The proposed framework establishes a noisy label detector by learning positive and negative textual prompts for each class. The positive prompt seeks to reveal distinctive features of the class, while the negative prompt serves as a learnable threshold for separating clean and noisy samples. We employ parameter-efficient fine-tuning for the adaptation of a pre-trained visual encoder to promote its alignment with the learned textual prompts. As a general framework, DeFT can seamlessly fine-tune many pre-trained models to downstream tasks by utilizing carefully selected clean samples. Experimental results on seven synthetic and real-world noisy datasets validate the effectiveness of DeFT in both noisy label detection and image classification.

Paper

Similar papers

Peer review

Reviewer mmEa7/10 · confidence 5/52024-07-08

Summary

This paper proposes a novel method for learning with noisy labels leveraging pre-trained foundation models. The method is motivated by new findings that prompt learning is more robust to noisy labels when fine-tuning CLIP. The paper then designs a simple detector by learning both positive and negative prompts for each class. This paper showcases strong empirical results on multiple datasets, outperforming many existing methods.

Strengths

1. The proposed method is well-motivated. The effectiveness of CLIP is often overlooked in previous works for learning with noisy labels, and this paper examines it thoroughly and finds a good way to adapt CLIP in downstream tasks even with noisy labels. 2. This paper makes an important contribution that prompt learning is robust to robust to noisy labels, while full fine-tuning is better on clean datasets. Based on this observation, the paper proposes a two-stage method. 3. This paper proposes a novel idea by jointly learning positive and negative prompts for classes, and leverage the negative learning loss to optimize learnable prompts. 4. Extensive experiments on both synthetic and real-world datasets show that the proposed method consistently outperforms existing baselines.

Weaknesses

1. It is unclear why positive and negative prompts can help detect noisy labels. It is suggested to provide some visualization examples.

Questions

1. Can examples detected as noisy label help improve the performance?

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have adequately addressed the limitations and potential negative societal impact of their work.

Reviewer FyAK6/10 · confidence 4/52024-07-08

Summary

This paper focuses on improving the finetuning performance of pretrained vision models by removing the noisy labeled data. Specifically the authors propose a two-stage method: the first stage is to learn a noise detector by prompt learning of the text encoder in CLIP. The second stage is to full-finetune the pretrained models (including CLIP and other pretrained models) . In the experiment, the authors include both synthetic datasets and real datasets and compare their model with different 6 different methods. They also use the filtered dataset to finetune different pretrain models.

Strengths

1. The paper is well writen and easy to follow. 2. The analysis of the relationship between noisy data and finetune methods is very insightful (Figure 1) 3. The proposed two-stage method is reasonable and results in good performance compared to other baselines. 4. It is great to show that the filtered data works for both CLIP models and non-CLIP models even though the filtering is done by the model based on CLIP.

Weaknesses

1. The optimization part is a bit unclear. (See Questions Below) 2. The proposed noisy detector cannot be reusable when working on a different dataset. Each downstream dataset needs a specific trained detector. 3. Noisy label problem is a more severe problem during the pretraining than the finetuning. But this paper only focuses on the finetuning, which makes the research scope narrow in this case.

Questions

1. I am confused about the optimization for the positive prompts. Are positive prompts learnt together with negative prompts and Visual PEFT by using $L_{dp} + L_{sim}$? Or are they learned separately? 2. Can a noisy detector learned on one dataset transferred to another dataset? For example, the noisy detector learned on ImageNet can be used to detect noise in CIFAR100 dataset since there are a great number of class overlapping in these two datasets?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have included the discussion of limitations in the supplementary.

Reviewer 6CXy5/10 · confidence 4/52024-07-10

Summary

The paper proposes a method for detecting noisy samples using vision-language models (CLIP). The main idea is to efficiently adapt (via prompt tunning) the clip model on noisy data and use this adapted model to select clean data. In the second stage, the clean data can be used to fully fine-tune a backbone model. The paper proposes the following contributions: a) using two learnable text prompts (positive and negative) to avoid using a threshold when detecting noise b) using negative learning [40] c) using a two-stage approach to adapt to the noisy distribution (first select clean samples then fully fine-tune using them). Experiments are done on multiple datasets with synthetic (CIFAR-100, Tiny-ImageNet, Stanford-Cars, CUB-200-2011) or real noisy labels (CIFAR-100N, Clothing1M, WebVision).

Strengths

* S1. Good direction of efficiently adapting CLIP models for noise detection. * S2. The general setting is sound, of using efficiently fine-tuned CLIP to select clean data and then fine-tuning using it. * S3. Good results on multiple noisy datasets.

Weaknesses

* W1. There needs to be more fair comparisons with baselines that are trained in the same settings and use similar models. * W2. Table 1 seems to compare DEFT which has a fine-tuning stage on the noisy dataset with the initial clip model that is not trained or adapted at all. Is this the case? Is the small-loss baseline using the initial CLIP model, without any adaptation to the current dataset? * W2.2. More appropriate baselines would be CLIP models that are adapted in standard ways to the current noisy dataset. Thus, one aspect that needs to be ablated is the fine-tuning method (none, fully fine-tuning, efficient fine-tuning, etc). Then, given a CLIP model, the second aspect that needs ablation is the selection method. Here the paper already compares against zero-shot and small-loss selection, but using the initial, pre-trained CLIP. The same should be done using an adapted clip. * W2.3. Another simple selection method that, like DEFT doesn’t require a threshold would be to select samples with $sim(I_i, T_k) > th$ (where th=0) where $T_k$ is the text features corresponding to the correct class. This selection should be different from the zero-shot selection because the threshold is applied directly on the similarity of the image and correct class features, without applying softmax, thus without taking into account if the correct class is the class predicted by the zero-shot model. * W3. Table 2 and Table 3 compare against multiple methods for training in noisy settings, starting with the same visual backbone, i.e. the visual part of CLIP. Is this correct? The question arises, is DEFT using additional information since it does the selection using both the visual and textual part of CLIP? Thus, the DEFT has a possible unfair advantage. * W4. What is the motivation for using two text prompts (positive and negative)? If only one prompt is used (as usual), we can still make a prediction and optimize this prediction to be the correct class and then apply a fixed threshold (>0.5). Why shouldn’t the simple approach work and what benefits does the dual prompt bring?

Questions

* Is the $L_{sim}$ loss (Eq.8) used for the adaptation in the second stage, or is it used after the selection in phase 1? If the latter, would this mean that this loss updates the PEFT module and the learnable prompts, and then the final model is used in the second stage?

Rating

5

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

The method is not sufficiently evaluated against fair baselines, see weak points for more details.

Reviewer g1PL5/10 · confidence 2/52024-07-12

Summary

The paper introduces a Denoising Fine-Tuning (DEFT) framework to address the challenge of noisy labels in vision-language models, particularly focusing on models like CLIP. The DEFT framework leverages the robust alignment of textual and visual features pre-trained on extensive image-text pairs to filter out noisy labels. This is achieved by learning class-specific positive and negative textual prompts. Positive prompts highlight distinctive class features, while negative prompts act as thresholds to differentiate between clean and noisy samples. The framework uses parameter-efficient fine-tuning (PEFT) to adapt the visual encoder to align with the textual prompts. Extensive experiments on synthetic and real-world noisy datasets demonstrate that DEFT significantly improves both noisy label detection and image classification performance.

Strengths

1. This paper proposes to combining textual and visual prompts for noisy label detection, which enhances the robustness of vision-language models to label noise. 2. The framework's generalizability to various pre-trained models and its parameter efficiency make it a versatile solution. 3. The experimental validation on multiple datasets, including real-world noisy data, provides strong evidence of the method's effectiveness. 4. The use of PEFT to maintain the generalization ability of pre-trained models while adapting to specific tasks is particularly noteworthy.

Weaknesses

1. The heavy reliance on pre-trained models may limit the framework's applicability in scenarios where such models are not available or suitable. 2. The discussion on the practical implementation and potential limitations in different real-world settings is relatively limited. 3. The computational overhead associated with maintaining and fine-tuning dual prompts could be a concern in resource-constrained environments.

Questions

1. How does DEFT perform when applied to pre-trained models with varying degrees of generalizability and domain relevance? 2. Can DEFT be adapted to scenarios with extremely high noise ratios or highly imbalanced datasets, and what modifications would be necessary to maintain its effectiveness? 3. What are the potential trade-offs between the computational overhead of DEFT and its performance gains, and how can the framework be optimized for deployment in resource-constrained environments?

Rating

5

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have discussed the limitations of this paper, and there is no negative societal impact.

Reviewer 6CXy2024-08-09

Response to rebuttal

I thank the authors for the response. I appreciate the clarification of the Small-loss baseline and the additional ablations. In W2.3. you show that the initial, pre-trained CLIP model is sensible to the threshold and it makes sense. How about the adapted CLIP model, as asked in W4? Overall I tend to increase my score to 5.

Authorsrebuttal2024-08-10

Dear Reviewer 6CXy, We sincerely appreciate your valuable comments and encouraging feedback. For the adapted CLIP model in W4, the minimum $sim(I_i, T_k)$ is $-0.443$, and the F1-scores for noise detection at different thresholds are presented in the table below. These results indicate that the adapted CLIP model is also sensitive to the choice of threshold. We will add the new baseline in the revised version. | Threshold | 0.0 | 0.1 | 0.2 | 0.3 | 0.4 | 0.5 | | --------- | :---: | :---: | :---: | :---: | :---: | :---: | | F1-score | 80.77 | 85.72 | 89.89 | 91.93 | 92.75 | 91.81 |

Reviewer FyAK2024-08-12

The authors address all my concerns and questions. Thanks for the effort in making the rebuttal. I changed my rating from 5 to 6.

Authorsrebuttal2024-08-12

Response by Authors

Dear Reviewer FyAK, Thank you for your thoughtful suggestions and the positive feedback on our work.

Authorsrebuttal2024-08-14

Official Comment by the Authors

Dear Reviewer g1PL, We appreciate your thorough evaluation and helpful suggestions and comments. In our response, we have provided point-by-point responses to your specific comments. We hope our response addresses all the concerns raised in your review. Since the author-reviewer discussion ends soon, we are happy to hear your thoughts if you need additional clarifications from the authors. Thank you very much. Best

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC