NoiseGPT: Label Noise Detection and Rectification through Probability Curvature

Thank you for pointing out perspectives that could be further improved.

• Claims of NoiseGPT:

  • Although there is a lack of works leveraging MLLMs to mitigate label noise, prior researchers have explored its capability in different fields. Huang [1] proposes MVT to supervise vision models using MLLM predictions in OOD scenarios. However, this method cannot be transferred to label noise detection scenarios as the predictions of MLLMs vary ubiquitously among different classes. It is hard to detect and rectify noisy labels from different classes with a unique threshold. Thus, we investigate and leverage the ICD curvature indicating the tendency of MLLM predictions under perturbations to distinguish between clean and noisy examples. On the other hand, CLIP, as a contrastive multi-modal model has been used to select clean examples according to its confidence[2]. And an adaptive loss is constructed for the classifier subsequently through combination with the transition matrix and the class frequency prior to mitigate the overfitting to label noise during training. Nonetheless, NoiseGPT is still distinctive for its way of utilizing pre-trained MLLMs.
  • About the memorization effects. There are significant differences between the memorization effects [3] that previous works use and the properties of MLLMs we leverage in NoiseGPT. Arpit defines memorization effects as behavior exhibited by DNNs to optimize faster on clean data than noise. Based on the fact that DNNs are prone to produce a higher loss for noise during the early stage of training, existing works [4,5,6] leverage an early stopping strategy or recurrently filter out noisy examples to mitigate the label noise. On the other hand, NoiseGPT leverages the generalization capability of MLLMs in a zero-shot manner. The properties of ICD curvature lie in its inherent optimization to associate visual features with corresponding text labels.

• Ablation study:

  • we add an experiment where different numbers of perturbations are tested. The performance of label rectification is shown in Figure A in the attachment file of the author rebuttal. It is noted that there is a marginal effect when $n$ increases to an extent and it would be unworthy since the computing time grows significantly.
  • We leverage CLIP ViT-L as a label corrector in comparison with MLLM backbones (including a smaller version of LLM) in Table A to demonstrate the superiority of NoiseGPT.
  • We tried feature interpolation in the earlier experiments instead of the proposed MoF to introduce noisy perturbation into query examples. The distribution of ICD scores would be different this way, which is illustrated in Figure C in the attachment. We make experiments on CIFAR-10N worse label, which turns out that the overlaps between clean and noisy scores of MoF are smaller compared with that of feature interpolation. Thus, it indicates a better capability to distinguish noisy labels.

[1] Huang Z, Liu C, Dong Y, et al. Machine vision therapy: Multimodal large language models can enhance visual robustness via denoising in-context learning[C]//Forty-first International Conference on Machine Learning. 2023. [2] Liang, et al. "Combating Label Noise With A General Surrogate Model For Sample Selection." arXiv preprint arXiv:2310.10463 (2023). [3] Arpit D, Jastrzębski S, Ballas N, et al. A closer look at memorization in deep networks[C]//International conference on machine learning. PMLR, 2017: 233-242. [4] Song H, Kim M, Park D, et al. How does early stopping help generalization against label noise?[J]. arXiv preprint arXiv:1911.08059, 2019. [5] Shen Y, Sanghavi S. Learning with bad training data via iterative trimmed loss minimization[C]//International conference on machine learning. PMLR, 2019: 5739-5748. [6] Chen P, Liao B B, Chen G, et al. Understanding and utilizing deep neural networks trained with noisy labels[C]//International conference on machine learning. PMLR, 2019: 1062-1070.

Dear Reviewer yYJy:

Thank you for your valuable time to review our paper. It has been a while since you last discussed with us. Here we provide a recap to help you read our rebuttal without feeling unfamiliar with this paper.

Your initial concerns can be summarized into two points:

• The relationship between memorization effects studied in previous LNL works and the ICD in our NoiseGPT.
• Lack of discussion on previous LNL works using MLLMs as experts.
• Lack of ablation studies to demonstrate the performance of each component of NoiseGPT and different label corrector backbones.
• There is a bias in ICD scores among different categories.

Our rebuttal carefully addresses these concerns by:

• Elaborating on the essential differences between NoiseGPT and previous LNL works.
• Discussing prior utilization of MLLMs as experts in LNL demonstrating the further insights and superiority of our works, and including them in the related works.
• Conducting experiments to explore the performance of each component in NoiseGPT, including the influence of hyperparameters $n$ and $\tau$, performance comparison with different backbones (CLIP&FLAN-T5-XL), the results are shown in the PDF attachment of our official author rebuttal.
• Since other reviewers also put forward concerns about the influence of bias for different categories in label detection and rectification. We conducted another experiment to further compare the biases between NoiseGPT and Pro-Mix. We calculate the proportion each class takes up in selected clean data on CIFAR sym. 90%. The results are illustrated below. We can see that our method shows a much lower variance, which demonstrates the effectiveness and superiority of our method. | Method\Class index| 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | Variance | | ----------- | ----------- | ----------- | ----------- | ----------- | ----------- | ----------- | ----------- | ----------- | ----------- |----------- |----------- | | Pro-Mix | 9.79% | 11.02% | 9.52% | 7.84% | 10.05% | 8.23% | 10.61% | 11.01% | 10.80% | 11.12% | 1.24 | | NoiseGPT | 10.43% | 10.67% | 8.95% | 10.63% | 10.91% | 9.04% | 10.75% | 8.46% | 9.98% | 10.18% | 0.68 |

It is really important to us that you could kindly read our rebuttal and provide further questions if there are any. Thank you so much and hope you have a good day.

Best, Authors.

Thank you for your suggestions.

• Complementary of noise detection & Baseline of noise detection using CLIP:

  • To directly compare the effectiveness of our noise detection over the baseline methods such as Pro-Mix and DividMix, we conduct experiments on CIFAR-10 sym. 80% dataset and show the results in Table B in the attachment of the author rebuttal. The compared detection scores are from noise detection modules of baselines. Note that the hyperparameters of NoiseGPT are fine-tuned to get the best scores, which validates the effectiveness of our NoiselGPT.
  • Moreover, we have made combination experiments on baselines more than M-correction and Pro-Mix, including DivideMix, Cross-Entropy, etc. Experiments show that all the baseline methods perform better after using cleaner datasets produced by NoiseGPT in most cases, verifying the enhancing effect of NoiseGPT. In this paper, we show two combinations that are assumed effective demonstrations of this property.
  • Meanwhile, there exist cases where complementarity between NoiseGPT and baselines is tested. In fact, MLLM tends to rectify some classes in a dataset more than others, as illustrated in Fig. 6 and Fig. 8. This bias leads to the unbalance in cleaned datasets produced by NoiseGPT, and degrades the performance of classification models in cases like CIFAR-100 sym. 0.2&0.5. To this end, we propose that adjust the threshold τ to prevent overconfident rectifications on clean examples. Results of this method are shown in Figure D in the attachment.
  • To better demonstrate the efficacy of NoiseGPT, we add an experiment where CLIP ViT-L is leveraged as a label rectifier to assign labels to examples considered noisy. The results are shown in Table A in the attachment in comparison with NoiseGPT on noisy CIFAR datasets.

• Experimental setting Webvision:

  • Although it would be more relevant to use the training set of Webvision which contains label noise crawled from the Internet, it would be difficult to analyze the performance of detection and rectification without ground-truth labels as reference. Thus, we leverage WebVision to demonstrate the scalability of NoiseGPT over larger datasets with more classes.
  • For the same reason, there has been hard evidence in experiments indicating diminished detection performance of MLLMs on these datasets.
  • To demonstrate the capability of NoiseGPT to combat noise web datasets, we visualize the distribution of ICD scores on 20000 examples selected from mini-WebVision[1], which is illustrated in Figure B in the attachment. The overall distribution is similar to what is shown in Figure 3, hopefully indicating the separation of clean and noisy examples.

[1] J. Li, R. Socher, and S.C.H. Hoi. DivideMix: Learning with Noisy Labels as Semi-supervised Learning.

Thank you for your suggestions.

• Relationship to previous loss curvature works:

  • The curvature of loss function is a common phenomenon in deep networks, existing works have studied in some fields but none of them have been conducted in LNL. Our work is the first to discover the effectiveness of learning curvature to benefit LNL, which can establish our contribution and novelty in this field.
  • Moreover, existing curvature studies are based on the classification probability, but our ICD measure considers the knowledge of in input prompt by carefully leveraging the in-context learning ability, which can benefit the effectiveness of the MLLM identification and rectification.
  • Furthermore, our ICD score does not need pre-training to be used, we only need to conduct zero-shot inference to implement this score, which is much more efficient and straightforward.

• Specification of model for curvature properties:

  • Our framework which leverages the ICL capability can only be implemented using MLLMs. Maybe in the future when vision models start to show ICL ability, we can also implement NoiseGPT to them.
  • Although the ICD curvature is restricted to MLLMs, smaller models of this kind also share this property, but with a drop in performance. For comparison, we adopt FLAN-T5-XL as another LLM backbone which has a size of 3B parameters and is smaller than FLAN-T5-XXL (11B). The performance of reducing noise is illustrated in Table A in the attachment file of the author rebuttal.

• Missing references and discussions of the related works:

  • Thanks for pointing them out, we will complement all the mentioned references, and a discussion is added in the related work section in our author rebuttal due to the character limitation here.

• Discussion on the scalability to large-scale datasets:

  • The computational cost required for perturbation will not increase as the scale of datasets grows. We select a few exemplars from each class in the dataset to construct a subset D^ex at first. All perturbations of subsequent MLLM input query examples are injected from this subset, and no more exemplars are needed when the scale of the dataset grows.
  • The computing time of the proposed method grows proportionally to the number of examples in a dataset. Thus, when faced with datasets of a much larger scale, we suggest that adjust the hyperparameter $n$ to reduce the number of perturbations for each input example. As the computing time also shrinks proportionally with the decrease of $n$, since $n$ determines how many times MLLM needs to generate an answer to produce ICD scores. By selecting a relatively smaller $n$, the performance of NoiseGPT can be maintained adequately while saving much of computing time.

• Parameter sensitivity analysis & Effect of perturbation numbers and how to reduce it for efficiency:

  • Threshold $\tau$: To further explore the influence of threshold on the overall noise mitigation, we make experiments on CIFAR-100 sym. 20% and 50% with varying τ. Besides, as $\tau$ only affects the noise detection of NoiseGPT, which is a binary classification scenario. We adopt the ROC curve and AUROC score to better evaluate changes in detection performance along with different thresholds, which reflects the changes in TPR and FPR as the threshold moves from the smallest to the largest. By analysing the ROC curve of a small subset of query examples, a suitable threshold can be inferred for the whole dataset.
  • Number of perturbations $n$: Theoretically, more perturbations make the normalized ICD score more robust and effective in distinguishing clean and noisy examples. However, there is a marginal effect when $n$ increases to an extent when it becomes computationally unworthy. In Figure A in the attachment, the sensitivity of hyperparameter $n$ is explored. In practice, we select the number of perturbations $n$ by balancing the computational cost and performance.

• Prompt design: We based our prompt settings on previous works [1] and made some modifications, including the sequence and the content.

  • The proposed prompt includes three sentences, a positive Q&A, a negative Q&A, and a query question whose answer is to be generated. In earlier experiments, we changed the order of the first two sentences which were negative, affirmative, and query. It was observed that MLLM tended to output a lower Softmax possibility of answering “true” for all query examples according to the regularity in former lines of prompts. Thus, we bypass this problem using the ICD score which depends on the relative magnitude of output possibilities between original and perturbated examples, instead of the absolute possibilities of “true” or “false”.
  • In experiments, we extended the prompts to different sizes in the same format shown in Sec. 3.2. And different sequences of these sentences were also investigated. It turned out that the tendency mentioned in the last paragraph was facilitated.
  • To completely mitigate the influence of prompts, we assume that the representation editing [2,3] methods could have possibly been a substitution for existing prompt learning methods. We are still working on this idea.

• Baseline using learning curvature with classifier model:

  • The loss function produced by classification models is related to the discrepancy between the prediction $y’$ and label $y$, which are essentially different from ICD. As a result, it would have been complicated and counterintuitive to investigate in probability changes of DNNs induced by perturbation.

[1] Huang Z, Liu C, Dong Y, et al. Machine vision therapy: Multimodal large language models can enhance visual robustness via denoising in-context learning[C] [2] Liu, S., Ye, H., Xing, L. & Zou, J. In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steering. [3] Kong, L. et al. Aligning Large Language Models with Representation Editing: A Control Perspective.

Dear Reviewer gxfG:

We want to express our appreciation for your valuable suggestions, which greatly helped us improve the quality of this paper. We are also glad that you agreed that our idea is novel and has the potential to be widely used.

We have made our maximum effort to address your concerns on scalability to larger datasets, the sensitivity of hyperparameters, and the unique properties of ICD curvatures in MLLMs, and etc.

Your further opinions are very important for evaluating our revised paper and we are hoping to hear from you. Thank you so much.

Best, Authors.

Thanks for your constructive comments.

• Limited novelty:

  • We make the first contribution to effectively detect and rectify label noise without the need for pre-training, providing a novel direction for learning with label noise, which has never been previously studied (Reviewer 81A3), and has the potential to be widely adopted (Reviewer gxfG). We hope the Reviewer could further provide on which aspect he thinks our novelty is limited, and we would be happy to try our best to address them.

• Missing previous works & several relevant works.

  • We will carefully cite all the mentioned works and discuss them in detail in the related works section. Should there be more references to consider, please feel free to let us know, thanks.
  • We would like to stress that the In-Context Discrepancy is a novel measuring score that is based on contextual prompt information, which can guide and formulate the output. But other measuring score mentioned by the reviewer is mostly based on classification probability. Therefore, our measurement is novel and different from existing works. We will clarify this in the future version.

• Typo.

  • Thanks for pointing them out, we will carefully address all of them.

• Efficacy.

  • By comparing to other well-known baseline methods, our method can further enhance the learning performance in most scenarios.
  • Only under CIFAR-100 dataset with low noise rate, such as 20% and 50%, the performance of NoiseGPT shows degradation. This is because the performance of NoiseGPT highly relies on the noise detection threshold, which is set fixed as 0.72 throughout Table 3. In fact, under low noise rates, the detection threshold should be higher than that of high noise rates, in order to prevent overconfidently selecting clean examples as noisy ones. Our experimental results in section 4.4 validate such a claim, and we further provide an ablation study in Figure D in the attachment file of the author rebuttal to show that when setting proper threshold, the performance of NoiseGPT can still be effective compared to Pro-Mix.
  • In challenging scenarios under high noise rates, NoiseGPT is quite effective in achieving 8.5% and 18.5% performance improvement compared to Pro-Mix under 80% and 90% noise rate, respectively.

• Related work on Sample Cleaning.

  • We have carefully provided the discussion of related works as mentioned in the author rebuttal.

Dear Reviewer TkfP:

We really appreciate your efforts to help improve this paper. We have carefully addressed the mentioned concerns, such as citations of relevant works, and the limitation of improvement in classification performance compared to Pro-Mix. Experiments have also been added to elaborate on these problems.

Having further discussions really helps to achieve consensus and clarify misunderstandings, we are eager to know if there are any remaining problems. We will try our best to address them.

Best, Authors.

During August 7-13 there will be an open discussion between the reviewers and authors. I observed that the author responded to reviewer questions elaboratively, and Reviewer 81A3 have already responded to the rebuttal (thank you!).

Reviewer TkfP, gxfG, QCqA --- Please use this opportunity to further clarify remaining concerns and confusions.

Thank you! AC