Learning Sample Difficulty from Pre-trained Models for Reliable Prediction

Summary

It is known that state-of-art deep-network based machine learning models are poorly calibrated. It is also somewhat widely known that the poor calibration is because we do not model data uncertainty while training. This work proposes to use CLIP pretrained model for estimating the sample difficulty, which when further used for modeling data uncertainty is shown to yield better calibrated models. Relative Maholonibis distance and leveraging a multimodal pretrained model such as CLIP for estimating the sample difficulty is the core contribution of this work.

Strengths

- Convincing experimental study. All the components of their proposed method: CLIP, RMD, alpha (hyperparameter) are well argued. - Strong evaluation. The proposed method is evaluated with positive results on various downstream applications: outlier detection, selective classification, ECE on OOD. - Writing is easy to follow.

Weaknesses

*Limited originality*: I enjoyed reading the paper but it cuts very close to the existing work. As mentioned in the paper, several previous methods attempted example weighting during training [39] based on their difficulty, and RMD is proposed and used earlier for outlier detection [45] (citation numbers from their paper). As I see it, the originality of the paper lies in using RMD and pretrained models for sample weighting. Although the paper is technically strong, I am somewhat unimpressed by its originality. *Universality of CLIP*. The proposed measure of sample difficulty is in essence how well CLIP can identify different classes. Although CLIP may have been trained on large and diverse data, it surely has its limits. We may then also need an additional measure that informs how difficult it is for the CLIP to provide a sample diffcilty for an example. For instance, the proposed measure most likely cannot be useful on a chest xray dataset for a downstream task of classifying pneumonia or normal. Comments on how such scenarios can be addressed are missing from the paper.

Questions

- Is the base model used for training on downstream models (which is mentioned as ResNet34) is pretrained? - What happens when CLIP itself is used as the base model? - Can some justification be provided on why we see better corruption robustness in ECE in Fig. 3? I would also appreciate if the authors can respond, if they can, to the weaknesses. *Minor comments*. - L171-182. I found the writing somewhat superfluous since we only care to see that the error rate dips as we move towards right in Fig 2. Perhaps there is no need to comment about why ViT is the best etc. Also, there seems to be some confusion between the text and Fig.2. Text interprets the y axis as accuracy while it is error rate in the figure. - L310: last sentence seems incomplete.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed 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

Yes. Edit after rebuttal: Please see my comment in the thread below.

Summary

To address the over-confidence problem of the uncertainty estimation in deep learning, this paper for the first time proposes to use the pre-trained large models to estimate the learning difficulty of each sample. Then, the estimated sample difficulty information is embedded in the final loss for training deep models. The authors evaluate the proposed framework on different datasets and with different model architectures, showing that it improves uncertainty calibration on both ID and OOD settings.

Strengths

1. The studied problem namely uncertainty calibration is important in modern deep learning. The presentation of this paper is good and the paper is easy to follow. 2. Different from some previous works that use pre-trained large models as the base models to fine-tune, this paper proposes to use large models as the tools to estimate sample difficulty, which makes the training process flexible since any model architecture can be used in the training. 3. I like the idea of using sample difficulty as a metric to guide the modeling of uncertainty. And based on the authors' observation, large models (especially the models learned with cross-modality) seems give more accurate estimation of sample difficulty.

Weaknesses

1. The proposed distance metric RMD in Eq. (5) is shown useful than MD in Fig. 6, but backs of the theoretic supports. 2. Two new hyperparameters ($\alpha$ and T) are induced in the final loss function (9), used as the weighting factors of the entropy penalty term, which act like the factor used in label smoothing and may reduce the robustness of the proposed method. 3. I wonder if the large model necessary for the estimation of sample difficulty? There must be some other metrics for sample difficulty, so if other metrics also yield good performance with the regularization-based loss function (9). Maybe some ablation studies can be conducted for this point.

Questions

1. If no theoretic support could be added, could you explain more about why RMD is better than MD. 2. Maybe experiments of hyperparameter study can be added for the second weakness.

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

4 excellent

Contribution

3 good

Limitations

N/A.

Summary

The authors propose to improve model calibration by leveraging information about a sample's difficulty. To do this, they cluster samples using embeddings obtained from large, pre-trained models, and then use a sample's distance to samples from the same class as proxy for difficulty. They show improvements on imagenet in calibration and slight improvements on accuracy.

Strengths

Originality: The approach is effectively a form of adaptive entropy regularization. The novelty is that the strength of the regularization depends on the "difficulty" of the samples. The approach makes sense and is (to the best of my knowlege) novel. Quality: Theoretical analysis is a bit lacking. Empirical results make sense and show a decent improvement in calibration scores. As such, the method does make sense. Clarity: Some choices the paper makes are unclear to me, both in terms of overall method design (why focus so much on RMD?) and in details (e.g. formulation of Eq. 10). See "details" for more. As far as language use goes, the paper is easily understandable. Significance: Results show that the idea works. ECE decreases, and the qualitative examples shown in the figures/supplementary make sense to me.

Weaknesses

* The main idea of the paper (estimate difficulty of samples on large pre-trained model, weight regularization with that) is a bit obscured by the fact that the authors make a big fuzz about using the "Relative Mahaloni Distance" (RMD), even though Table 7 shows that even doing simple K-Means clustering works as well (even better than normal Entropy Regularization, if one compares with Table 4). The paper's main idea would be much more obvious if the paper first established that the idea works even using a simple clustering like K-Means, and then shows that RMD is a sensible improvement upon this. As such, I don't really understand why RMD was chosen: The paper also never goes into explaining WHY RMD beats K-Means, or why doing Mahaloni Distance (MD) does not work as well. The authors merely say that RMD worked well for near-OOD detection.

Questions

* Why take the maximum over i in Eq. 10? * Table4: misses a description of what the columns represent. The text says that they are different regularization strenghts, but this should also be made explicit in the table itself. * Authors state that all results are averages, but never talk about the variances between runs. It would be nice to know if error bars overlap. * What is the runtime of calculating RMD? It would be nice to know if this method is applicable in practice or is prohibitively expensive. * How well does the Gaussian assumption fit? An exploration of how well the RMD describes the actual representation space would be interesting.

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

Authors do not adress limitations of their approach, and I'd encourage them to add a discussion of those!

Summary

This paper proposes a difficulty-aware uncertainty regularization approach, which first pre-defines the difficulty of each training sample and then differently regularizes training samples during training. To quantify the sample difficulty, the authors utilize pre-trained large models like CLIP to extract feature vectors, and the relative Mahalanobis distance is computed on a collection of these feature vectors.

Strengths

1. The paper is well-written and exhibits clarity in its presentation. The visual results help to comprehend the concept of sample difficulty. 2. The effectiveness of considering sample difficulty in classification tasks has been empirically verified against other uncertainty regularization methods. In particular, the ablation study conducted on the regularizing strength hyperparameter $\alpha$ in Section 5.4 clearly demonstrates that the proposed sample-specific regularization approach outperforms ER, which performs global regularization.

Weaknesses

1. The current version of the paper solely presents the average value obtained from five trials without including information about the standard deviation. It is highly recommended to include error bars. 2. The proposed methodology heavily depends on CLIP as a key component. Specifically, when using ViT-B and MAE-ViT-B, the accuracies achieved in Table 5 are 73.59% and 73.81%, respectively, which are lower than the accuracy of 74.02% achieved by the Poly baseline presented in Table 1. At this stage, it is challenging to assert that the proposed method can be applied universally to any large-scale pre-trained models.

Questions

1. The overall scheme that a pre-trained model provides supervision for a given data reminds the knowledge distillation. However, the authors stated "knowledge distillation will not solve the problem" in lines 29-30. Could you make an additional statement on this? 2. The reason for the absence of the CRL as a baseline in the main tables, specifically Tables 2 and 3, is unclear or not readily apparent. 3. Is there a specific justification for not including C10 in the misclassification detection experiment? Miscellaneous: 1. Section 5.4. "roubustness" should be "robustness." 2. Table 4. "under different on ImageNet1k" should be "under different $\alpha$ on ImageNet1k." 3. It would be nice to update the figures in the paper to utilize colors that are accessible for individuals with color blindness. Specifically, figures 4, 7, and 11 might pose challenges for individuals with red-green color blindness.

Rating

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

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The authors did not address the limitations.

Summary

In settings where deep neural networks are used for critical tasks, it is crucial to ensure that they are calibrated and capable of reliable predictions. It is desirable to have the ability to measure the confidence of the model's predictions and reject those that have high uncertainty. To achieve this, the authors suggest using neural networks that have been pre-trained on large and diverse datasets as calibration auxiliaries for downstream tasks. Through extensive experimentation on benchmarking datasets such as ImageNet, the authors demonstrate that using pre-trained models such as CLIP to define the relative Mahalanobis Distance in feature space is a valid measure of sample difficulty. Additionally, they introduce a novel regularization term that accounts for the difficulty of individual samples. In section 5, they provide an extensive evaluation of their proposed learning objective in comparison to other standard regularizers. While difficulty-aware regularization can be somewhat costly, it ultimately reduces the expected calibration error.

Strengths

Some strengths of the work: **Originality**: Researchers are actively studying how to measure the difficulty of individual instances in machine learning to improve its reliability. The authors have introduced and proven the effectiveness of RMD as a reliable method for measuring this difficulty. **Clarity**: I found the paper well-written and straightforward to understand. The authors have provided ample empirical evidence to support their idea of sample difficulty, and they have included a comprehensive discussion while comparing it to relevant baselines. **Quality**: Understanding the limitations of existing algorithms on complex problems requires measuring sample difficulty. This research introduces innovative regularizers that enhance the models' calibration in downstream training. **Significance**: Alongside establishing robust benchmarks and regularizers for calibrated training, the authors also find that self-supervised learning algorithms (such as MAE) learn richer representations that better estimate hardness.

Weaknesses

1. When it comes to measuring sample difficulty, the effectiveness of using large-scale pretrained datasets depends on the degree of domain/distribution shift in the downstream task. For example, it remains uncertain whether the same measures can be applied to medical imaging tasks as compared to natural image classification.

Questions

1. The authors explore example difficulty in the context of classification tasks. Can similar measures be extended to dense-prediction tasks?

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

2 fair

Limitations

It's important to mention that measuring the difficulty of samples adds extra computing workload. To provide clarity to readers, the authors are encouraged to discuss this topic and any potential limitations of their proposed work.

Looking forward to further feedback

Dear Reviewers, Thank you again for your thorough reviews. We are glad that both Reviewer 93FC and Ye8r found our rebuttal satisfactory and increased the score accordingly. We invite the remaining three reviewers (FmxN, YeMZ, Lxfs) to check our rebuttal. Please kindly note that the discussion phase is close to the end (Aug 21st). If you have any further concerns or requests, we would be very happy to address them in the author-reviewer discussion period. If all your concerns have been resolved, it is much appreciated if you may raise the rating of our work. Best, The Authors

Thanks for the responses.

Thank you for your responses. I have thoroughly reviewed the feedback, including the additional experimental results and the illustrative figure provided in the attached rebuttal PDF. The results from the ablation studies align with the stated claims. However, I still maintain concerns regarding the issue of hyperparameters. Given the various choices available for the combination of $\alpha$ and $T$, and the uncertainty regarding the method's sensitivity to $T$, it remains challenging to definitively assert that larger models consistently outperform others. As the utilization of larger models stands as a pivotal contribution within this paper, I am maintaining my score.