Prototype-based Aleatoric Uncertainty Quantification for Cross-modal Retrieval

Cross-modal Retrieval methods build similarity relations between vision and language modalities by jointly learning a common representation space. However, the predictions are often unreliable due to the Aleatoric uncertainty, which is induced by low-quality data, e.g., corrupt images, fast-paced videos, and non-detailed texts. In this paper, we propose a novel Prototype-based Aleatoric Uncertainty Quantification (PAU) framework to provide trustworthy predictions by quantifying the uncertainty arisen from the inherent data ambiguity. Concretely, we first construct a set of various learnable prototypes for each modality to represent the entire semantics subspace. Then Dempster-Shafer Theory and Subjective Logic Theory are utilized to build an evidential theoretical framework by associating evidence with Dirichlet Distribution parameters. The PAU model induces accurate uncertainty and reliable predictions for cross-modal retrieval. Extensive experiments are performed on four major benchmark datasets of MSR-VTT, MSVD, DiDeMo, and MS-COCO, demonstrating the effectiveness of our method. The code is accessible at https://github.com/leolee99/PAU.

Paper

References (66)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer hqQV6/10 · confidence 3/52023-07-03

Summary

The paper focuses on the unreliable prediction of cross-modal retrieval methods caused by low-quality data. The authors introduce evidential theoretical framework to quantify the uncertainty arising from inherent data ambiguity. To provide trustworthy uncertainty predictions, they propose a novel Prototype-based Aleatoric Uncertainty Quantification (PAU) framework. Extensive experiments demonstrate the effectiveness of their methods.

Strengths

- To the best of my knowledge, this paper is the first to combine the evidential framework with cross-modal retrieval methods, so this work is quite novel. - The idea of introducing uncertainty quantification to mitigate the negative impact of low-quality data on retrieval is interesting and sound. The experimental evaluations in the paper also provide strong evidence to support the claims made in the paper. These experiments are wide-ranging and cover several different datasets. - The paper is well written and easy to follow. However, there is room for improvements as I mention below.

Weaknesses

1. While the experiments are quite wide-ranging, the details provided in the paper are quite scarce, such as how many epochs are trained and how to choose hyperparameters. Code is also not provided. 2. Although the authors provide related work on cross-modal retrieval and uncertainty quantification, it is not clear how these works differ from their proposed method. 3. There are some unclear points in the paper. - What is the meaning of the histogram in Figure 1(b)? - Does PAU only contain the component of $L_{uct}$, $L_{div}$ and Re-rank? How does it constrain whether $M''$ is predicted accurately? Providing pseudocode would help readers better understand the proposed method.

Questions

1. What is the effect of using different initialization methods for all prototypes? 2. What is the time complexity of the algorithm? How does it compare to other algorithms? 3. The authors argue that the uncertainty loss forces the prototypes into learning the rich semantics of subspace to realize accurate uncertainty quantification. But I'm confused that the uncertainty loss only limits the consistency of U and H, why it can achieve accurate uncertainty quantification?

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

The authors point out the limitations of their work.

Reviewer b4gq8/10 · confidence 5/52023-07-05

Summary

This paper addresses the aleatoric uncertainty quantification in cross-modal retrieval. The authors point out that the model predictions are often unreliable due to the aleatoric uncertainty, which is induced by low-quality data. They first define the inherent aleatoric uncertainty in multi-modal data. Then K prototypes are constructed as the belief masses to employ DST and SL, two classical theories in uncertainty quantification. On top of this, the uncertainty loss and diversity loss are proposed to help prototypes represent the overall semantic space. Extensive experiments are conducted on multiple tasks and benchmarks.

Strengths

1. This work presents a novel insight for the uncertainty of cross-modal retrieval. They classify uncertainty into aleatoric uncertainty and epistemic uncertainty, and give a clear definition of aleatoric uncertainty for multi-modal data. I think it will inspire further uncertainty-aware work in cross-modal retrieval or multi-modal learning. 2. In Appendix, the proof of “the ambiguous multi-modal data highly associated with multiple semantics” is intuitive and reasonable. 3. The experiments greatly prove the claims, especially Fig. 3 and Fig. 4. These two experiments reflect that the uncertain data indeed have a significant negative impact on results, and PAU could provide more trustworthy predictions. This phenomenon may be able to give an inspiring insight for robust training.

Weaknesses

1. In the reranking process, beta_1 and beta_2 are two learnable parameters. But it is unclear how these parameters are trained in this paper. 2. Prototypes are expected to represent the overall semantics space. Do they really represent the overall semantics space after training? 3. The paper only mentions uct loss and div loss, which seem to be proxy-based loss functions, and can not optimize the prediction directly like cross-entropy loss or triplet loss. I wonder if other basic losses are used in the training process. 4. The Table 7 in appendix shows that the exponential function is the best method to generate evidence. Why there is a large performance gap between ReLu function and exponential function? Or what is the advantages of exponential function. 5. The legend in Figure 1 (b) is empty.

Questions

Refer to weaknesses.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

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

4 excellent

Limitations

Yes

Reviewer tGbU3/10 · confidence 4/52023-07-06

Summary

This paper focuses on improving video-text retrieval by considering the uncertainty (or confidence) when computing image-text similarity. The high-level idea is to estimate the data uncertainty and then perform re-ranking to improve the performance of retrieval. The authors designed a PAU framework, Prototype-based Aleatoric Uncertainty Quantification with CLIP encoders, to estimate the uncertainty score of the input instances. The authors conducted task-specific fine-tuning on the target datasets, and reported marginal performance improvements.

Strengths

- The authors addressed an interesting problem of estimating the uncertainty of data, which is used to improve retrieval performance. - The proposed PAU framework is presented clearly. - Both the uncertainty loss and diversity loss are technically valid and straightforward. - The authors conducted several experiments on multiple benchmarks, and demonstrate marginal performance improvements.

Weaknesses

- The goal and task definition of this paper seem not clear. Both the introduction section and Figure 1 are somewhat confusing, which makes me difficult to get the main point. An image (or a single video frame) should contain rich semantics. Typically, an image is paired with a caption (or sentence), forming a image-text pair, to describe the semantics. However, the authors seem to argue that a single-scene video should match only a single word (i.e., talking), which is counter-intuitive. The multi-scene video may have richer semantics than single-scene video, but the authors argue that matching multiple semantics is not ideal. Overall, their key idea is counter-intuitive. - The authors conducted experiments solely on downstream fine-tuning. The fine-tuning datasets are usually cleaner and with higher quality. A majarity of the noisy image-text paired data should be found in VL pre-training dataset. While downstream fine-tuning data (like youcook2 or MSRVTT) are human annotated/collected, the pre-training data is web-crawled noisy imgae-text pairs. To study the capability of uncertainty estimation for low-quality data, it would be much more interesting to experiment their proposed method in the pre-training stage. The proposed method may have potential to greatly improve the pre-training, but it remains unknown in current stage. - The paper is not written well, making it difficult to understand the motivation. This is probably due to unprecise or uncommon wording in the writing.

Questions

see weakness

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed 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

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

No potential negative societal impact

Authorsrebuttal2023-08-20

Follow-up discussion

Thank you for your valuable feedback on our submission, particularly your suggestion to **train on noisy datasets**. The insightful suggestion enhance the quality of our work and better strengthen our claims. We hope that these improvements will be taken into consideration. If we fully address your concerns about our paper we would be grateful if you could re-evaluate our paper. If you have additional concerns, we remain open and would be more than happy to discuss with you.

Area Chair Q5682023-08-20

Dear Reviewer tGbU, This is another friendly reminder to acknowledge that you have read the rebuttal and the other reviews. Please also share how they change your view on the paper, if at all. Thanks again for your service! Best, AC

Reviewer VBVJ6/10 · confidence 5/52023-07-07

Summary

This paper proposes prototype-based aleatoric uncertainty quantification (PAU) to represent aleatoric uncertainty (i.e., data uncertainty) for cross-modal retrieval tasks. The proposed method uses the CLIP backbone, where the initial similarity matrix M is computed by the output of the CLIP encoders. Then, "evidence" $e$ is computed by the similarity between each modality embedding output and the prototype vectors. Using $e$, the belief $b$ and the uncertainty $u$ is computed. As far as the reviewer understood, the computed $u$ represents how the current embedding is *different from* the prototype vectors, i.e., if the current embedding is specifically different (in terms of the cosine similarity) from the prototype vectors, then its uncertainty becomes large. Using the learned uncertainty $u$, the proposed method re-rank the retrieved items by cosine similarity of the output of the CLIP encoders. Two losses are introduces to make the prototype vectors better, the uct loss (minimize the distance between average embedding vectors and prototypes) and the diversity loss (orthogonal regularization for prototype vectors). The effect of the proposed method is evaluated on four datasets.

Strengths

- Measuring aleatoric uncertainty for multi-modal learning is an interesting and important problem. - The proposed method looks efficient and easily adaptable to various methods. - The proposed method is a somewhat straightforward extension of Subjective Logic (SL) [27] from the fixed class labels to the learnable prototypes. Hence, it may have the good theoretical properties shown by the SL paper [27], although it is not explicitly described in the paper.

Weaknesses

My concerns can be summarized as: - It is not clear that replacing fixed class information of SL with learnable prototype embeddings will ensure a proper uncertainty estimation. - This paper misses the important previous work on the aleatoric uncertainty estimation method for cross-modal retrieval [A]. - The comparisons of MS-COCO 5K would not be fair (only the proposed method uses the pre-trained CLIP model), and many recent state-of-the-art methods are missing. ### Can the proposed method provide a proper uncertainty estimation? The proposed method highly depends on SL [27], where SL uses fixed class information; hence, there is no learnable prototype. As this paper introduces new learnable prototypes, the proposed method requires two additional objective functions, Eq (5) and Eq (6). I am not sure how Eq (5) and Eq (6) will guarantee the quality of the proper uncertainty measure. Eq (6) looks okay, but it enforces the orthogonality of the prototype embeddings, not the mutual exclusion among the embeddings. The latter one would be derived from independence, not orthogonality. Note that independence and orthogonality are highly related and somewhat exchangeable in many cases, I am not sure that smaller orthogonality regularization (i.e., Eq (6)) can ensure the weaker dependency between the prototype vectors. Moreover, I am not sure that Eq (6) is specifically better than widely-used orthogonality regularization, such as $\| I - Z^\top Z \|_F^2$, where $I$ denotes the identity matrix, and $Z$ denotes the full prototype vectors. Another possible alternative is using Maximum Mean Discrepancy (MMD) or Hilbert-Schmidt Independence Criterion (HSIC), whose values directly induce the strength of the dependency between two probabilistic densities. In my opinion, Eq (6) could be better, but Eq (6) itself is okay and can be somewhat sound regularization. To me, the meaning of Eq (5) is ambiguous. It seems that Eq (5) behaves as minimizing the distances between prototypes and the embeddings; thereby, it enforces to make the value of $e_K$ larger (i.e., make the uncertainty value $u$ smaller). I am not sure how Eq (5) helps the proper uncertainty estimation. I think both Eq (5) and (6) can be explained with some in-depth discussions with minimal changes in the manuscript, but as of now, I am not fully confident whether the proposed method can represent proper uncertainty estimations despite of using learnable prototypes rather than fixed class information. ### Missing related works on uncertainty estimation This paper misses many directly related works on uncertainty estimation and cross-modal retrieval. Especially this paper did not consider PCME [A], a probabilistic embedding approach to represent aleatoric uncertainty (i.e., the quality of the data) for cross-modal retrieval. - [A] Chun, Sanghyuk, et al. "Probabilistic embeddings for cross-modal retrieval." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021. I think [A] should be discussed in the related work section and in the experiment section if possible. There are also more aleatoric uncertainty estimation methods based on probabilistic embeddings, such as: - Shi, Yichun, and Anil K. Jain. "Probabilistic face embeddings." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019. - Park, Jungin, et al. "Probabilistic representations for video contrastive learning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022. - Neculai, Andrei, Yanbei Chen, and Zeynep Akata. "Probabilistic compositional embeddings for multimodal image retrieval." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022. - Sun, Jennifer J., et al. "View-invariant probabilistic embedding for human pose." Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part V 16. Springer International Publishing, 2020. I would like to suggest adding discussions of the comparisons of the SL-based aleatoric uncertainty estimation and the probabilistic embedding-based aleatoric uncertainty estimation. ### Missing comparison methods for MS-COCO 5K It could be a minor issue, but I think Table 4 misses many relevant and state-of-the-art methods for cross-modal retrieval. For example, as my previous comment, PCME [A] is a highly related work to this paper. In terms of using multiple diverse embeddings, [B] could be another relevant work to the proposed method, which is more relevant to PVSE [49]. I also would like to suggest adding more recent state-of-the-art cross-modal retrieval works, such as VSE$\infty$ [C]. - [B] Kim, Dongwon, Namyup Kim, and Suha Kwak. "Improving Cross-Modal Retrieval with Set of Diverse Embeddings." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. - [C] Chen, Jiacheng, et al. "Learning the best pooling strategy for visual semantic embedding." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021. Especially, Table 4 can be unfair because all other comparison methods do not use the pre-trained CLIP backbone; hence their performances are much worse than a simple "baseline" of this paper. Therefore, in my opinion, Table 4 should contain the comparisons with other methods using the pre-trained CLIP backbone. However, I fully understand that it is not easy to compare with other cross-modal retrieval methods, especially when they apply different backbones, pre-extracted features, and optimization techniques. After a quick Google search, I found a very recent work, that uses the CLIP ViT-B/32 backbone for cross-modal retrieval: - [D] Chun, Sanghyuk. "Improved Probabilistic Image-Text Representations." arXiv preprint arXiv:2305.18171 (2023). I think the experimental results of [D] can be directly comparable with Table 4 of this paper, including VSE $\infty$, PCME, and InfoNCE. I would like to suggest adding the results of [D] to the main paper if the authors are willing to update their table to contain more recent papers. *Note that [D] is not a direct comparison method of the proposed method because it arxived after the NeurIPS submission deadline*; therefore, I don't argue that [D] is a significant related work that should be compared with the proposed method, but I bring [D] because [D] uses the same CLIP ViT-B/32 backbone with the experiments in this paper, and it reported [A] and [C] with the ViT-B/32 backbone, listed in my review. ### Minor comments - The implementation details of the overall framework are somewhat hard to catch at first glance. I suggest adding the detailed process for computing $e$ and $u$ in Figure 2, and adding the details of $e$ (Table 7 in Appendix) to the main paper.

Questions

Please check my comments in the weakness section. My major concerns are (1) I am not sure whether applying SL to learnable prototypes can lead to proper uncertainty estimations (2) missing significant related works, such as [A] and [B] ([B] could be considered as a contemporary work because [B] is a CVPR 2023 paper, that is published after the NeurIPS deadline).

Rating

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

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

2 fair

Contribution

2 fair

Limitations

I don't think this paper has a potential negative societal impact.

Reviewer qbVB5/10 · confidence 4/52023-07-09

Summary

This paper presents the Prototype-based Aleatoric Uncertainty Quantifications (PAU) framework, which enhances the reliability of a pre-trained vision-language model (CLIP) by leveraging evidential knowledge. The paper introduces learnable prototypes as semantic categories and maps images and texts to these categories for uncertainty quantification. The proposed method improves model training and ranking in image-text retrieval tasks.

Strengths

- The paper addresses the issue of model reliability and provides a reasonable approach to quantify data uncertainty in image-text retrieval tasks. - The introduction of trainable prototypes complements contrastive learning-based vision-language alignment by incorporating evidence knowledge. This leads to improved reliability of the pre-trained model. - The paper provides extensive experimental results and a detailed ablation study, demonstrating the effectiveness of the proposed method compared to existing image-text retrieval approaches.

Weaknesses

- The loss function defined in Equation 5 raises confusion. The uncertainty is computed in the K category domain, while the mean similarity is computed in the N category domain. Could the author explain why these two domains are mapped together? - The paper includes an ablation study on the number of trainable prototypes (K), showing that the best R@1 performance on MSR-VTT is achieved when K=8. Concerning the scalability of the method, would K vary on different datasets to achieve the best performance? - Figure 5 visualizes the uncertainty of videos and captions. However, the connection between the data pairs is not clear. Could the authors explain the rationale behind selecting these specific examples? - The paper lacks a clear introduction to the baseline models, including their training process and model structure. It would be helpful to provide more details in this regard. Also, the reporting of experimental results seems inconsistent or incomplete.

Questions

Please refer to my question in the Weaknesses section.

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

2 fair

Contribution

2 fair

Limitations

N/A

Authorsrebuttal2023-08-20

Follow-up discussion

Thank you for your valuable feedback on our submission, particularly your suggestions to **further explain the Eq. 5** and explore **K vary on different datasets**. These insightful suggestions enhance the quality of our work and better strengthen our claims. We hope that these improvements will be taken into consideration. If we fully address your concerns about our paper we would be grateful if you could re-evaluate our paper. If you have additional concerns, we remain open and would be more than happy to discuss with you.

Area Chair Q5682023-08-20

Dear Reviewer qbVB, This is another friendly reminder to acknowledge that you have read the rebuttal and the other reviews. Please also share how they change your view on the paper, if at all. Thanks again for your service! Best, AC

Reviewer VBVJ2023-08-16

Thanks to the authors for answering my questions. The response looks great to me. Especially, I think the additional experiments for comparing PCME and PAU in terms of uncertainty estimation will make the submission stronger. I am a bit surprised by the additional ablation study result for Eq. 6. It will be great if the authors can add a related discussion in the revised paper (e.g., what could be other options, why the proposed Eq 6 is better than others). Assuming that all the experiments and the discussions covered by the responses will be included in the revised paper, I will raise my score to "weak accept". Also, I strongly recommend including related works (e.g., aleatoric uncertainty estimation methods based on probabilistic embeddings, [A, B, C], and maybe [D]) listed in this comment.

Authorsrebuttal2023-08-17

Thanks to the professional Reviewer VBVJ for the positive response and efforts to improve our work. The results of Eq. 6 also surprised us. We will conduct the experiments to explore the reason, test other methods such as Hilbert-Schmidt Independence Criterion (HSIC), and discuss them in the revised paper. All the supplemental experiments and discussion is helpful to improve our paper, we are willing to contain them in the revised paper. The related references you mentioned are really necessary for our paper, we will add all of them in the final paper.

Authorsrebuttal2023-08-17

Thanks to the Reviewer hqQV for the positive response. All the additional experiments and discussion will be supplemented in our revised paper. If you have any other concerns, we are happy to address them.

Authorsrebuttal2023-08-17

Thanks to the Reviewer b4gq for the positive response. All the additional experiments and discussion will be supplemented in our revised paper.

Reviewer qbVB2023-08-21

Post after rebuttal

Thanks for providing more results. While the intuition behind pushing $u_q$ to $h_q$ remains unclear, the new experimental results partially address my previous concerns, and the method seems to consistently work well on different datasets. Thus, I tend to raise my score.

Authorsrebuttal2023-08-21

Response to Reviewer qbVB

Thanks to the reviewer qbVB for the positive response. As for $u_q$ and $h_q$, we can imagine that a multi-scene video can usually match multiple semantics, such as the video B in Fig. 1(a). It can match to 3 semantics of "talking", "shadow", and "cavern". By contrast, the single-scene video A in Fig.1 (a) approximately only match the semantics of "talking". Thus, the video B will have higher average similarity $h_q$ with textual modality than video A, because it not only have high similarities with the descriptions related to "talking", but also with the descriptions related to "shadow" and "cavern". That is to say, if a video matches more semantics, it becomes more uncertain, because it is harder to be summarized using a simple description. Therefore, the more uncertain the instance, the higher the average similarity with another modality. $u_q$ and $h_q$ should be positively correlated. This is the reason of pushing $u_q$ to $h_q$. Hope this can address your puzzle. All the additional experiments and discussion will be supplemented in our revised paper.

Authorsrebuttal2023-08-21

Response to Reviewer tGbU

Thanks for your efforts in reviewing. We have carefully read your comments and provided detailed answers. Your comments about testing on noisy image-text pairs well encourage us to further consider the application of PAU. Now, we have revised several parts of our paper to make our claim clearer. If possible, we hope you can clarify your remain concerns or puzzles to further improve our paper quality. We are happy to address these concerns.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC