Deep Correlated Prompting for Visual Recognition with Missing Modalities

Large-scale multimodal models have shown excellent performance over a series of tasks powered by the large corpus of paired multimodal training data. Generally, they are always assumed to receive modality-complete inputs. However, this simple assumption may not always hold in the real world due to privacy constraints or collection difficulty, where models pretrained on modality-complete data easily demonstrate degraded performance on missing-modality cases. To handle this issue, we refer to prompt learning to adapt large pretrained multimodal models to handle missing-modality scenarios by regarding different missing cases as different types of input. Instead of only prepending independent prompts to the intermediate layers, we present to leverage the correlations between prompts and input features and excavate the relationships between different layers of prompts to carefully design the instructions. We also incorporate the complementary semantics of different modalities to guide the prompting design for each modality. Extensive experiments on three commonly-used datasets consistently demonstrate the superiority of our method compared to the previous approaches upon different missing scenarios. Plentiful ablations are further given to show the generalizability and reliability of our method upon different modality-missing ratios and types.

Paper

Similar papers

Peer review

Reviewer YNgk5/10 · confidence 4/52024-07-11

Summary

The paper proposes a prompt optimization approach to the missing modality issues in multimodal learning. Inspired by the missing-aware prompt (MMP), this paper adds more prompts, including correlated, dynamic and modal-common prompts, to each encoder to improve the performance. The experiment on three datasets shows the effectiveness of the proposed method.

Strengths

The missing modality issue in multimodal learning is a practical challenge. The designed method is clearly presented.

Weaknesses

1. The novelty of the proposed method is limited since the MMP has proposed the prompt optimization approach to solving the missing modality issue. Compared with MMP, this paper adds more parameters in the form of prompt tokens from different inputs and functions. 2. The empirical comparison with MMP is probably not quite fair as the proposed method uses more additional parameters compared with MMP. According to Line 337, this method adds 2.4% additional parameters, while MMP only adds 0.2%.

Questions

What is the specific contribution of this paper compared with MMP, other than adding more parameters and functions? Will MMP's performance be better or comparable when MMP uses the same number of parameters as the proposed method?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes.

Reviewer BZuF6/10 · confidence 4/52024-07-15

Summary

The model proposes prompting strategy where both modalities (image and text) are prompted, and the prompt for both modalities are correlated. The strategy is to use multiple prompts, namely correlated prompts, dynamic prompts, and modal-common prompts. As the backbone itself is multimodal (CLIP), it is a good idea to consider synchronized multi-modal prompts to fully harness the model capabilities when prompting it. The model surpasses multiple multimodal SoTAs on multiple datasets and also has proven to be effective in handling missing modalities in training and inference.

Strengths

1. The strategy of using multiple types of multimodal prompts, along with the correlation strategy, is logically sound as the multimodal backbone itself is trained to understand the relationship between image and text modalities. 2. The modal surpasses multiple SoTAs on multiple benchmarks with considerable score improvement. 3. The ablation studies are sufficient to understand the justification of the network design.

Weaknesses

1. Ablation studies regarding the multimodal backbone, e.g. using other model than CLIP or use dedicated unimodal encoders for each modality, highly recommended to increase paper quality. 2. In table 4, what are the performances when either image or text modalities are completely missing?

Questions

1. Please elaborate further on how modal-common features are disentangled. 2. If possible, show the layer J-th in Figure 1 (framework overview) 3. Minor suggestion: The phrase "abundant ablations" in the introduction is a bit overboard, I suggest to write it as just "Ablation studies are further given..."

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The limitations are discussed in the appendix, including that only text and visual modalities are tested with this model and the number of tested models.

Reviewer 64zT5/10 · confidence 5/52024-07-16

Summary

This paper addresses the challenge of generalized missing modalities in multimodal learning, where a modality can be absent during any learning phase (e.g., training, testing, or both). he authors investigate prompt learning with missing modalities and propose deep correlated prompts designed to capture various types of correlations between prompts and input features across different modalities. Specifically, the proposed prompts include mechanisms for perceiving beneficial information from preceding layers, dynamically generating prompts based on input characteristics, and leveraging the complementary information from multimodal inputs. These designs improve the robustness of large multimodal models (e.g., CLIP) to missing modalities. Extensive experiments and ablation studies demonstrate consistently superior performance and verify the effectiveness of the proposed method.

Strengths

1. This paper addresses a more challenging missing modality setting, where modalities may be absent during both training and testing phases, making it highly practical and essential for real-world applications. 2. The paper is well-motivated. The authors highlight the weaknesses of prior work and propose several designs (e.g., deep correlated prompts, dynamic prompts, common prompts) to improve robustness. 3. The paper explores various types of correlations between prompts and input features across different modalities, and the proposed designs for each are technically sound. 4. Extensive experiments show great improvement on the baseline and consistently superior performance compared to other methods across all benchmarks. 5. Comprehensive ablation studies are conducted to validate the effectiveness of each proposed component.

Weaknesses

1. The paper lacks a detailed explanation or discussion on the efficacy of different prompt designs. In Figure 2, it shows that sequentially adding different designs improves the baselines, but it does not discuss the individual improvement gains for each design. Additional discussion on each design could help validate whether the increasing gains from sequentially adding designs are not merely due to more learnable parameters. 2. The paper lacks visualization of each learnable prompt (e.g., deep correlated prompts, dynamic prompts, and common prompts). Visualizations could help validate whether the different components work as expected. For example, do dynamic prompts genuinely capture the different characteristics of inputs, or do they merely distinguish between different missing cases, which might be easier to learn due to the obvious absence of a modality? 3. For each available modality, it seems there are a total of $(3*(2^M-1))$ prompts for each missing modality case. This could lead to an exponential increase and redundant prompts as more modalities are considered (i.e., M>2). For example, in a vision-and-language task, in the case of complete and missing-image, the text modality is available for both cases. However, it requires two separate prompt sets for the text encoder, which may actually learn the prompts for the same “text-available” case.

Questions

1. In Table 4, I noticed that some values of the related work MMP are the same as the figures recorded in the paper. For example, the settings with: - missing rate = 70% (100% image and 30% text) in MM-IMDb, - missing rate = 70% (30% image and 100% text) in Hateful Memes, - missing rate = 70% (65% image and 65% text) in Hateful Memes. As far as I know, the MMP backbone model is the multimodal transformer ViLT. The authors state they re-implemented MMP on their setting (i.e., CLIP) for a fair comparison. It seems that the numbers should not be the same since they use different backbone models. Can the authors clarify why the values are identical despite using different backbone models? 2. According to the design of prompts, it seems that the proposed method is not limited to two-stream models (e.g., it could be applied to single-stream models without using Eq. (5)). Generalizing the method to single-stream models and comparing it with related works could be helpful in verifying the generalizability of the proposed method. Have the authors tried it for single-stream models? If so, what were the results? 3. I am willing to revise my rating if the authors also address the concerns mentioned in the weaknesses.

Rating

5

Confidence

5

Soundness

4

Presentation

3

Contribution

3

Limitations

One limitation is that the proposed method requires modality-specific deep correlated prompts for each available modality, which could be challenging to extend to more modalities (e.g., five or more modalities).

Reviewer xVBv4/10 · confidence 5/52024-07-17

Summary

This paper proposes to address the missing modality problem for the multimodal recognition model (i.e. the multi-modal data could be incomplete). There are three techniques of prompting being proposed (while the recognition model, i.e. two-stream multimodal method CLIP in this paper, is kept fixed), including: 1) correlated prompts, where a part of the prompts in the input-level are firstly selected according to the missing scenario (e.g. complete, text-only, or image-only), then the prompt in each of the following network layers are predicted from the multimodal prompt of its preceding layer; 2) dynamic prompts, the input-level prompts contain a portion generated according to the input sample; 3) modal-common prompts, where the rest of the input-level prompts is stemmed from a common component shared across modalities. The combination of the aforementioned three techniques experimentally shows better performance in comparison to various baselines (mainly the SOTA method from MMP [17]).

Strengths

+ The proposed method provides superior performance with respect to various baselines and its proposed techniques (i.e. correlated prompts, dynamic prompts, modal-common prompts) are experimentally shown to benefit the model performance. + The extensive experiments are conducted on multiple dataset with various experimental settings. + The presentation is clear and easy to follow.

Weaknesses

- The modal-common prompts and the dynamic prompts actually are not directly connected to the missing modality problem (or being irrelevant to different cases of missing modality). While excluding these two prompting techniques from the proposed method (in which such variant becomes "Ours (A)" in Figure 2), the improvement with respect to the state-of-the-art approach of handling missing modality (i.e. MMP[17]) would become marginal (please include MMP[17] into the ablation study shown in Figure 2 or directly provide the tabular quantitative results for the ablation study). Similarly, while we only consider the technique of correlated prompts as the manner in the proposed to tackle the missing modality, it becomes the only difference in the proposed method compared to MMP [17] (in terms of methodology), thus leading to the concern of limited novelty. Furthermore, there should be a baseline of integrating the modal-common prompts (acting as a basic component of prompt) and dynamic prompts into MMP[17] to better highlight the contribution of the proposed correlated prompting technique (which is the main technique in the proposed method to be connected with the missing modality challenge). Moreover, as modal-common prompts and the dynamic prompts introduce additional learnable parameters (in comparison the correlated prompts), there should be further detailed analysis/comparison in terms of number of learnable parameters versus model performance. - Though the proposed dynamic prompts do experimentally shown to improve the overall performance under various missing modality cases, such prompting technique is actually not new, where we can see its similar application in various research problems (e.g. Wu et al., IDPG, NAACL'22; Lu et al., PromptPG, ICLR'23; Qiu et al., FedTPG, FL@FM-NeurIPS’23).

Questions

Although currently the proposed method seems to provide superior performance with respect to various baselines and its proposed techniques (i.e. correlated prompts, dynamic prompts, modal-common prompts) are experimentally shown to benefit the model performance, there are concerns regarding limited novelty (where only the correlated prompts are considered to be related to missing modality while the other two techniques, i.e. dynamic and modal-common prompts, are not) and detailed analysis for the number of learnable parameters versus model performance, (as listed in the weaknesses), in which the the authors are highly encouraged to make the corresponding clarifications in the rebuttal.

Rating

4

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

no potential negative societal impact is found.

Reviewer nnhg6/10 · confidence 4/52024-07-17

Summary

This paper proposes a new method to handle missing modalities in visual and language recognition systems. The paper proposes a very similar method to the one proposed by MMP [17] but using different way of getting the prompts to feed them into the transformer layers. Comparison with other works show that the method seems to be effective and some ablations studies are performed to study the different design choices. The method is validated using the most common datasets for this task.

Strengths

- The method seems to work when compared with other state-of-the-art models. - The paper presents results on several datasets and with different settings of the model.

Weaknesses

- The main weakness of the paper is clarity. There are three different sets of prompts that are appended to the intermediate representations. However, the only difference between them seems to be the type of architecture the method uses to compute them. The explanation is very limited and Figure 1 does not illustrate where do these prompts come from. Without the clarity of this explanation it becomes really hard to understand how the motivation of each type of prompt fits the design. What are exactly correlated prompts, dynamic prompts, and modal-common prompts? What make them correlated, dynamic and modal-common? This is not clear in the paper at all. - It is not clear what is baseline. What does dropping features when modality is missing? The input sequence become shorter and coming from only a single modality? If that's the case, what is trainable and what is not? Please explain well this part. I would expect that this baseline is: training with the same number of parameters as the base method, by simply adding learnable prompts at each layer and training using mod-drop (dropping modalities randomly when training, dropping modalities can be done by inputting noise instead of tokens, the average of the modality tokens, zeroes, or not passing the missing modality at the input, it is a design choice that needs to be explained). If it is not what I'm thinking, please explain well, since this is a key experiment. - When comparing with MMP, how did the authors do it? Please explain exactly how was this re-implementation. Also, to be fair, the authors should have applied their method using ViLT instead of CLIP, in that way there is no doubt that this method is better than the very similar MMP. - What is the zero-shot performance of CLIP on these datasets?

Questions

- Please explain well the mechanism of the different types of prompts, input, output at train and test time for each one of them. It could have been done easily with a figure, but at least with a few sentences it could become clearer. - What makes a "dynamic" prompt "dynamic"? - What does baseline mean and how was implemented? - How was MMP implemented on your framework? - What if using ViLT instead of CLIP, would still your method be better than MMP? - What is the zero-shot performance of CLIP on these datasets? it is important since this might be a robust method that does not suffer from missing modality. It can be implemented using nearest neighbor to each of the class embedding using either modality, and combining them when both are present.

Rating

6

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

Limitations have been addressed.

Reviewer nnhg2024-08-11

Thanks to the authors for the thorough reply. I think the figure makes the contribution of the paper way more clear, and I suggest that it is included in the main manuscript. It is still not 100% clear to me what is baseline for you. I apologize for not being clear with my question in the first place. **Fine-tuned CLIP** in the last Table that you showed in the rebuttal is the same baseline model that you used for Figure 2 and Figure 3 in the main paper? Was this baseline trained with missing modalities at train time (as a sort of "augmentation"), i.e. modality dropping at train time? That's my only remaining question, in order for me to give my final evaluation on the paper.

Authorsrebuttal2024-08-11

Response for Reviewer nnhg

Many thanks for your reply. We will try to save space to include the figures in our manuscript. For the baseline, the Fine-tuned CLIP in the last Table in the rebuttal is the same baseline model that we used for Figure 2 and Figure 3 in the main paper. The baseline is trained with the same missing modalities at training time to keep fair comparison with other methods. We also keep the identical training settings for other methods to make a fair comparison.

Reviewer nnhg2024-08-11

Authors have addressed most of my concerns. I still think the paper could be clearer at explaining the different model components, and be more thorough when explaining the different setups for baselines and other methods. The included experiments in the rebuttal helped clarifying most of the doubts. However, the paper will need a substantial effort at explaining better all the unclear points that I and other reviewers have raised. I will increase my score to 6, and I hope the authors work on the manuscript to improve its clarity and especially its reproducibility in terms of experiments, experimental setup, baseline definition and definition of model components. Thank you for your great work in this rebuttal.

Authorsrebuttal2024-08-11

Response for Reviewer nnhg

Many thanks for your careful comments. We will incorporate detailed explanations in the manuscript to clarify the improvements of our method compared to the baseline and existing methods. We will also make the code publicly available.

Reviewer BZuF2024-08-12

Thank you to the authors for addressing the questions. The authors have addressed most of my concerns, and I hope the answer can be further added in the final version of the paper either in the main paper or in the appendix. Overall, I stand with my initial rating of Weak Accept as the authors have demonstrated that their proposed method is competitive with logical explanation and proof to back it up.

Reviewer YNgk2024-08-13

Thanks for the response. My concern about the experimental comparison is addressed, so I will increase the score to 5. The novelty concern still holds but not a major flaw as the novelty concern is possibly not quite objective as it always is, so I wouldn’t fight for rejection.

Authorsrebuttal2024-08-13

Response for Reviewer xVBv

Dear reviewer, thanks for your time and efforts in reviewing our manusript. We have provided a point-to-point response regarning your concerns and we are looking forward to receiving your valuable feedback on the points we addressed in the response. If you have further concerns, place let us know and we will respond to you as soon as possible.Thank you for your dedication to the review process. Sincerely, Authors

Authorsrebuttal2024-08-13

Response for Reviewer 64zT

Dear reviewer, thanks for your time and efforts in reviewing our manusript. We have provided a point-to-point response regarning your concerns and we are looking forward to receiving your valuable feedback on the points we addressed in the response. If you have further concerns, place let us know and we will respond to you as soon as possible. Thank you for your dedication to the review process. Sincerely, Authors

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC