RegBN: Batch Normalization of Multimodal Data with Regularization

Recent years have witnessed a surge of interest in integrating high-dimensional data captured by multisource sensors, driven by the impressive success of neural networks in the integration of multimodal data. However, the integration of heterogeneous multimodal data poses a significant challenge, as confounding effects and dependencies among such heterogeneous data sources introduce unwanted variability and bias, leading to suboptimal performance of multimodal models. Therefore, it becomes crucial to normalize the low- or high-level features extracted from data modalities before their fusion takes place. This paper introduces a novel approach for the normalization of multimodal data, called RegBN, that incorporates regularization. RegBN uses the Frobenius norm as a regularizer term to address the side effects of confounders and underlying dependencies among different data sources. The proposed method generalizes well across multiple modalities and eliminates the need for learnable parameters, simplifying training and inference. We validate the effectiveness of RegBN on eight databases from five research areas, encompassing diverse modalities such as language, audio, image, video, depth, tabular, and 3D MRI. The proposed method demonstrates broad applicability across different architectures such as multilayer perceptrons, convolutional neural networks, and vision transformers, enabling effective normalization of both low- and high-level features in multimodal neural networks. RegBN is available at \url{https://github.com/mogvision/regbn}.

Paper

Similar papers

Peer review

Reviewer 6Kg35/10 · confidence 3/52023-07-03

Summary

This paper focuses on mitigate the influence of unwanted variability and bias in multimodal data. To this end, the authors proposed a novel non-parametric regularization technique, i.e., RegBN. The proposed method utilizing Frobenius norm to constrain cross modality consistency, enabling model to discern underlying patterns. Comprehensive experiments show the usefulness of proposed approach in various datasets from different research areas.

Strengths

1. Paper is well-written and follows a good structure. 2. Comprehensive performance evaluation are shown, involving various recent multimodal methods and seven datasets, that proposed approach is working in practice. 3. Mitigating unwanted variability and bias of multimodal data is an important topic and few methods have touched this field. The proposed method is interesting. Improving the reliability of multimodal learning by normalizing multimodal feature is a new perspective.

Weaknesses

Major - Method: The proposed method mitigates the influence of unwanted variability and bias by minimizes the linear relationship between multimodal layers. In Figure 1, the authors claim that "by leveraging RegBN, these data sources can be rendered independent, thereby enabling a multimodal network to discern underlying patterns and optimize its performance". To the best of my knowledge, exploring consistency and complementarity[1] may be two mainly underlying reason why multimodal model outperforms the unimodal. Does the proposed method ignore to explore the cross-modal consistency? - Analysis: In line 108, the authors claim that "The goal is to find potential similarities (mainly caused by confounding factors and dependencies during data collection)". In my view, this is an essential assumption for this paper. I think the authors should provide more discussion about why potential confounding factors and dependencies cause these similarities. It could be necessary to give some explanation and citations. - Comparisons: In line 33 "For instance, the race or gender of speakers in audio classification, backgrounds in video parsing, or the education level of patients in dementia diagnosis is commonly recognized as confounders.". This statements seems very closely to the motivation of DRO[2]. It is appreciated to give some discussion about the relationship between RegBN and DRO. Minor - Ambiguous notations: In line 101 and 102, what is the meaning of N and M? [1] Deep Partial Multi-View Learning, Changqing Zhang, Yajie Cui, Zongbo Han, Joey Tianyi Zhou, Huazhu Fu and Qinghua Hu, IEEE Transactions on Pattern Analysis and Machine Intelligence (IEEE T-PAMI) [2] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781–6792. PMLR, 2021.

Questions

Please address the issues raised in the weakness section.

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

3 good

Limitations

Please address the issues raised in the weakness section.

Reviewer JscA5/10 · confidence 3/52023-07-05

Summary

This paper introduces an approach for the normalization of multimodal data called RegBN. The RegBN aligns multimodal features by a learnable projection matrix with regularization in the form of the Frobenius norm. The regularization strength is updated with the L-BFGS optimization algorithm. Extensive experiments on various modalities are conducted to verify the method.

Strengths

- The paper is well-written and easy to follow. - Aligning multimodal data from the perspective of the normalization method can be promising. - Extensive experimental results show the effectiveness of the proposed RegBN.

Weaknesses

- The method is not very novel for the reviewer. Essentially, RegBN is an l2 regression with regularization in the form of the Frobenius norm on weight. Moreover, L-BFGS has been widely adopted to update the regularization strength. - From the extensive experiments, RegBN is actually a multimodal feature fusion module. Other than baselines of normalization methods such as MDN and PMDN, more comparisons with other multimodal feature fusion modules such as [1] and [2] should be provided. - Why the method is called RegBN. It is not clear how RegBN is connected with the normalization method. [1] Yikai Wang et al. Learning Deep Multimodal Feature Representation with Asymmetric Multi-layer Fusion. 2021 [2] Yikai Wang et al. Deep Multimodal Fusion by Channel Exchanging. 2020

Questions

Please see the weaknesses. Overall, this paper proposes a simple but effective scheme to better fuse multimodal features. The technical contribution is limited. The experimental results are comprehensive and convincing. It would be better to provide more comparative experiments as suggested.

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

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Yes

Reviewer saMu7/10 · confidence 4/52023-07-06

Summary

The paper introduces a novel approach called RegBN for normalizing multimodal data before fusion in neural networks. The integration of heterogeneous multimodal data poses challenges due to confounding effects and dependencies among different data sources, which can introduce variability and bias. RegBN addresses these issues by using the Frobenius norm as a regularizer term. The method generalizes well across multiple modalities and eliminates the need for learnable parameters, simplifying training and inference. The effectiveness of RegBN is validated on eight databases from various research areas, covering diverse modalities. The proposed method shows broad applicability across different neural network architectures, enabling effective normalization of both low- and high-level features in multimodal models. Overall, RegBN offers a promising solution for improving the performance of multimodal models by addressing normalization challenges.

Strengths

The paper presents a novel idea regarding multi-modal fusion. The proposed method is model-agnostic, i.e., the authors provide details about how to apply RegBN within different frameworks, including late fusion, layer (intermediate) fusion, and early fusion. This method does not require significant modification of the original framework and can be plugged in regardless of the encoder's architecture. Extensive experiments are conducted, including eight datasets covering applications in multimedia, affective computing, healthcare diagnosis, and robotics. On all datasets, the proposed method outperformed the considered baseline: vanilla model, or model with normalization in other manner such as BatchNorm or PMDN. This work provides a relatively unique exploration of the topic of multi-modal learning which I believe is worth presenting to the community.

Weaknesses

Authors conducted experiments between with/without RegBN and also compared with other normalization methods. It will be interesting to see how does RegBN help regardling of the problem of imbalanced modality utilization, as pointed out in recently papers [1, 2, 3] [1] Wang, Weiyao et al. “What Makes Training Multi-Modal Classification Networks Hard?” 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019): 12692-12702. [2] Huang, Yu et al. “What Makes Multimodal Learning Better than Single (Provably).” Neural Information Processing Systems (2021). [3] Wu, Nan et al. “Characterizing and overcoming the greedy nature of learning in multi-modal deep neural networks.” ICML 2022.

Questions

- Typo in the caption of Figure 3. Subfigures are indexed incorrectly. - The authors should remember to correct the misstatement in the main paper about SMIL for MM-IMDb.

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

Authors mentioned that the current algorithm works only on single GPU.

Reviewer eB1v5/10 · confidence 4/52023-07-12

Summary

This paper proposes RegBN for the normalization of multimodal data. RegBN uses the Frobenius norm as a regularizer term to address the side effects of confounders and underlying dependencies among different sources of data. The proposed method generalizes well across multiple modalities and eliminates the need for learnable parameters, simplifying training and inference. Experiments on eight databases from five research areas demonstrate the effectiveness of the proposed method.

Strengths

- This paper proposes a novel RegBN method for the normalization of multimodal data. - Extensive experiments demonstrate the effectiveness of the proposed method.

Weaknesses

- My main concern is that the proposed method lacks comparison with other competitive normalization methods such layer norm and group norm. - Using eight databases from five research areas is not necessary. The authors should focus on two or three datasets. - Synthetic multimodal dataset should be provided to thoroughly evaluate the effectiveness of the proposed method.

Questions

See weaknesses.

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

3 good

Contribution

3 good

Limitations

See weaknesses.

Reviewer Fqi76/10 · confidence 3/52023-07-29

Summary

The paper presents a novel regularization approach tailored for multimodal/heterogenous data fusion, which can be applied off-the-shelf to modern neural network architectures. The approach, namely RegBN, targets the reduction/removal of confounding factors and partial dependencies that are usually present among features extracted from different modalities. To do so, RegBN is implemented as a regularization technique that promotes feature independence across data modalities. The regularization is iteratively optimized over mini-batches using a separate solver not relying on back propagation and gradient descent. This separation is critical as otherwise the network could simply learn to use the confounding factors instead of removing them. Experiments are performed on several multimodal datasets based on different modalities and spanning across different applications, such as multimedia, affective computing, healthcare diagnosis, and robotics. Moreover, a synthetic dataset is designed to further highlight the impact of RegBN. The results showcase a positive impact in all the test cases.

Strengths

1. The approach is novel and grounded with theoretical motivation 2. The method can be plugged off-the-shelf to most multimodal architectures, possibly leading to vaste adoption 3. Wide breadth of experiments that explore several multimodal scenarios 4. Positive results in all scenarios

Weaknesses

1. No integration with recent widely adopted multimodal foundation models, such as CLIP, ImageBind, and many others. Despite I understand that RegBN purpose contrasts with the multimodal alignment objective driving those models, I believe it would be valuable to bring them to the discussion. It would also be beneficial for the positioning of the paper. 2. The boost is sometimes minor, e.g see Table 2, 5, I.3, and in some settings even slightly negative, e.g., see Table 5 and I.4 (video fusion rows). Also, it is unclear the high difference between AV-MNIST results reported in the main paper (~ 99% accuracy) and the ones reported in table I.3 (~71% accuracy). Furthermore, why for the experiments on LLP, training metrics are reported in Table 2 and validation metrics in the supplementary? 3. As RegBN requires a separate optimization, its integration affects the time complexity of the model training. The Authors provides multiple examples on how RegBN could/should be integrated in existing architectures, see Figure I.1, however, in most cases a single RegBN layer is added to the network. It is unclear whether this choice is suboptimal for performance in favor efficiency, or whether it is useless to have multiple RegBN as in I.1c. 4. Minor: 1. The writing is for sure enjoyable, however there are some parts that could have better explained / could be less ambiguous. For instance, in section 3: _"Ideally, the residual layer ... RegBN minimizes the linear relationship between these layers via: equation (2)"_, but then equation (2) does only contain $f$ and $g$, and thus it is completely tacit the key insight on why solving equation(2) would provide $f_r$ orthogonal to $g$, as well as it is not explicit that $f_r$ corresponds to the normalization of $f$. Moreover, Figure 1 provides a general idea of the impact of RegBN, but it could have been more aligned with the formulation, e.g., by adding $f$, $f_r$, and $g$; and in general the caption could be expanded to better describe it — what does the grey band represent?. Finally the positioning could have been more clear by stating why recent multimodal foundation models like CLIP are out of the scope. 2. “multimodAl” often mistyped as “multimodEl”; Table I.5, 3rd column, video fusion, (70.6) vs (81.4) seems a bit suspicious; Color gradient is not explained in Figure I.9; wrong x axis label in Figure I.10;

Questions

1. Do the Authors envision any use/application of RegBN for large foundation models, e.g., is it meant for whenever they are employed for downstream tasks? 2. What could be the reason why RegBN affects negatively performance? Is it due to the fact that when evaluating some dataset in-distribution learning spurious correlation might provide a boost? Please comment also on the last two sentences of weakness # 2. 3. Is it enough to apply RegBN once as shown in I.1a,b,d or might be worth to apply it to several layers? 4. Consider slightly modifying sec. 3 and Figure 1, as well as adding somewhere (introduction or related work) a few positioning sentences wrt foundation models. Fix other typos/mistakes.

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

Societal impact is not discussed. I believe that RegBN could provide a benefit in reducing stereotypical biases that currently affect multimodal models.

Reviewer JscA2023-08-19

Thanks to the authors for the detailed response. The rebuttal addresses my concerns properly. I would like to increase my score to 5.

Reviewer Fqi72023-08-19

Thanks to the Authors for the additional clarifications. I confirm my score

Reviewer 6Kg32023-08-20

Thanks for the authors' response, my main concerns about the motivation and underlying reason of effectiveness has been addressed. Taking other reviewers' comments into account, I tend to increase my score to 5/10.

Reviewer saMu2023-08-20

I appreciate the authors in especially extending the experiments and I am very excited to see the results presented. Given my expertise and experience in doing very relevant research on the same topic, I highly recommend this paper be accepted.

Area Chair kE7g2023-08-20

Dear authors, Thank you for providing the response. I acknowledge that this response has been read and will be fully considered. Regards, AC

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC