NeurIPS 2020

Deep Multimodal Fusion by Channel Exchanging

Review 1

Summary and Contributions: This paper examine the idea how to fuse multimodal inputs by interleaving channels of two networks by replacing dead-channels (i.e. when the BN param \gamma = 0) by the other network activations. After providing some theoretical insight, they evaluate their approach on semantic segmentation with a RGBD inputs and Image-to-Image translation. [Note: the reviewer is experienced in multimodal learning, and visually grounded language tasks. The reviewer is only familiar with the RGBD literature.]

Strengths: The current approach moves away from the classic late-fusion paradigm in deep learning (i.e. fusing modalities in a final neural block) by interleaving the channels of the different modalities at different levels. On a personal note, I think this is a promising line of research, and the proposed method is quite simple: 1) create sparse activation by using a L1 norm over the batchnorm scaling factor 2) replacing the activations. It is pleasant to see a simple and potentially working method. However, as discussed in the "Relation to Prior work", this line of research share many similarity with Conditional Batch Norm, and it is very surprising that no links are drawn between methods. After looking at the NYUDv2 literature, it seems that the table2 results are quite strong. Why are there two columns Commonly setting vs Same with our setting, while it is said l201 that you follow the standard setting.

Weaknesses: The authors compare their work to concatenation and alignment baselines. However, I am missing additional method-based related benchmarks. For instance, another mid-fusion baseline (or ablation) could be: - summing activation (only keep the second part of Eq6) - summing activation and adding a soft/hard gating mechanism - performing CBN \gamma = gamma_m1 + gamma(f(x_m2)) - turning the mixture of expert into a fusion block In the end, this paper shows the interest of performing non-naive mid-fusion, but some fundamental part of the approach remains obscure. Can we exchange channels without l1-normalization (e.g. thresholds) ? How does it compare with the modulation approach? Can we unshare the convnet? Why reglazing on half-of the params? As discussed below, the theoretical motivation has a few weaknesses, and it seems to not correlate with the experimental setting.

Correctness: In Table1: you show that sharing the convolutional parameters provide an important boost. In the studied tasks, it makes sense as RGB and Depth channels may share many similarities, and this can be done. What happens if you do not share the convnet but perform Channel exchanging? Do you observe an alignment of the channels? If yes, what is the final performance, if no, can you elaborate on this strong limitation. In Table4: how does it compare with concurrent methods of the literature? The author tried to provide a theoretical intuition of the proposed method. - Theroem1: I am not sure to understand the interest of this theorem. By construction, a lasso-regularisation over the parameters make the parameters sparse. Why demonstrating it? A simple reference to The Elements of Statistical Learning would have been enough. - Corrolarry1: I am very skeptic by both the underlying concept and the proof; Adding channel exchanging would reduce the loss. Why designing a specific f'_{1:M} is enough to prove the corollary. If we define f'_{1:M} to be a constant function equals to 1e6, then the loss should be higher. Besides, Table6 in the appendix shows that exchanging features on all the channels is less efficient than exchanging half-of-the channels. (Note that this is not due to the l1 regularisation as observed in the same table).

Clarity: Overall the paper is not hard to follow. My main concern is that there are multiple unjustified claim (l35-40), l60. I am quite concerned with the claim that the method can be extended to other modality l174. As language networks are based on RNN (with no batchnorm) and transformers, why do you think it would work?

Relation to Prior Work: There is a large portion of the multimodal learning literature based on modulation (Dumoulin et al. 2018) that is never mentioned in this paper. This is even more surprising as the authors refer to instance normalization, which was among the seminal work for modulation applied to style-transfer. Besides, Conditional Batch Norm (DeVries et al. 2017) have clear links and motivations with the current paper: this is critical missing reference. In CBN and its variant, the parameters of the batchnorm are conditioned on the other modalities. I would recommend updating the paper with the following references: - l34: early/late fusion is a pretty old concept (which is ill-defined in deep learning). You can find different definitions with: (Atrey et al.,2010; Bruni et al., 2014; Hall and Llinas, 1997; Snoek et al., 2005). The standard definition is often taken by Snoek et al. (2005). There have been new definition in the deep learning literature (Kiela 2017, Lazaridou et al. 2015, DeVries et al. 2017). In any case, two recent citations is not enough to define early/late fusion :) - Introduction 3rd paragraph: multiple claims are not supported by the literature, e.g, "aggregation-based fusion is prone to underestimating ... have been aggregated" is not justified. Same remark l39-40. - Related work: the authors speak about aggregated fusion and alignement based fusion. This taxonomy was defined by (Baltrušaitis et al. 2019), and it is clearly a missing citation - There exist other work that examines the impact and interest of batchnorm parameters,and how to handle dead channels such as (Wenqi et al. 2020), which is related to the current approach - the exchange of modality has also been explored with other modalities. cf Baltrušaitis et al. Section 7.1 paragraph2, but none of these methods are mentioned in the paper Overall, I strongly encourage the authors to frame their work in the current literature better, as several key references are missing. Dumoulin, Vincent, et al. "Feature-wise transformations." Distill 3.7 (2018): e11. C. G. M. Snoek, M. Worring, and A. W. M. Smeulders. Early versus late fusion in semantic video analysis. In Proceedings of the ACM International Conference on Multimedia, 2005 P. K. Atrey, M. A. Hossain, A. El Saddik, and M. S. Kankanhalli. Multimodal fusion for multimedia analysis: a survey. Multimedia systems, 16(6):345–379, 2010. D. L. Hall and J. Llinas. An introduction to multisensor data fusion. Proceedings of the IEEE, 85(1):6–23, 1997. E. Bruni, N.-K. Tran, and M. Baroni. Multimodal distributional semantics. Journal of Artificial Intelligence Research, 49: 1–47, 2014 D. Kiela. Deep embodiment: grounding semantics in perceptual modalities. Technical report, University of Cambridge, Computer Laboratory, 2017. A. Lazaridou, E. Bruni, and M. Baroni. Is this a wampimuk? cross-modal mapping between distributional semantics and the visual world. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2014. H. de Vries, F. Strub, S. Chandar, O. Pietquin, H. Larochelle, and A. Courville. GuessWhat?! Visual object discovery through multi-modal dialogue. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017 T. Baltrušaitis, C. Ahuja, and L.-P. Morency. Multimodal machine learning: A survey and taxonomy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(2):423–443, 2019. Shao, Wenqi, et al. "Channel equilibrium networks for learning deep representation." ICML (2020). ----------------------------------- Thank you for this excellent rebuttal and the new baselines. I am stil not convinced that the method can be easily extended to other modalities (a simple MLP may not easily output a convnet activation), but the methods remain sound for RGB-D and related vision tasks. I thus updated the paper score.

Reproducibility: Yes

Additional Feedback: Small remarks: - Abstract: "Sufficient" ablation studies have also been carried out. "Sufficient" is quite subjective, and I would remove it - Eq6: theta -> 0

Review 2

Summary and Contributions: This paper presents a channel-exchanging-network, a parameter-free multi-modal fusion model, to dynamically exchange channel information between sub-networks of different modalities. Experiments are conducted on two tasks to validate the effectiveness of the proposed method.

Strengths: --The model is interesting to me. The importance of each channel is measured by the scaling factor of BN or IN and weak channels are replaced by the averaged channels of other modality, and effective fusion of different modalities is thus achieved. --The method has sufficient theoretical analysis and thorough experiments. --The model is likely to be of interest to a large proportion of the community.

Weaknesses: --Some experiments are missing. The comparison of the method against the random exchanging in ShuffleNet should be reported. In addition, if those unimportant channels are judged, you directly discard them. What are the advantages of the proposed method over this operation? The experiment should be conducted to show the effectiveness of channel exchanging over directly discarding the unnecessary channels. --In equation (6), the current channel is replaced with the mean of other channels if its scaling factor is smaller than a certain threshold. Why use the mean of other channels? Are there any other choices?

Correctness: Yes.

Clarity: Needs improvements.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback:

Review 3

Summary and Contributions: It has been observed that multi-modal fusion often does not lead to performance improvement with a simple fusion approach. The hypothetical reason for this non-trivial fusion issue is deemed due to the difficulty of balancing the trade-off between the multi-modal information. This submission proposes a novelty way of tackling this balancing problem in multi-modal fusion, i.e., channel exchange. The authors argue that this approach is a parameter-free approach to adjust the balance. In order to select the channels to be exchanged between two modalities, the authors measure channel importance by the magnitude of the Batch-Normalization scaling factor, denoted as \gamma, which is motivated by the channel pruning [38]. But the authors extend the understanding by Theorem 1. Also, Eq. (6) (channel exchange) is interesting in that, according to Corollary 1, there always exists a linear combination of other modality channels that decrease or at least unchange the loss (it is intuitive though). The proposed architecture is compact by sacrificing the applicability: the proposed method is only applicable to the problems where the input data shape of modalities have the same (compatible) form. This submission is a strong submission w.r.t. writing, technical contribution, and evaluation. =================Updated after rebuttal================== This reviewer has read all the reviews and the authors' rebuttal. This reviewer supports to accept this work. In the following, this reviewer has a few comments. While it was good to confirm that Theorem 1 still holds (I knew), this reviewer's point was not on the validity of Thoerem1, but on the gap between practice and theory. With \beta!=0, the probability that Theorem 1 holds is not necessarily with high probability. Thus, without \beta=0, the claim in L184-186 is not a guaranteed claim. This can be reflected in the final version. Also, this reviewer agrees with R1's comment on the related work and baselines. The authors are recommended to reflect the comments properly in the final version.

Strengths: - This submission address a multi-modal fusion balancing issue - The property of the algorithmic behaviors are adequately formalized - The performance is noticeably increased while keeping compact network architecture - The experiment results demonstrate the effectiveness of the proposed method adequately.

Weaknesses: - Theorem 1 is derived and interpreted under the assumption that \beta is ignorable (L178-179), which introduces the gap between Theorem 1 and practice. But the authors empirically show that the result still holds. (see the additional feedback below) - The proposed method is not applicable to a completely heterogeneous form of multi-modal data, e.g., image + audio.

Correctness: - Except the assumption (\beta = 0) and over-generalized interpretation to the practical case (\beta!=0), everything looks good.

Clarity: - The paper is well-motivated, and the description is clear.

Relation to Prior Work: - The baselines compared in this work are deemed to be sufficient.

Reproducibility: Yes

Additional Feedback: - L178-179, L186-188: Although the authors demonstrate the empirical result that many \gamma is stuck into 0 (a local minimum), the reason that \gamma is converged to zero during training (not at a convergence point) may not be necessary due to a small gradient if \beta is not zero. In an extreme case that \beta != non-zero, \gamma == zero, and x-mu != 0, then x' = \beta; thus, there always exists a descent direction dL/dx' that is non-zero. This implies that there is at least a descent direction to \gamma that is escapable from a basin of attraction. In short, Theorem 1 with the assumption of \beta == 0 introduces the gap between the theory and the actual implementation. In this regard, it would have been interesting to see the algorithmic behavior of the model implementation with no \beta. However, this cannot be the reason to discount this submission if the authors discuss this part properly in the paper. Anyway, resultantly, the result happens to be held, as shown in Fig. 5 in the supplementary material due to some other reasons. This reviewer would like to hear the authors' feedback on this part.

Review 4

Summary and Contributions: This paper proposes a new approach for  multimodal fusion. The method is parameter-free and only uses the coefficient of batch normalization (BN) as a criterion to replace some unimportant channels. The proposed architecture is relatively straightforward and composes BN as the importance measurement of each corresponding channel. It has better performance on semantic segmentation and image translation. In addition, it includes some theoretical analysis and nice ablation studies which help explain why it works.

Strengths: 1. The main contribution is the method on network pruning for multi-modal fusion, which allows channel-wise information to be exchanged between modalities but also retain intra-modal features. 2. The paper reinterprets modality-common and modality-specific patterns in the context of intra-modal and inter-modal information fusion. 3. The paper is demonstrated through well-designed experiments.

Weaknesses: 1. While the paper does achieve superior results, large chunks are devoted to highlighting BN and arguing that BN approaches can do better than other approaches. To this end, the novelty seems to be limited to putting the similar idea with [38]. 2. Authors make a point about exchanging channel — but exchanging channel is only achieved using BN. I am not really sure if such a contributionis incremental. In my opinion, the rationale of Theorem 1 are mostly based on [38], and Theorem 1 seems like a simple fact. 3. It would be better if the authors provide a figure of the percentage of exchanged channel for every layer, which will help readers better understand the model. 4. As stated in the paper, all modalities must be homogeneous.

Correctness: The claims and method are correct.

Clarity: The paper is well written while as a research paper, the contributions are not so significant.

Relation to Prior Work: The difference between the proposed method and prior work is clearly discussed.

Reproducibility: Yes

Additional Feedback: Overall, the idea is simple yet compelling. But I do have the following questions: 1. In eq(6), the current channel is replaced with the mean of other channels if the scaling factor is smaller than a certain threthod. If there are three modalities (represented by A, B, and C) and two of these (suppose they are B and C) are lower than the thredthod, the x of B is the mean of A and C, or just be replaced by A? And do all experimental settings specify two modalities? Does it make effect for three or more modalities? 2. I failed to run the segmentation code through your github link. The error is “CUDA out of memory”, but my GPU’s memory is 12G. Thus I suggest mentioning the GPU device type clearly in the paper. 3. Does the L1 penalty really work? Beacause I noticed that in Table 6 of the supplementary material, the mean IoUs shown in the third row are similar to those shown in the sixth row (46.2/38.4/48.0 vs. 46.0/38.1/47.7).