A Unified Framework for U-Net Design and Analysis

U-Nets are a go-to, state-of-the-art neural architecture across numerous tasks for continuous signals on a square such as images and Partial Differential Equations (PDE), however their design and architecture is understudied. In this paper, we provide a framework for designing and analysing general U-Net architectures. We present theoretical results which characterise the role of the encoder and decoder in a U-Net, their high-resolution scaling limits and their conjugacy to ResNets via preconditioning. We propose Multi-ResNets, U-Nets with a simplified, wavelet-based encoder without learnable parameters. Further, we show how to design novel U-Net architectures which encode function constraints, natural bases, or the geometry of the data. In diffusion models, our framework enables us to identify that high-frequency information is dominated by noise exponentially faster, and show how U-Nets with average pooling exploit this. In our experiments, we demonstrate how Multi-ResNets achieve competitive and often superior performance compared to classical U-Nets in image segmentation, PDE surrogate modelling, and generative modelling with diffusion models. Our U-Net framework paves the way to study the theoretical properties of U-Nets and design natural, scalable neural architectures for a multitude of problems beyond the square.

Paper

References (99)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer go4z7/10 · confidence 3/52023-06-19

Summary

This paper proposes a formal definition of U-nets, a crucial building block of modern deep learning pipelines such as diffusion models. Thanks to this definition it is possible to both get theoretical results explaining some of the u-nets behaviours, and generalize u-nets to settings more exotic than the 2D images. Experiments are then conducted on various modalities to validate the theory.

Strengths

- I agree with the main selling point of the paper : U-Nets have been under studied, especially in the context of diffusion models, and it is more than necessary to put an end to that. - I think the formalism effort presented in this work is great, and it was indeed, to the best of my knowledge, lacking. - the diversity of the modalities studied showcases the potential impact of better understanding U-Nets

Weaknesses

- **inductive bias**: theorem 2 is the cornerstone of the theoretical explanation of this work as to why U-Nets are effective in diffusion. Yet I think that it has 2 problems. First, the result is about the Haar decomposition of the diffusion process, and therefore it applies to all models that have a multiscale structure with average pooling, not specifically U-Nets. Second, and in my view much more problematic, diffusion pipelines often use an "epsilon-type" prediction framework, in which the role of the U-Net is not to predict the denoised image (or signal in general), but rather the noise. There are multiple examples of this that can be found either in the examples of the diffusers library of HuggingFace or in the seminal papers of Ho and Song: - in the unconditional generation example: https://github.com/huggingface/diffusers/blob/main/examples/unconditional_image_generation/train_unconditional.py#L576, when prediction type is "epsilon", the loss is computed as the MSE between the output of the model and the noise, and one can see that this is the default mode (L229), not changed in the way the script is called (https://github.com/huggingface/diffusers/blob/main/examples/unconditional_image_generation/README.md) - in the text-to-image generation example, the prediction type is taken by default from the noise scheduler which has it by default has "epsilon" (https://github.com/huggingface/diffusers/blob/716286f19ddd9eb417113e064b538706884c8e73/src/diffusers/schedulers/scheduling_ddpm.py#L121), not changed in the way the script is called (https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/README.md) - In [7], algorithm 1 clearly shows that it's the noise that is learned - In the implementation of [34], the noise `z` is used used in the MSE with the model output (https://github.com/yang-song/score_sde/blob/main/losses.py#L115) As a consequence I think the theoretical result does not cover the practice. Indeed, it is said "This inductive bias enables the encoder and decoder networks to focus on the signal on a low enough frequency which is not dominated by noise." while the network's role is precisely to focus on subspaces which feature the noise in order to return it. While I understand that some pipelines still use a "sample" type prediction, I think that the understanding of the U-Net inductive bias is not understood if it doesn't take into account both types of predictions, especially if the explanation of one type is contrarian to the other type. - **clarity and space use**: I think that some sections of the paper are very unclear, in part because some essential parts were left out in the appendix. For example, Algorithm 1 for staged training should be in the core paper. Similarly the practical construction of a U-Net that guarantees boundary conditions should be in the core paper. For now, we just get an idea of what the decoder subspaces should look like, but not how the decoder should be designed in order to achieve these subspaces. Right now it is not clear to me how these decoders are designed. A last example (although there are some more) is that the formal definition of the U-Net is hard to grasp at first read. There should be a correspondence between the elements introduced and the typical 2D image segmentation U-Net architecture. - **prior work comparison/mention**: I noticed two parts that needed more references to prior work/comparison to it. First, the connection between U-Nets and wavelets has been a topic of many works among which [A, B]. Second, diffusion on complicated geometries including manifolds has been recently covered by works such as [C-F]. For both of these, mentioning as well as highlighting the differences would be super important. *Minor*: - To me preconditioning means finding a matrix close to the Hessian inverse in an optimization procedure in order to better take into account the geometry of the problem when optimizing, not finding a good initialization. Here it seems that preconditioning is more used to describe a good initialization rather than taking into account geometric information. - "It also explains why in practice, encoders are often chosen significantly smaller than the decoder [26]". I am not sure about this fact, so I think it needs much more evidence (see for example in the diffusers library https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d.py#L198 the decoder only has one more layer than the encoder). Also [26] is about VAEs, not U-Nets. [A]: Ramzi, Z., Michalewicz, K., Starck, JL. et al. Wavelets in the Deep Learning Era. J Math Imaging Vis 65, 240–251 (2023). [B]: Liu, P., Zhang, H., Zhang, K., Lin, L., & Zuo, W. (2018). Multi-level wavelet-CNN for image restoration. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops (pp. 773-782). [C]: De Bortoli, V., Mathieu, E., Hutchinson, M., Thornton, J., Teh, Y. W., & Doucet, A. (2022). Riemannian score-based generative modeling. arXiv preprint arXiv:2202.02763. [D]: Fishman, N., Klarner, L., De Bortoli, V., Mathieu, E., & Hutchinson, M. (2023). Diffusion Models for Constrained Domains. arXiv preprint arXiv:2304.05364. [E]: Huang, C. W., Aghajohari, M., Bose, J., Panangaden, P., & Courville, A. C. (2022). Riemannian diffusion models. Advances in Neural Information Processing Systems, 35, 2750-2761. [F]: Liu, L., Ren, Y., Lin, Z., & Zhao, Z. (2022). Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778.

Questions

- What is meant by self-similarity? - In the code, for MNIST experiment, a network called GNet is used rather than a U-Net : what is it and does it fit with the experiments reported in the paper? - When deriving Theorem 2, it seems that each subspace W_i is an image subspace. However, in practice the U-Net acts on feature spaces, i.e. spaces with much more feature channels than the image itself, how would you take that into account?

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

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

2 fair

Contribution

3 good

Limitations

N/A

Reviewer ndAy5/10 · confidence 3/52023-06-25

Summary

This paper proposes a framework for designing and analyzing general UNet architectures. Theoretical results are presented that can characterize the role of encoder and decoder in UNet and conjugacy to ResNets is pointed out via preconditioning. Furthermore, this paper proposes Multi-ResNets, UNets with a simplified, wavelet-based encoder without learnable parameters. In addition, this paper presents how to design novel UNet architectures which can encode function constraints, natural bases, or the geometry of the data. Experiments of Multi-ResNets on image segmentation, PDE surrogate modeling, and generative modeling with diffusion model are conducted and demonstrate competitive performance compared to classical UNet.

Strengths

- Provide the mathematical definition of UNet, which enable identifying self-similarity structure, high-resolution scaling limits, and conjugacy to ResNets via preconditioning. - Based on theoretical analysis, authors propose Multi-ResNets, a novel class of UNet with no learnable parameters in its encoder. - Multi-ResNets achieves competitive or superior results compared to a classical U-Net in PDE modeling, image segmentation, and generative modeling with diffusion models.

Weaknesses

- The presentation of the paper is not smooth and clear, which needs significant improvement. For example, the authors claim to propose a unified framework for UNet design, but I don’t see how this model can help researchers design specific UNet for different tasks. It’s more like the authors improve UNet on three chosen tasks and provide some theoretical analysis. Please correct me if I misunderstood anything. - This paper did a poor evaluation of the PDE part. In the PDE surrogate modeling, UNet is not the state-of-the-art model (authors made a wrong and misleading claim about the performance of UNet in PDE modeling) and performs much worse than the Fourier Neural Operator (FNO) [1][2][3]. So I am not convinced that it is useful to only demonstrate the effectiveness of Multi-ResNets compared to UNet in PDE modeling. In addition, no other representative baselines are compared within the NS and shallow water experiments, such as FNO[1], MPPDE[4], FFNO[5], etc. [1] Li, Zongyi, et al. "Fourier neural operator for parametric partial differential equations." arXiv preprint arXiv:2010.08895 (2020).\ [2] Takamoto, Makoto, et al. "PDEBench: An extensive benchmark for scientific machine learning." Advances in Neural Information Processing Systems 35 (2022): 1596-1611.\ [3] Helwig, Jacob, et al. "Group Equivariant Fourier Neural Operators for Partial Differential Equations." arXiv preprint arXiv:2306.05697 (2023).\ [4] Brandstetter, Johannes, Daniel Worrall, and Max Welling. "Message passing neural PDE solvers." arXiv preprint arXiv:2202.03376 (2022).\ [5] Tran, Alasdair, et al. "Factorized fourier neural operators." arXiv preprint arXiv:2111.13802 (2021).

Questions

Please refer to the weakness part.

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

2 fair

Contribution

2 fair

Limitations

The limitations of this paper are well discussed.

Reviewer Q8QJ7/10 · confidence 2/52023-06-30

Summary

The paper presents a unified framework for U-net design, by generalizing the overall structure of U-nets into different components. The framework is then investigated from a theoretical point of view. At its core, the paper presents Multi-Resnets which is a novel class of U-nets, which is then tested rigorously in the experimental section, in relation to generative modelling, PDE modelling and image segmentation.

Strengths

It is always great to see papers that take a step back and tries to unify different work into a framework, to conceptualize what is up and down in a certain field. In that sense this paper success in doing this. In particular, I think sections such as 5.1 are some of the stronger sections of the paper because they conceptualize what the role of the encoder is. Originality: The framework presented in the paper seems to be original compared to other work. Quality: The paper seems to be of a high quality, with both theoretical results to backup the claims of the paper but also empirical evidence to support. This in especially true, when the appendix of the paper is taken into account as it provide much more information, especially on the experimental section of the paper. Clarity: The paper is fairly clear in what it is trying to achieve, however since there are a lot of moving parts to the paper the overall structure sometimes suffer from this and the overall story become a bit "muddy". Significance: While the conceptualization of U-nets indeed is refreshing, I doubt this paper will have a huge impact on domains that use U-nets as most established fields seems to build incrementally upon each others instead of redesigning architectures. That said, only time will tell.

Weaknesses

I did not in particular find any weaknesses of the paper.

Questions

Maybe out of scope for the paper, but let me ask anyway: The original reason for resnet being introduced was to solve the problem of vanishing gradients. Do you see a connection between this problem and your work?

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

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central 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

The authors provide a lengthy explanation of what the limitations of the framework is in the appendix of the paper, however I think the authors should have definitely included this in the main paper. To me, while the framework generalizes U-Nets it does not necessarily guarantee that the design process becomes easier, as the authors also touch on L162-166. This indeed does seem to be a limitation of the framework, when the initial goal was to "..designing and analysing general U-Net architectures".

Reviewer 7MgH6/10 · confidence 4/52023-07-02

Summary

The work proposes a unified mathematical framework to analyze and design U-Net. Authors highlight the importance of preconditioning and provides several highlights to design unet. Authors propose a new parameter-free encoder based on wavelet space.

Strengths

1. Sec 2 presents a unified mathematical framework for U-Unets. 2. The discussion of precondition parts and how to design various precondition for different tasks is insightful, such as sec 3.2, 3.3. 3. Authors also conduct various experiments to support the their claims and highlight the importance of precondition and network design. Some experiments are quite interesting to me, such as Sec 5.1 and 5.3.

Weaknesses

1. Though it shows improvement in other tasks, there are minimal improvements in diffusion model tasks. Authors should report performance of Multi-ResNet for diffusion models and compare it with existing U-Nets used widely in diffusion models. 2. I would like authors to comments on how to design Unet for diffusion model so that it can recover high frequency details in high resolution image generation task. The average pooling will filter high frequency noise, but also filter the high frequency data information.

Questions

See 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

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

See weakness

Reviewer go4z2023-08-11

Thanks for your answers

Before answering, I would like to thank the authors for engaging in the review process in what I view as a very positive and intellectually honest manner. I think that aside from the 2nd point about theorem 2, I don’t have major blockers left, granted the improvements mentioned in the rebuttal are implemented (which I don’t think is a big issue). *1st point on Th. 2., about specificity*: But then how does this result relate to e.g. the one by Donoho in “Denoising by soft thresholding” (1995)? He also looked at the variance of the noise in wavelet coefficients. *2nd point on Th. 2., about epsilon prediction*: I think the clarification that the decoder can focus on the high frequency is important, and to me it sounds contradicting with what was written in the paper “This inductive bias enables the encoder and decoder networks to focus on the signal on a low enough frequency which is not dominated by noise.”, but maybe it’s because I am understanding the word “focus” in this context in the wrong way. Still I think this demands clarification. Further, my main take-away from this response is that the inductive bias of the U-net basically (and maybe I am simplifying too much here) is that it has the ability to provide “a lower-frequency, low-noise [version of the] signal”. I don’t understand then how we can understand that this provides a good inductive bias, i.e. is “valuable compressed information that aids the decoder in generating its output”. To me this is the crux of the matter, rather than showing that the U-net’s encoder can extract this information. I realize that my problem is not necessarily on Theorem 2., which I see as valid (not having checked the proof thoroughly but it matches what I have seen in the past in wavelet works), but rather on the interpretation made thereafter. In the abstract it says "In diffusion models, our framework enables us to identify that high- frequency information is dominated by noise exponentially faster, and show how U-Nets with average pooling exploit this.". While I agree that the first part of this sentence is backed up in the text, to me the second part is not as I explain above.

Authorsrebuttal2023-08-13

Reply to: 'Thanks for your answers' (1/3)

We very much share the positive sentiment about this useful and enlightening discussion with the reviewer. We are thankful for the questions the reviewer is raising which allow us to improve and produce a clearer, more accurate explanation of our results. We are glad we could resolve the reviewer’s concerns in most points through the explanations and improvements provided, and would like to discuss the remaining questions below. We would be thankful if you could consider the outcome of this discussion towards the final recommendation and score of your review. > “1st point on Th. 2., about specificity: But then how does this result relate to e.g. the one by Donoho in “Denoising by soft thresholding” (1995)? He also looked at the variance of the noise in wavelet coefficients.” We appreciate the reviewer for highlighting this relevant article. We recognise a notable similarity between the message of our Theorem 2 and the analysis of coefficient decay in the pertinent Besov space in [Donoho1]. Under the forward diffusion process analysed in Theorem 2, the signal of our data predominantly concentrates on the low-frequency wavelet coefficients. In light of [Donoho1], this observation can also be articulated as: when examining the appropriate sequence of wavelet coefficients in a sequence space, the high-frequency coefficients, in expectation, tend to zero considerably faster than their low-frequency counterparts. Therefore, restating Theorem 2 in the context of the Besov space would indeed enhance clarity, especially in elucidating how this property can be generalised to other geometries, given that all analysis is anchored in the Besov space. We propose to incorporate this analysis into the appendix of our manuscript, with reference to the given article, complemented by a more generalised statement of Theorem 2 for users working with U-Nets in complicated geometries. Another intriguing idea would be to utilise this sequence space information from a specific dataset to define the noise schedule of a diffusion process and to determine the size of each decoder block based on these coefficients. While delving deeper into this would require additional research, we are confident that presenting this information in a straightforward manner could be an interesting extension towards better U-Net designs *tailored to individual datasets*. We believe that presenting Theorem 2 in this expanded manner, coupled with our general U-Net framework, marks an initial step towards such research. The crux here lies in understanding how this sequence space decays, because as demonstrated in [Donoho1], this can provide insights into our model's design. The reviewer may also be referring to the wavelet shrinkage procedure on page 3 first proposed in (Donoho, 1992). In step (2), wavelet coefficients are translated and then thresholded (the operator $(\cdot)_+$ is in our understanding the ReLU function) which effectively results in a ‘continuous shrinkage to zero’ of wavelet coefficients, where some coefficients are indeed set to zero. This is in contrast to the ‘hard projection’ in average pooling, where coefficients are either their identity or set to zero. It is also worth noting that this procedure makes the assumption that the reconstruction is at least as smooth as the ground-truth function that shall be approximated (see (1.3)), an assumption which Theorem 2 does not make. On a final note: [Donoho1] is a long paper with many results and we admit that it is slightly challenging at present. If we should take into consideration another specific result from this paper that we have not discussed above, we would appreciate it if the reviewer could point us to it. Lastly, we would like to point out that there are two versions of this paper. We carefully read the version cited below [Donoho1]. We thank the reviewer for this very interesting reference to relevant literature, and the useful discussion on it. [Donoho1] https://web.stanford.edu/dept/statistics/cgi-bin/donoho/wp-content/uploads/2018/08/denoiserelease3.pdf

Authorsrebuttal2023-08-13

Reply to: 'Thanks for your answers' (2/3)

Below, we will respond to several comments coherently. > “2nd point on Th. 2., about epsilon prediction: I think the clarification that the decoder can focus on the high frequency is important, and to me it sounds contradicting with what was written in the paper ‘This inductive bias enables the encoder and decoder networks to focus on the signal on a low enough frequency which is not dominated by noise.’, but maybe it’s because I am understanding the word ‘focus’ in this context in the wrong way. Still I think this demands clarification. To resolve this confusion, we would like to cite and explain the two sentences following the cited sentence above: “[...] As the subspaces are coupled via preconditioning, the U-Net can learn the signal which is no longer dominated by noise, added on each new subspace. This renders U-Nets a computationally efficient choice in diffusion models [...].” Please allow us to explain this with reference to Sec 2. Say a U-Net $U_i$ is making an approximation on resolution $i$. Then, the decoder $D_i$ on resolution $i$ and output space $W_i$ is making an approximation *preconditioned* (i.e. receiving as input) the output of the U-Net $U_{i-1}$ on resolution $i-1$, which is also the output of the decoder on resolution $i-1$ (see Eq. (1)). Hence, the decoder $D_i$ on the resolution we wish to approximate receives as input a ‘lower resolution approximation’ of the output it wishes to produce, and can in this sense focus on the added information (i.e. the high-frequency details) relative to what it is preconditioned on (i.e. the ‘lower-resolution approximation’ of $U_{i-1}$. The encoder $E_i$ facilitates this via its skip connection by providing a ‘lower-dimensional, low-noise input’, as the reviewer correctly stated, or more precisely, a representation in a changed basis thereof. This input is useful for the learning task of the decoder $D_i$ regardless of whether it is outputting noise (epsilon-type) or signal (sample-type) as we outlined in our rebuttal. We hope this provides an explanation of what we mean by the decoder being able to “focus on high frequency” details in the paper. What we intended to express with the sentence quoted by the reviewer is that the encoder and decoder networks can focus their ‘learning task’ on only the frequency information added on a certain resolution, in light of preconditioning. The decoder produces the actual approximation of the output on resolution $i$, preconditioned on the previous resolution’s approximation. Crucially, it also requires the encoder $E_i$’s output as input. If we choose $P_i$ as average pooling, as a first approximation of the encoder, we project to lower-resolution subspaces, because average pooling is conjugate to projection in a Haar wavelet basis (see Theorem 2 in [5] for a detailed explanation), which is also the basis of choice in Theorem 2. According to Theorem 2 in the submission, these subspaces correspond to ‘low noise, high signal’. The encoder $E_i$ can then learn a change of basis thereof and pass it to the decoder. In this sense, both the encoder and decoder can focus on a “low enough frequency which is not dominated by noise”, enabled via average pooling. Based on this discussion, we suggest to reformulate the quoted sentence to: ***‘This inductive bias enables the decoder to have access to signal-dominated, low-frequency information, which is not dominated by noise.’*** We thank the reviewer in particular for enabling this improvement.

Authorsrebuttal2023-08-13

Reply to: 'Thanks for your answers' (3/3)

> “Further, my main take-away [...] is that the inductive bias of the U-net [...] it has the ability to provide ‘a lower-frequency, low-noise [version of the] signal’. I don’t understand then how we can understand that this provides a good inductive bias, i.e. is ‘valuable compressed information that aids the decoder in generating its output’. [...]” > “In the abstract it says ‘In diffusion models, our framework enables us to identify that high- frequency information is dominated by noise exponentially faster, and show how U-Nets with average pooling exploit this.’. While I agree that the first part of this sentence is backed up in the text, to me the second part is not as I explain above.” We hope that given our response above, the usefulness of average pooling, it being a beneficial inductive bias has become more apparent. Another way of putting it is the following argument: our goal in diffusion models is to separate the signal from the noise, both in the epsilon-type and sample-type formulation. ***If we perform average pooling many times, we get lower-resolution coefficients where the signal-to-noise ratio is high, i.e. where it is easy to separate the signal from the noise. This is exactly what we mean by ‘beneficial inductive bias’: The network’s architecture facilitates solving the approximation task.*** We can easily and computationally efficiently make a good approximation on the lower-resolution, and one resolution after the other (via preconditioning) focus on higher resolutions where separating the signal from noise is significantly harder. In light of this discussion, we acknowledge that the interpretation of Theorem 2 could be clearer and extended. Towards a potential camera-ready version, we will use the above discussion to clarify Theorem 2, and very much thank the reviewer for pointing this need out to us. > “I realize that my problem is not necessarily on Theorem 2., which I see as valid (not having checked the proof thoroughly but it matches what I have seen in the past in wavelet works), but rather on the interpretation made thereafter.” We are glad that we seem to be able to resolve some of these questions on the applicability of Theorem 2 through this discussion, and agree that the remaining questions are mostly on the interpretation of Theorem 2 (which is crucially important!). While we tried very hard, explaining mathematical statements in words is always challenging and can only get epsilon-close to the truth. Hence we very much appreciate this discussion as these insights are far from straight-forward, and are happy to clarify any further questions, if required.

Reviewer go4z2023-08-14

I repeat again here that I am extremely happy with the way the conversation is happening, and I thank the authors ocne more for engaging in such a positive way. I think my comment on the first point has been well received. I am happy with how the theorem 2 is planned to be discussed from now on in the camera-ready version of the paper. For all of these reasons I am increasing my score to 7, Accept. The reason I am not giving an even higher score is because I think that Theorem 2 shares too many similarities with the existing result by Donoho. I have 2 last points: - I think it would be valuable to add in a comment (maybe in appendix as an extended discussion) on how max pooling can also extract low resolution information (can it?). - Another interesting reference might be: “DEEP CONVOLUTIONAL FRAMELETS: A GENERAL DEEP LEARNING FRAMEWORK FOR INVERSE PROBLEMS“ by Ye et al. 2018.

Authorsrebuttal2023-08-15

Reply to: Official Comment by Reviewer go4z

We would like to thank the reviewer for the valuable and profound discussion. We truly appreciate the reviewer’s efforts and the questions raised, particularly on Theorem 2, which contribute to and allow us to provide an improved, clarified version of our submission in a potential camera-ready version. We thank the reviewer for bringing up the topic of max pooling and we will comment how this fits within our framework in an updated appendix as well as discuss the relevant paper suggested in the related work accordingly. Lastly, we are grateful to the reviewer for viewing to accept our paper on the basis of this discussion. We have found this interaction to be very beneficial to the improvement of our work and are thankful to the reviewer in this regard.

Reviewer Q8QJ2023-08-15

Thanks for your answers

I thank the authors for their lengthy answer, to not only my own review but also my fellow authors. Especially, thanks for being very honest in your answers being aware of your papers weaknesses and still defending the strengths of your paper. The mentioned improvements to the potential camera-ready manuscript sounds like great improvements to the paper. I have no further questions and comments for the paper. I will keep my score at 7 as it is already in higher end based on the other reviews.

Authorsrebuttal2023-08-16

Response to: "Thanks for your answers"

We thank the reviewer for their useful questions and remarks which have helped to improve the quality of our submission towards a potential camera-ready version. We are grateful for the efforts in carefully reviewing our work. Lastly, we thank the reviewer for maintaining the high score and viewing to accept our article.

Reviewer ndAy2023-08-18

Most of my concerns have been addressed by the author's response. Regarding the performance comparison of FNO and UNet, PDEArena and PDEBench draw the opposite conclusion, so the current benchmark paper on PDE is not a gold standard yet. So it's better for the authors to run PDE baseline models, such as FNO, by themselves and compare with their modified UNet model (after all one study case is PDE modeling, so it is better to not only have UNet result but also include other baseline models.). I understand that this paper is not solely about PDE modeling, and given the fact that my other concerns are addressed, I will increase my score to 5.

Authorsrebuttal2023-08-18

Response to: Official Comment by Reviewer ndAy

We thank the reviewer for their suggestions which have improved the quality of our paper, and for raising their score in light of the discussion. In a potential camera-ready version, we will include the experiment we presented in the author rebuttal PDF, where we conducted an independent FNO baseline comparison to U-Nets. In this experiment, we observed that, for a comparable model size and on two datasets (Navier-Stokes, Shallow water), U-Nets outperform FNO in terms of r-MSE (across multiple random seeds), in some cases by an order of magnitude. We will also include a more thorough literature review on the topic.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC