Understanding the Expressivity and Trainability of Fourier Neural Operator: A Mean-Field Perspective

In this paper, we explores the expressivity and trainability of the Fourier Neural Operator (FNO). We establish a mean-field theory for the FNO, analyzing the behavior of the random FNO from an edge of chaos perspective. Our investigation into the expressivity of a random FNO involves examining the ordered-chaos phase transition of the network based on the weight distribution. This phase transition demonstrates characteristics unique to the FNO, induced by mode truncation, while also showcasing similarities to those of densely connected networks. Furthermore, we identify a connection between expressivity and trainability: the ordered and chaotic phases correspond to regions of vanishing and exploding gradients, respectively. This finding provides a practical prerequisite for the stable training of the FNO. Our experimental results corroborate our theoretical findings.

Paper

References (37)

Scroll for more · 25 remaining

Similar papers

Peer review

Reviewer gSyo6/10 · confidence 3/52024-07-06

Summary

In this work, the authors investigate the expressive power and trainability of the Fourier Neural Operator (FNO). A mean-field theory is adapted for the FNO to examine the behavior of randomly initialized FNOs from the perspective of the 'edge of chaos'. The investigation focuses on understanding the expressive behavior of randomly initialized FNOs by examining the transition between ordered and chaotic phases. This phase transition demonstrates specific characteristics unique to the FNO, influenced by mode truncation, and shares similarities observed in densely connected networks and CNNs. Additionally, the authors establish a correlation between expressive power and trainability: the ordered and chaotic phases correspond to regions where gradients vanish or explode, respectively. The authors identify the necessity of initializing FNO near the edge of chaos for stable training theoretically and experimentally.

Strengths

1. **Theoretical novelty:** The paper adapts mean-field theory specifically for Fourier Neural Operators (FNOs), extending existing theories from other neural network architectures such as densely connected networks and CNNs. 2. **Practical implications:** The theoretical findings lead to concrete recommendations for initializing FNOs to ensure stable training, particularly for deep FNOs. 3. **Comprehensive experiments:** The authors validate their theoretical results across multiple PDE-solving tasks, including advection, Burgers, Darcy Flow, and the Navier-Stokes equations. All these datasets are publicly available.

Weaknesses

1. The adaptation of mean-field theory to FNOs, while a notable endeavor, may not sufficiently break new ground in the field. 2. The significance of the theoretical aspects that FNOs share similarities with CNNs should be further discussed. Overall, this work addresses how the network should be initialized for stable training of FNOs; however, its contribution might be somewhat limited. A stability analysis beyond the initialization phase, such as on the architecture, would be interesting to see. 4. There are barely any experimental results shown in the main text, and the discussion of how their theoretical analysis is useful in practical applications is lacking. The discovery in their work, the He distribution for FNOs, has already been applied in prior works such as [1] (see their GitHub implementation, `reset_parameters()` method under models); hwoever, it is always good to have a rigorous justification, which is shown in this work. [1] Helwig, J., Zhang, X., Fu, C., Kurtin, J., Wojtowytsch, S., and Ji, S. Group equivariant Fourier neural operators for partial differential equations.

Questions

**Major Questions:** 1. If we treat FNO simply as a CNN, where global convolution occurs in the frequency domain, with Parseval theorems, it also implies, though less rigorously, that we can initialize the (complex) weights based on the He Distribution. This also hints on the similarities between FNOs and CNNs. Could the authors comment on this perspective? 2. In line 122, it states: > "two learnable weights $\Theta^{(\ell, k)} \in \mathbb{R}^{D \times D}$ and $\Xi^{(\ell, k)} \in \mathbb{R}^{D \times D}$". If I understand correctly, these are complex weights for spectral convolution. Firstly, why is $\Xi^{(\ell, k)} \in \mathbb{R}^{D \times D}$? Is this a typo where a $K$ is missing? Secondly, why are there two sets of learnable weights? Isn't there just one for the Fourier transform of the convolution kernel? Do you treat the real and complex parts separately, resulting in one set of parameters for the real and another for the complex parts? 3. The weights in spectral convolution are complex. By initializing the weights according to $\mathcal{N}\left(0, \frac{\sigma^2}{4 D}\right)$, do you treat the real and complex parts separately? 4. The FNO is implemented with $L = 64$ layers, which is particularly deep for FNOs involving $64$ forward and backward FFTs. Is such a large number of layers practical in real-world applications? If not, how are your findings useful for shallower FNOs? **Minor/Optional Questions and Suggestions:** 1. Some abbreviations are introduced before their full names, such as DCN in line 25, whereas its full name is introduced in line 37 (densely connected network, DCN). 2. Given your insights into FNO expressivity and trainability, do you have any thoughts on potential architectural improvements or variations that could enhance FNO performance? 3. It would be beneficial to provide a clearer explanation and discussion of the theoretical results and their implications for readers without a background in mean-field theory.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes, the authors adequately addressed the limitations.

Reviewer ePA15/10 · confidence 3/52024-07-09

Summary

This paper leverages mean-field theory, developed for neural networks, to study the expressivity and trainability of random Fourier neural operators (FNOs). It provides practical insights for the appropriate initialization of FNOs.

Strengths

The authors address a significant question in the neural operators literature: why does the performance of Fourier neural operators degrade as the number of layers increases? The authors take a first step toward answering this question by adopting an original theoretical perspective.

Weaknesses

- The main body of the paper is not self-contained. Additionally, the technical section can be difficult to follow due to the need for back-and-forth referencing between multiple sections and references. Overall, the paper could benefit greatly from improved clarity. To enhance readability, I recommend reorganizing some parts of the text. - For example, the covariance matrix $\Sigma_{\alpha,\alpha'}^{(l)}$ should be introduced in section 3.2 where it is first examined. - At first glance, one might question if the iterated map in equation (6) is a definition. However, its origin becomes clear upon reviewing the supplementary material. - Lemma A.1 should be included in the main text, as it is a central result. - Some parts, such as the detailed description of the datasets, can be moved to the appendix to make space. - I recommend providing further motivation for all concepts and referencing previous works where appropriate. For instance, the importance of studying the fixed point of the iterated map should be emphasized. Most importantly, please refer to the supplementary material for proofs when available. - The proposed study focuses solely on the expressivity of random FNO at initialization, a rather limited scenario. This limitation should be more explicitly emphasized, particularly in the title. I am not convinced that the findings offer practical insights into the stable training of FNO, despite the preliminary numerical experiments provided. Additionally, the term _trainability_, as used by the authors, only pertains to the vanishing or exploding gradients of random FNO. - One might question the impact of the optimizer, the initial learning rate, and its scheduler on the results. An ablation study addressing these factors would have been beneficial.

Questions

- Does the fixed point of the iterated map always exist? - The representation of Fig.1(a) is unclear. Does it depict the gradient norm averaged over the training epochs? I assume it shows the values only at initialization, averaged over multiple seeds. Additionally, is the standard deviation relatively small? Clarification on these points would be helpful.

Rating

5

Confidence

3

Soundness

2

Presentation

2

Contribution

2

Limitations

I understand that error bars are omitted from the figures for the sake of clarity. However, including standard deviations in the tables would enhance the presentation of the results. Additionally, it is concerning that the results are averaged over only three seeds, which is insufficient for robust statistical analysis.

Reviewer xhAf8/10 · confidence 4/52024-07-10

Summary

This paper presents a comprehensive theoretical analysis of the Fourier Neural Operator (FNO) using mean-field theory. The main contributions to me at least are: 1. Establishing a mean-field theory for FNO, analyzing its behavior from an edge of chaos perspective. 2. Examining the ordered-chaos phase transition of random FNO based on weight distribution. 3. Identifying a connection between expressivity and trainability, showing that ordered and chaotic phases correspond to regions of vanishing and exploding gradients respectively. 4. Providing practical initialization requirements for stable FNO training.

Strengths

1. Theoretical Depth: The paper provides a rigorous mathematical analysis of FNOs, which has been lacking in the literature. The authors adapt and extend existing mean-field theory techniques to the specific structure of FNOs. 2. Novel Insights: The paper reveals FNO-specific characteristics in the phase transition, induced by mode truncation, while also showing similarities with densely connected networks. This bridges the gap between FNOs and more traditional neural network architectures. This was very very insighting to me! 3. Practical Implications: The theoretical analysis leads to practical guidelines for initializing FNOs, addressing a known issue of training instability in deep FNOs. This demonstrates the real-world value of the theoretical work. 4. Comprehensive Experiments: The authors validate their theoretical findings through extensive experiments on various PDE-solving tasks, showing the practical relevance of their analysis. 5. Clarity: Despite the technical nature of the content, the paper is well-structured and clearly written, making it accessible to a broader audience in the machine learning community.

Weaknesses

1. There are no specific weaknesses that I see however, while the experiments are comprehensive, the authors don't fully explore very deep FNOs or a wide range of hyperparameters (in particular sweeping over the number of modes). Additional experiments on larger-scale problems or more complex PDEs could further strengthen the empirical validation. 2. Insufficient analysis of the impact of different Fourier modes: The paper doesn't provide a detailed analysis of how the number and distribution of Fourier modes affect the expressivity and trainability. This is crucial as the choice of modes is a key hyperparameter in FNOs. 3. Insufficient exploration of error propagation: For time-dependent PDEs, there's limited analysis of how errors propagate through time steps and how this relates to the expressivity and trainability of FNOs.

Questions

1. I know this paper tackled FNO's, but I am curious if this could easily be adapted to other variants of FNOs like U-FNO, F-FNO etc 2. The paper doesn't provide a detailed analysis of how different PDE characteristics (e.g., stiffness, non-linearity, multi-scale behavior) might affect the expressivity and trainability of FNOs. Could you explain more on this point?

Rating

8

Confidence

4

Soundness

3

Presentation

3

Contribution

4

Limitations

The authors have described their limitations!

Reviewer Jbgf7/10 · confidence 3/52024-07-12

Summary

This work explores the expressivity and trainability of Fourier Neural Operators (FNOs) by establishing a mean-field theory and analyzing their behavior from an edge-of-chaos perspective.

Strengths

The study examines the ordered-chaos phase transition of the network based on weight distribution, highlighting unique characteristics induced by mode truncation and similarities to densely connected networks. It also identifies a connection between expressivity and trainability, noting that ordered and chaotic phases correspond to regions of vanishing and exploding gradients, respectively. This insight provides a practical prerequisite for the stable training of FNOs. The experimental results support the theoretical findings.

Weaknesses

1. The summary of related work can be further improved, particularly by including recent neural operator architectures from recent conferences or journals. 2. While Figure 1.(c) confirms that the original setup for initialization is reasonable, our experience suggests that the performance of FNOs in both training and test accuracy is highly dependent on the initialization variances for different hyperparameters and tasks. The derivation and results in this paper do not provide sufficient insights into this issue. 3. The code is not available.

Questions

1. Can the derivation in this paper for FNOs be generalized to more general neural operators? 2. Is it possible to identify (approximately) how the initialization depends on different datasets?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Reviewer xhAf2024-08-09

Response to the authors

Thanks for running some additional experiments for deeper FNOs and a wider range of hyperparameters. The potential applicability of your theory to various FNO variants (geo-FNO, G-FNO, WNO) is insightful, and your explanation of the challenges with U-FNO and F-FNO due to skip-connections is clear! Super cool work. The suggestion to incorporate mean-field theory for DCNs with skip-connections is a promising direction for future work. Overall, I thank the authors for clarifiing several key points and highlighted potential directions for extending this work, strengthening the paper's contributions to understanding FNOs! I will keep my score however: )but great work!

Reviewer gSyo2024-08-09

Thank you for the reply. Some of my concerns have been addressed. However, I still feel that the impact of this work might be limited, and the practical side of this work needs to be further explored. I have adjusted my score.

Authorsrebuttal2024-08-10

Dear Reviewer gSyo, Thank you very much for your thoughtful feedback and for adjusting your score. We appreciate your recognition of the improvements we have made. Your insights are invaluable, and we are grateful for your guidance in enhancing the quality and relevance of our work.

Reviewer ePA12024-08-11

I would like to thank the authors for their detailed answer. While most of my concerns have been addressed, I agree with Reviewer gSyo’s conclusions. Consequently, I have increased my score to 5.

Authorsrebuttal2024-08-13

Results of additional experiments

Below are the results of the experiment conducted with two different optimizers (SGD and RMSProp) on the Navier-Stokes dataset with $\nu=1 \times 10^{-5}$. The average results from three different seeds are presented in the table below. The learning rates for SGD and RMSProp were determined through grid search, with SGD set at 0.00001 and RMSProp at 0.0005. The table indicates that, similar to the Adam optimizer, the deep FNO can only be effectively trained around $\sigma^2 = 2.0$. This suggests that our theoretical results are consistent and independent of the choice of optimizer. | SGD (train loss) | 0.1 | 0.5 | 1.0 | 2.0 | 3.0 | 4.0 | |------------------|--------|--------|-------|-------|-------|-------| | 4 | 12.654 | 4.136 | 3.143 | 2.901 | 2.876 | 2.733 | | 8 | 21.237 | 12.560 | 4.143 | 2.493 | inf | inf | | 16 | 21.237 | 21.236 | 8.304 | 2.203 | inf | inf | | 32 | 21.237 | 21.237 | 8.191 | 1.957 | inf | inf | | RMSProp (train loss) | 0.1 | 0.5 | 1.0 | 2.0 | 3.0 | 4.0 | |----------------------|--------|--------|--------|-------|-------|-------| | 4 | 0.205 | 0.191 | 0.179 | 0.176 | 0.168 | 0.174 | | 8 | 0.214 | 0.225 | 0.216 | 0.173 | 0.185 | 0.197 | | 16 | 21.237 | 21.237 | 0.210 | 0.172 | inf | inf | | 32 | 21.237 | 21.237 | 21.237 | 0.196 | inf | inf | Dear Reviewer ePA1 Thank you very much for your detailed feedback and for taking the time to consider our responses. We are pleased to hear that we were able to address most of your concerns and that you have decided to increase your score. We understand and respect your agreement with Reviewer gSyo’s conclusions. We believe that our work provides a solid foundation and has the potential to make meaningful contributions to the field. We are committed to further refining our research and exploring additional aspects to enhance its impact and practical applications. Your feedback has been invaluable in helping us improve our paper, and we are grateful for your constructive comments and support.

Reviewer Jbgf2024-08-11

Thanks for those detailed responses for possible new references and two questions. Now, I have better understanding about the limit and potential of this theoretical analysis. I will keep my score.

Authorsrebuttal2024-08-13

Dear Reviewer Jbgf Thank you very much for your thoughtful feedback and for taking the time to review our responses in detail. We are glad to hear that our explanations have helped clarify the limits and potential of our theoretical analysis. Your insights have been instrumental in refining our understanding and presentation of the research. We are committed to further exploring the potential applications and implications of our findings to advance the field.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC