Riemannian Residual Neural Networks

Recent methods in geometric deep learning have introduced various neural networks to operate over data that lie on Riemannian manifolds. Such networks are often necessary to learn well over graphs with a hierarchical structure or to learn over manifold-valued data encountered in the natural sciences. These networks are often inspired by and directly generalize standard Euclidean neural networks. However, extending Euclidean networks is difficult and has only been done for a select few manifolds. In this work, we examine the residual neural network (ResNet) and show how to extend this construction to general Riemannian manifolds in a geometrically principled manner. Originally introduced to help solve the vanishing gradient problem, ResNets have become ubiquitous in machine learning due to their beneficial learning properties, excellent empirical results, and easy-to-incorporate nature when building varied neural networks. We find that our Riemannian ResNets mirror these desirable properties: when compared to existing manifold neural networks designed to learn over hyperbolic space and the manifold of symmetric positive definite matrices, we outperform both kinds of networks in terms of relevant testing metrics and training dynamics.

Paper

References (53)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer nyBA4/10 · confidence 3/52023-07-01

Summary

EDIT: Having read everything here. I am increasing my score slightly. However, I still think the paper is not clearly explained. It is unclear why you first introduce the method without the feature map. It seems like there are two different versions of the method and it is not clear which is which. I am also confused by the experimental set up for datasets such as CORA. Do you first embed the graphs into a manifold? Additionally, I think that requiring a closed form expression for geodesics is somewhat strong and limits the applicability. The authors claim to attempt RESNET structures to hyperbolic manifolds

Strengths

The paper is applicable to a wide variety of manifolds

Weaknesses

The proposed methods seem to not really leverage the manifolds intrinsice geometry and depend entirely on the embedding on the manifold into ambient space $R^D$. Indeed $n_i$ is defined on all of $R^D$. This means that there is no gauranteed that there would be any notion of consistence if $M$ was embedded into $R^D$ and $R^{D'}$ two different ways (where $D'$ may or may not equal $D$). This seems to be a major limitation of this that is not properly discussed. at a bare minumum, there should be some level of invariance to, e.g., Rotations and Translations of the manifold in $R^D$ after its been embedded. Additionally, much of the paper is hard to understand such as the construction of the feature maps, which appears to take place in local coordinate systems which will not be consistent across the manifold.

Questions

N/A

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good 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

1 poor

Contribution

2 fair

Limitations

N/A

Reviewer 5HC16/10 · confidence 4/52023-07-02

Summary

This paper extends the well known residual network which are usually applied to Euclidean data to a variant defined on manifold. The main novelty is to replace the "addition / plus" operation in Euclidean space to exponential operation. Specifically, given an input point on a certain manifold (e.g., hyperbolic or SPD), it first learns a vector in the tangent space of the given input point through a neural network layer, and then maps the learned vector back to the manifold using exponential. By utilizing pushforward and pullback operations, the proposed method can transform input between different manifolds with different dimensions. Experiments are conducted on hyperbolic and SPD spaces to demonstrate the superior performance of the proposed method over HNN and SPDNet.

Strengths

This paper proposed to extend residual network from Euclidean space to non-linear manifold by replacing the conventional addition/plus operation with manifold exponential operation. The theoretical part is sound and the experiments are effective in supporting the proposed method.

Weaknesses

The biggest weakness, as also mentioned in by the authors at the end of the paper, is whether the proposed method is only applicable to hyperbolic and SPD matrix? Is it possible to apply this residual network to other non-linear manifolds that have closed-form exponential manifolds? Furthermore, Is it possible to apply this residual network to other non-linear manifolds that do $\textbf{not}$ have closed-form exponential manifolds? If so, please list such manifolds.

Questions

In general: 1. Except for Euclidean, hyperbolic and SPD, any other manifold are applicable to the proposed residual network? Method part: 2. For the function h_i defined in Line 202 parameterized by a neural network, given an input on a mainfold M_{i-1}, how to ensure the output is also on a manifold M_i? 3. For SPD matrix, what's the advantage of the proposed Riemannian Residual Neural Networks over SPDNet (the AAAI paper) in theory? 4. In Line 225 - 231, what's the relationship between f_i and g_{\theta_i}? 5. Still in Line 225 - 231, why does $\nabla g_{\theta_i}$ is a map from M to TM? 6. In SPD case of Line 258 - 260, since g_k does not contain any learnable parameters, how to train such a network for SPD? Experiment part: 7. How many links on average per graph used in the experiment in Section 5.1.1? It is important to verify the proposed method could perform well on medium-to-large datasets. 8. In Table 2, the standard deviations of the proposed method is obviously larger than SPDNet, does it imply the proposed method is unstable than SPDNet? (Similar phenomena also happen in Table 1, though not that obvious.) Implementation part: 9. As some parts (e.g., Line 225 - 231) are not that obvious to implement, will the code be released?

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors mentioned the limitations of the proposed method in Line 357 - 359.

Reviewer 1vkd6/10 · confidence 4/52023-07-05

Summary

The paper generalizes the ResNet layer to non-euclidean geometries by replacing the Euclidean sum with the exponential map. The theory is general and applies to any smooth metric. They propose a way to parametrize a vector field on the manifold which is more geometrically principled than the trivial vector field embedding. Empirically they show improvement in performance on hyperbolic datasets and PSD spaces compared to some baselines.

Strengths

The paper theory is general and applies to any smooth manifold metric. In ResNet a neural network produces a vector field and the output of the ResNet is the input plus such vector field. The proposed generalization, assuming access to a vector field on the manifold, defines the output of the ResNet as the exponential map of the input in the direction of such a vector field. This is consistent with the Euclidean case. A vector field in Euclidean space can simply be the output of a neural network, while on a manifold more care is required. Using an embedded vector field is straightforward, but, as correctly pointed out by the authors, it is not very principled geometrically. The authors propose a computationally tractable and geometrically principled way of defining a parametric vector field on a manifold. The idea is to make use of a collection of $\mathbb{R}$ valued functions (obtained as a projection on hyperplanes or similar) to define, through push forward and pullback, a vector field on the manifold. Extensive specific examples of such a collection of functions are given for the hyperbolic and PSD manifolds. Experiments are also performed in these cases, and the proposed approach appears to outperform the current state of the art. In the general manifold case (Appendix B.5), the idea of projection on pseudo-hyperplane is appealing and well-argued. And the further generalization to “hyper-disks” allows proper formal extension also to the non-geodetically-complete manifolds.

Weaknesses

The definition of the vector field (Appendix B.4) is not sufficiently formal and contains mistakes. Specifically, the author assumes access to a smooth function $f:M\rightarrow \mathbb{R}^k$, a so-called “feature map”. The differential $D_x f$ is then a linear map from the tangent space in $x$, $T_x M$ to the tangent space in $f(x)$, $T_{f(x)} R^k = R^k$. Observing that the dual of $R^k$ is isomorphic to $R^k$ itself, the pullback of the differential $(D_x f)^*$ can be seen as a map from $R^k$ to $(T_x M)^*$, for every $x\in M$. This function is evaluated in $f(x)\in R^k$ such that the vector field (as defined in line 215) is a map $$ l_f: x \rightarrow (D_x f)^*(f(x)) $$ which, as we saw, take values in $(T_x M)^*$ and NOT on $T_x M$, as line 215 is saying. We would like to see this inconsistency explained. Is this based on the observation that both $(T_x M)^*$ and $T_x M$ are isomorphic to $\mathbb{R}^{dim(M)}$? And also, can you reason about the choice of evaluating the pullback of the differential (line 215) in $f(x)$? Is this somehow principled? The equivalence with Euclidean ResNet shown in Appendix D is a proper extension in the case of an embedded vector field, but it is rather difficult to follow in the case of the feature map. Specifically, Proposition 1 is trivially proved for an embedded vector field, but the same argument should also apply to the case of a feature-map-induced vector field. It would be helpful if the discussion on $g_{w,b}$ could e.g. focus on the case of axis-aligned planes (i.e. each $w$ should be an element of a standard basis and $b = 0$), such that the differential $D_x f$ reduces to an identity. We found this to be significantly more intuitive. In the general manifold case (Appendix B.5), the idea of projection on pseudo-hyperplane is, although well explained, not at all investigated. First, it is not clear how to practically implement such projections of these pseudo-hyperplanes (not in the geodesically complete case, and even worse in the general case). Second, there are no experiments regarding general manifolds, and there is also no dissertation on the increased computational complexity with respect to hyperbolic and PSD cases. This reduces “contribution 3” in the statement in lines 72-76. *Minor:* * The related work section argues that the proposed construction is different from a neural ODE (which also generalizes ResNets), but honestly, we found this argument to be incomplete. It does seem like the proposed construction is a neural ODE. * Line 126, we found it unclear which "specific structure" is being exploited in methods using Frechet means (these averages apply on practically all manifolds). *Regarding the score:* We are willing to increase the score if these concerns are appropriately discussed in the rebuttal.

Questions

Please reply to the weaknesses above. Furthermore: Major: * Line 215 and Appendix B4. The function $l_f(x)$ is inconsistent. In the main paper is a map $M \rightarrow T M$ while in the appendix is a map $M \rightarrow (T M)^*. Can you explain the inconsistency? Minor: * Line 189 $f_nn$ should be $f$? * The Log-Euclidean metric amounts to a Euclidean metric in the tangent space at the identity. In that case, does $f$ reduce to being $f: T_I M \rightarrow R^k$? If so, is your construction just a standard Euclidean ResNet in a pre-specified tangent space? That would be good to state explicitly, if so. * It's quite common to run into numerical instabilities (esp. on hyperbolic manifolds). Is that something you face?

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

The parting "Limitations" paragraph discusses a relevant assumption. However, we feel that a discussion of the tractability of (projections onto) "pseudo-hyperplanes" (Eq. 17 in the appendix) is lacking.

Reviewer 1LjF6/10 · confidence 2/52023-07-10

Summary

The paper proposes an extension of standard ResNets called Riemannian Residual Neural Networks. The extension is done based on Riemannian manifolds as discussed in Equation (2). Some numerical results on node classification problems are presented in section 5 to show the improvements of the proposed generalization of ResNets. -- Post-rebuttal Review Update -- I thank the authors for the detailed responses to my comments. I find the responses satisfactory and raise my score to 6.

Strengths

1- The idea of Riemannian ResNets sounds interesting and as the numerical results suggest could help improve the performance of ResNet models in applications where the chosen Riemannian geometry suits the dataset.

Weaknesses

1- The paper's presentation remains abstract in the main body, and I do not find the current presentation accessible enough to deep learning practitioners. For example, Section 3 spends about 2.5 pages explaining Riemannian geometry but does not discuss a concrete example where the exponential map and vector fields can be discussed. The examples in section 4.2.1 appear late in the draft and also do not derive the expression for the exponential map that appears in Riemannian ResNets. 2- Since the paper has not discussed the algorithmic steps of training and evaluating a Riemannian ResNet, it is not that easy to see how the network can be trained for non-Euclidean Riemannian geometries. I suggest adding one or two algorithms to the draft to discuss the steps of training a Riemannian ResNet for the cases discussed in Section 4.2.1.

Questions

How did the computational costs of training a Riemannian ResNet compare to that of a standard ResNet? Could the Riemannian ResNet demand more computational power for training than the normal ResNet?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

Please see my previous responses.

Authorsrebuttal2023-08-20

Follow-up

Dear reviewer, We would like to ask if we have addressed your concerns, and if so, if it would be possible to raise your rating for the paper? If any additional questions have arisen, please let us know.

Reviewer 1vkd2023-08-14

Thanks for the rebuttal

We thank the author for the reply. However, we would like to ask for further clarification. > One motivation is that in the Euclidean case, because the maps ... will produce a standard Euclidean residual neural network. This is very interesting and can be a sufficient motivation for the choice of evaluation on $f(x)$. Can the author provide step-by-step reasoning and proof of this statement? Why will $l_f$ reduce to a standard linear layer?

Authorsrebuttal2023-08-15

Further clarification

We can certainly do so. Note for the Euclidean case, that our feature map $f : \mathbb{R}^n \rightarrow \mathbb{R}^k$ will, for example ($b=0$, $W$ has normalized row vectors), take the form $f(x) = Wx, W \in \mathbb{R}^{k \times n}$. Then note that we have $Df = W$ and $(Df)^* = W^T$. We see for the standard feature map-based construction, our vector field $\ell_f (x) = (D_x f)^* (f(x))$ takes the form $\ell_f (x) = W^T W x$. For the learnable case (which is standard for us, given that we learn Riemannian residual neural networks), note from Lines 217-219 that we have $\ell_{f,\theta} (x) = (D_x f)^* (n_\theta (f(x)))$ for $n_\theta$ a neural network. Hence we have $\ell_f (x) = W^T n_\theta (W x)$. For the case that you mentioned before, i.e., when the feature maps are trivial projections (onto axis-aligned hyperplanes), we have $W= I$ and $\ell_f (x) = n_\theta(x)$. Thus our construction can be viewed as a generalization of a standard neural network.

Reviewer 1vkd2023-08-16

Increasing score

Thank you for the follow-up. This should settle our last concerns, and we will increase our score to recommend acceptance. Should the paper get accepted, we hope that the above discussions will be incorporated into the final paper.

Authorsrebuttal2023-08-20

Thank you

We would once again like to thank you for your detailed comments and your time! We will incorporate these discussions into the paper and ensure the paper is clear with respect to the above mentioned details.

Reviewer 5HC12023-08-20

Thanks for the authors' response. It addressed my concerns. As the authors promised to release their code, I will keep my original rating.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC