Neural Isometries: Taming Transformations for Equivariant ML

Real-world geometry and 3D vision tasks are replete with challenging symmetries that defy tractable analytical expression. In this paper, we introduce Neural Isometries, an autoencoder framework which learns to map the observation space to a general-purpose latent space wherein encodings are related by isometries whenever their corresponding observations are geometrically related in world space. Specifically, we regularize the latent space such that maps between encodings preserve a learned inner product and commute with a learned functional operator, in the same manner as rigid-body transformations commute with the Laplacian. This approach forms an effective backbone for self-supervised representation learning, and we demonstrate that a simple off-the-shelf equivariant network operating in the pre-trained latent space can achieve results on par with meticulously-engineered, handcrafted networks designed to handle complex, nonlinear symmetries. Furthermore, isometric maps capture information about the respective transformations in world space, and we show that this allows us to regress camera poses directly from the coefficients of the maps between encodings of adjacent views of a scene.

Paper

References (55)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Sfsw4/10 · confidence 3/52024-06-30

Summary

The paper proposes an autoencoder framework that encodes the input symmetries into isometries in the latent space. The equivariance in the latent space is captured by a functional map $\tau$, which is regularized to be an isometry. Instead of hard constraints, the equivariance of the system is encouraged via the optimized $\tau$ and the loss function design. The proposed method is evaluated on a homography-perturbed MNIST, conformal shape classification, and camera pose estimation tasks. The experiments show the proposed method performs on par with the handcrafted equivariant baselines and outperforms a baseline with a similar approach (Neural Fourier Transform). I’m positive about the paper. However, there are some main concerns that I would like to know the answers to. I’m willing to raise the scores once they are addressed.

Strengths

1. The paper is well-written and easy to follow. 2. The paper proposes a novel framework that softly models the latent space equivariance as functional maps. 3. Unlike handcrafted equivariant networks that are limited to certain groups, the proposed framework can model different input transformations. 4. The proposed method is shown to perform on par with the baselines in homNIST and camera pose estimation tasks, and outperforms the baselines in the conformal shape classification task.

Weaknesses

1. Although being shown to be effective in the experiments, the constraint of $\tau$ being an isometry seems a bit heuristic. The paper can benefit from some theoretical investigation on why constraining to being isometries helps the performance. 2. Following up on the previous point, can all transformations be modeled as isometric functional maps in the latent space? That being said, does restricting $\tau$ to isometries affect the types of transformations the network is able to model? 3. Since the equivariance is only obtained softly via the loss function design, it’s essential to see how much equivariant is maintained/lost in the latent space. I recommend the authors report some measures of equivariance and compare them with the handcrafted methods. (For example, equivariant error/loss in the latent space.) 4. There are several heuristic designs in the proposed framework, and some of them affect the performance significantly. For example, it is not explained clearly why a smooth multiplicative mask is required to ensure $\tau_\Omega$ is semi-diagonal. The effect of such approximation is also not discussed/evaluated. The second is the multiplicity loss, which greatly affects the performance of the framework. 5. I appreciate the comparisons with handcrafted equivariant networks. However, I believe the paper can be greatly strengthened by comparing it with the same autoencoder framework while not enforcing the latent equivariance and training with data augmentation. This can demonstrate the benefits of the proposed latent equivariant design.

Questions

1. Since there are no hard-coded constraints (except for the regularization) in the system, the proposed method can work with different input transformations. Do you think this work can be applied to automatic symmetry discovery tasks? 2. I understand that the inner product needs to be preserved to construct an isometry, but why does $\tau$ have to commute with $\Omega$? 3. Ln 207, “we show in experiments that a major benefit of our isometric regularization is that our multiplicity loss (promoting a diagonal-as-possible and thus sparse and compact $\tau_\Omega$) can serve as an effective substitute for access to triples.” Is this true for all three experiments, or is it only validated on homNIST?

Rating

4

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

As mentioned in the paper, the proposed network is unable to perform graph-based tasks. In addition, the equivariance in the latent space is not evaluated. As a result, it is unclear how the equivariance is preserved in the latent space. Lastly, by restricting the functional maps to be isometries, the transformations that the network can model might be limited.

Authorsrebuttal2024-08-11

Address remaining questions in discussion period?

Dear Reviewer Sfsw, Thank you again for your comments and feedback! As the end of the author / reviewer discussion period is fast approaching, we would love to hear your thoughts and see if we can address any remaining questions in the remaining time. Please let us know whether we addressed your comments and questions appropriately! Thank you! Best, the authors

Reviewer Sfsw2024-08-11

I thank the authors for their efforts in the rebuttal. Many of my concerns have been addressed. It's good to see NIso can achieve a similar equivariance error to the handcrafted methods. It's also interesting to see patterns similar to spherical harmonies appear in the SO(3) case. It is also great to see NIso perform better than AE with augmentation. The added experiments strengthen the paper. In addition, I kindly disagree with other reviewers on the need to solve occlusion/partiality. Occlusion and imperfect symmetry modeling can be a different research topic on its own in the field. However, I recommend that the authors refrain from claiming that NIso can generally handle occlusion/partiality robustly, as it requires additional proof and verification that is not presented in the paper. Although most of my concerns have been addressed, there are still some that remain. The multiplicity loss still seems like a heuristic trick to me, and it has a significant impact on the system's performance. I would love to see a more theoretical probe into it in future work. Secondly, although it is shown experimentally that the proposed method can deal with transformations that are not unitary, it is still not guaranteed theoretically. The authors provide a plausible explanation in response to Reviewer FFcP (there exists a domain in which these transformations are either unitary or isometric, and can be mapped to via a sufficiently expressive autoencoder). Although I personally agree with such an explanation, this is still a speculation without proper proof. I again suggest the authors refrain from making such a claim. Overall, the paper is strengthened after the rebuttal, and I believe the paper introduces a novel way to softly model equivariance in the latent space, which is a great addition to the community. Therefore, I have raised my score to reflect the above points.

Authorsrebuttal2024-08-12

Thank you for your response and for updating your score, we are glad that you find NIso will contribute to the NeurIPS community! Following your comments, we will update our paper to make sure to not overclaim. Specifically, we will clarify that (1) NIso can only exactly model unitary transformations; and (2) That our experimental results only suggest the potential of NIso to model more complex, non-unitary transformations, and that a rigorous theoretical investigation combined with extensive empirical results is necessary to convincingly demonstrate broad generalization. We will also clarify that (3) NIso does currently explicitly handle partiality. We agree that a deeper investigation into the multiplicity loss will strengthen our method and we are actively investigating this in follow up work. We would also like to note that an important novel benefit of our formulation is that it allows for us to regularize for block-diagonality *without backpropagating these gradients through the solve for the estimated transformation.* Specifically, the prior methods including the NFT attempt to impose a block-diagonality loss on the estimated transformation itself. We discussed with the authors who confirmed that this is only possible in a post-processing step as it otherwise destabilizes training. In contrast, by learning to parameterize an operator $\Omega$, we instead impose the regularization for block diagonality on the mask $P_\Lambda$ which is stable during training. Our comparisons with the NFT (see Tab. 3, center left) suggest that this is a key feature enabling self-supervised discovery of sparse and condensed representations, and we are excited to investigate this further. Thank you again for your comments, which will help to clarify the paper!

Reviewer FFcP4/10 · confidence 2/52024-07-08

Summary

This paper introduces Neural Isometries, which is an autoencoder framework learning to map the observation space to a general-purpose latent space wherein encodings are related by isometries whenever their corresponding observations are geometrically related in world space. Several experiments, including camera pose estimation, are conducted.

Strengths

The idea of transforming complicated equivariances in observation space into isometries in latent space is intuitive and interesting. Experiments have demonstrated the potentials of this method in practical tasks. The paper is well written and easy to understand.

Weaknesses

This is an interesting paper but in-depth discussion is lacking. My main concern is that there is no quantitative or theoretical description of the equivariance. For example, 1. How equivariant is the model? i.e. How large is the equivariance loss? 2. What type of equivariance in the observation space can be modeled? Will the model fail when the pertubations are too strong? In addition, the experiment results are relatively weak. Only 3 small scale experiments are conducted, and only a few baselines are considered.

Questions

1. If the observation space is already isometric, e.g. rotate by 90 degrees, what can be said about the latent sapce? 2. How robust is this model? For example, how will the latent vector change when the inputs are partially observed (masked)?

Rating

4

Confidence

2

Soundness

3

Presentation

3

Contribution

2

Limitations

See the weakness section.

Authorsrebuttal2024-08-10

Response (1/2): Modeling Equivariance

Thank you for your response. We are glad to hear that you view our experimental results as convincing. ## Which Types of Equivariance can be modeled? We apologize for the lack of clarity in our previous response. In the following, we will make two points: (1) We show that NIso can exactly model equivariance for any unitary transformations, which make up a large chunk of previously proposed equivariant ML architectures. (2) We provide detailed reasoning for our empirical results that clearly suggest that NIso can learn to be equivariant even to certain non-unitary groups. ### Equivariance to unitary transformations NIso can learn to be equivariant to any transformation that preserves the norm of its input. Examples for this include: Shifts on the torus, as demonstrated in Fig. 3 of the main paper. SO(3) with spherical inputs, as demonstrated in Fig. 1 of the rebuttal PDF. The group of 90 degree rotations mentioned by the reviewer, which we empirically validated but were not able to include in the rebuttalPDF due to space constraints (we will include this in the camera-ready paper!) We note that prior work in equivariant ML at top conferences is often equivariant to only **a single one** of these transformations, with significant expert crafting. For instance, SO(3) is a non-abelian group and its IURs are substantially complex – nevertheless we are able to recover a basis and representation of transformations with *exactly* the same properties. The fact that NIso learns to be equivariant to all these transformations is a significant strength - we are not aware of any previous work in symmetry discovery that has demonstrated this capacity. We hope that this is absolutely un-ambiguous: **NIso can learn to be equivariant to a transformation that preserves a norm of its input**. ### Equivariance to non-unitary transformations We agree that the statement “Any transformation can be approximated with a sufficiently strong encoder” was too loose. We specify it in the following. As discussed above, NIso can closely approximate **unitary** transformations on its input. For other more complex transformations, our experiments suggest that a sufficiently expressive autoencoder can, in fact, discover a latent space where the manifestation of the transformation is approximately unitary and even isometric. We believe this is a surprising and interesting result that is perhaps being overlooked. We note that it is not understood which transformations or groups can be embedded in this way, and point out that this is an exploration of fundamental math, beyond the scope of any ML paper. However, a potential explanation of this phenomenon could be related to the fact that for many challenging transformations, there exists a domain in which these transformations are either unitary or isometric. For example, Mobius transformations are isomorphic to the isometries of the hyperbolic ball with their action on the sphere representing their restriction to the boundary. Similarly, camera motions in the image plane are in fact projections of the isometric action of SE(3) on the underlying 3D scene. We hypothesize that by forcing the network to find an isometric representation for the observed transformations we are implicitly regularizing the learned latent space to discover information about this underlying geometry. We note that our ability to recover reasonable camera poses from the estimated transformations where other representation learners fail is indicative of this. An additional illustrative example supporting this hypothesis can in fact be seen in Figure 7 of the supplement, where we show examples of failure cases in which the pose extraction from $\tau_{\Omega}$ fails catastrophically. Here we observe that these failure cases often coincide with sequences of frames where background and foreground objects are respectively far from and close to the camera – both cases where depth-estimation pipelines also often fail. ### Conclusion All in all, NIso is a novel *first step* towards a new paradigm for symmetry discovery and equivariant ML. To the best of our knowledge, no prior work in symmetry discovery has been able to show, for instance, self-supervised discovery of approximate equivariance to a variety complex transformations, some of which do not even form a group in the observation space (such as the action of camera motions on the image plane). Using the additional page of the paper, we will add above discussion to the “Methods” section of the paper - we agree with the reviewer that this will make the paper stronger!

Authorsrebuttal2024-08-10

Response (2/2): Analysis of Partiality Robustness

## Analysis: Robustness to Partiality In response to the reviewer’s concerns, we are happy to report that we performed a theoretically grounded experimental evaluation of robustness which we will be happy to include in the revision using the additional page of the camera-ready paper. Specifically, we seek to evaluate the robustness of the isometry-estimation module in isolation in the presence of occlusions. We consider the encoder-free paradigm – consisting of the projection to the learned basis, the estimation of the isometric map as in Equation (8), followed by unprojection – which is the same setup in the toric and spherical laplacian experiments. **This setup allows us to study exactly the effect of masking on the latents (as requested by the reviewer) which in this case are the input images.** To do so, we consider two observations $\psi$ and $T\psi$ which only partially correspond and denote $O_\psi$ and $O_{T\psi}$ to be the diagonal overlap masks such that the $i$-th diagonal element of $O_{\psi}$ is 1 if $i$ is the overlap with $T\psi$ and $0$ otherwise, with $O_{T\psi}$ defined in the same manner. We observe that the equivariance error under the partiality can thus be defined by the magnitude of the difference between the components of $O_{\psi} \psi$ that get mapped to $ O_{T\psi} T\psi $ under $\tau$. That is, we define the partiality equivariance error to be $ \lVert O_{T\psi} \tau O_{\psi} \psi - O_{T\psi} T\psi \rVert^2 / \lVert O_{T\psi} T\psi \rVert^2$. Using the toric laplacian experiments as a base, we consider two models of partiality. In the first, we no longer consider the domain to be toric and instead shifted images are clipped at the boundaries with the resulting empty pixels masked to zero – corresponding to the type of partiality often observed in video. In the second, we randomly mask out $2 \times 2$ patches in the shifted image. We train five instances of our model under both partiality regimes, masking out approximately 10%, 20%, 30%, 40%, and 50% of pixels in each instance and measuring the resulting partial equivariance error on the test set. The results are shown in the table below. | Percent Occluded | 10% | 20% | 30%| 40%| 50% | | :------------------------ | :-----: | :-----: | :-----: | :-----: | :-----: | | Shift Mask | 5.66% | 9.42% | 14.43% | 17.42% | 20.47% | | Patch Mask | 6.69% | 13.62% | 23.54% | 34.14% | 45.35%| Notably, we see that the partial equivariance error is consistently lower in the presence of shift-based occlusions, and increases less as the percentage of occluded pixels increases. We observe that the principal difference between these two regimes is that the unoccluded pixels exist in a contiguous block under the shift mask, whereas the region of unoccluded pixels is fragmented under the patch masking. Intuitively, the contiguous matches between large blocks (comprising the majority of the image) act as a strong regularizer that de-prioritizes matching the occluded areas. Conversely, when the correspondence is interrupted and fragmented it offers weaker regularization and the resulting map is more affected by the spurious matches induced by the occlusion.

Authorsrebuttal2024-08-12

Additional question remaining in discussion?

Dear FFcP, Thank you again for your response and feedback! We believe we were able to address your stated remaining concerns, including a straightforward experimental evaluation of robustness. As the end of the author / reviewer discussion period is fast approaching, we would like to be sure you are satisfied by our discuss and see if we can address any remaining questions. Please let us know whether we addressed your comments and questions appropriately! Thank you! Best, the authors

Reviewer JQqD6/10 · confidence 3/52024-07-14

Summary

This paper proposes a generic equivariant ML framework by learning latent representations are are modelled to be related by an isometry. There are several important design choices made by the authors: (1.) An autoencoder framework that keeps the spatial structure (i.e. images get encoded into images) (2.) The isometric map is represented compactly with a basis that is the outcome of an eigen- -decomposition of a PSD operator (3.) The isometry manifests as an orthogonal matrix with a sparse block diagonal structure in its reduced projected form. Experiments are shown for (1.) Learning the spectrum of a Toric Laplacian and comparing it qualitatively to the real one in Fig 3 (2.) Homography perturbed MNIST (3.) Conformal shape classification and (4.) Camera Pose Estimation from real-world video. The results (1) show a good proof of concept and (2), (3) and (4) demonstrate a clear superiority over the closest conceptual baseline NFT. (2.) Also makes the case for achieving comparable performance with simpler equivariant architectures using the NIso latent space.

Strengths

- Overall the biggest strength of this paper is that it was an enjoyable read. I found the hypothesis (i.e. learning latent codes related by an isometry for arbitrary transformations in the observation space) to be interesting and reasonable. The overall writing and description of the method is very nice. - The Experiments especially (1.) was a good proof of concept.

Weaknesses

- I am actually quite confused on how the operator \Omega and Mass matrix M is learned? The Laplace Beltrami is a very structured object. See all requirements enumerated in Wardetzky, Max, et al. "Discrete Laplace operators: no free lunch." Symposium on Geometry processing. Vol. 33. 2007.). It is unclear what the nature of this operator is and its learning feels ad-hoc and unmotivated. - Do the authors make sure that the optimization for the functional map is well-posed? Typically in the original formulation a "good" functional map is the outcome of using many-many descriptors in addition to regularization constraints like commutativity with laplacian. If I understand correctly the proposed formulation seems to be confident in learning a structured functional map (like Fig 2) using just one pair of corresponding functions ? - As a high-level opinion: Overall the experimental section could be stronger. The "discovery" of the toric laplacian was nice, but a demonstration on roto-translations perhaps with scale and/or spherical images with SO(3) would be extremely convincing. I would also recommend an evaluation that focuses beyond NFT and actually quantifying the complexity of different equivariant models. - As mentioned by the authors, the requirement of generalizing to domains with diverse connectivity is indeed a hard one, but important.

Questions

- The basis functions visualized in Fig3 (third column) are learned or from the actual toric laplacian? what are the various rows? - It would also be useful to report some feature of complexity in Table 1 indicating "meticulously-engineered, handcrafted networks" - More examples of evidence showing good basis functions - What is a perturbation in Section 5.1?

Rating

6

Confidence

3

Soundness

2

Presentation

4

Contribution

3

Limitations

I think the authors make a fair declaration of the limitations of their work in Section 6. I would make a stronger vote after the rebuttal. I think this paper has many ingredients to warrant acceptance: clever idea, good argumentation, and decent proof of concept. It lacks clarity in some places and the evaluation does not appear equally general as is the description of the method. For this reason, I maintain a borderline acceptance with a full intention to re-assess in the next round.

Authorsrebuttal2024-08-11

Address remaining questions in discussion?

Dear JQqD, Thank you again for your comments and feedback! As the end of the author / reviewer discussion period is fast approaching, we would love to hear your thoughts and see if we can address any remaining questions in the remaining time. Please let us know whether we addressed your comments and questions appropriately! Thank you! Best, the authors

Reviewer JQqD2024-08-12

Rebuttal response

Thank you for a refreshing rebuttal. After reading through the rebuttal and other reviews, I am quite happy to raise my score to a clear acceptance. This is an interesting paper with some fascinating observations. However, there is still some way to go before this can really be generalized and put into meaningful action. Sustaining drawbacks are some very hard questions like - how much data is needed for this approach? Equivariance to multiple non-trivial (perhaps non-unitary) transformations of the input etc. But I think as a good proof of concept, I believe this paper tells a story that would be interesting at Neurips. I would highly recommend including the spherical image experiment in the main draft (also what kind of architecture did you use for the encoder?)

Authorsrebuttal2024-08-12

Thank you for your response and we are extremely glad to hear you are happy to recommend a clear acceptance! We completely agree that NIso represents a first step towards identifying underlying symmetries and that further investigation is necessary to support claims of broad generalization. The reviewer is correct to point out that answering difficult questions like those regarding data efficiency and effectiveness in the presence of multiple, different non-unitary actions are important prerequisites. We will amend the revision to reflect this and make sure that we communicate limitations on partiality and transformations modeled clearly. We will incorporate the spherical harmonic transform experiment into the main body of the revision as the results are very compelling (thank you again for suggesting this experiment!). In both this experiment and the toric laplacian experiment, the encoder and decoder are just the identity map. We pass the images directly to the transformation estimation module, learning only the parameterization of the operator with the goal of evaluating the abilities of the module in isolation.

Reviewer 2Npe4/10 · confidence 4/52024-07-15

Summary

The main motivation behind Neural Isometry is as follows: most of real world transformations in vision and geometry processing lack identifiable group structure and therefore challenging for prior work in equivariant learning that assumes such knowledge apriori. The paper proposes an auto encoder framework that maps observations (related by some geometric transformations) to a latent space where embeddings are related by a linear transform. The framework leverages the existing functional map framework and corresponding regularization to learn this latent space. The paper validates and compares their performance with the baseline NFT (Neural Fourier Transform) that assumes the knowledge of the group a priori.

Strengths

- The problem of learning structured latent space in a self supervised equivariant learning set up is relevant. - The proposed framework is formalized clearly and in detail in Section 4.

Weaknesses

As mentioned many times in paper, Instead of theoretical justification ( as in NFT), this paper seeks to validate the efficacy of approach experimentally on real world tasks. - Limited Experimental Set up: The datasets used are still very small scale (MNIST,) or synthetic (SHREC) and therefore far from the real world data. The submission mainly follows and compares with one baseline (NFT) that was implemented by authors themselves. The submission could motivate the problem/solution by probing/playing with the learned structured space e.g. in geometry processing tasks [1]. - Limitation of Functional map: One of the biggest limitation of functional map framework is its applicability to real world data where partiality is ubiquitous (due to occlusion etc). Therefore, Neural Isometry by design also inherits this limitation. experiments are shown on MNIST/SHREC where there is no occlusion or partiality. In the third benchmark, instead of skipping the odd frame, does the performance decays rapidly between distant frames? - Presentation: The submission contains several typos, convoluted sentences (Line 85-88) and statements without a citation. e.g. -- Line 49: < Lorentz transformations with d'almbert operator of Minkowski space> citation is missing here. Also, why do we need to know this fact in the introduction of this paper? -- Line 42: <by preserving the spatial dimensions> not sure what does this mean after reading the next 2 lines. Please clarify -- Line 89: <orthogonal relaxation of FM> citation or please detail what those are. - Related work: The submission cites a dozen variants of deep functional map paper [2] but not the deep functional map paper itself [2] that inspired them. Please justify how [20,21,22,39] are related to this work or inspired this work given other deep functional pipeline cited in paper. The reviewer would instead relate/distinguish this work with [3]. 1. Composite Shape Modeling via Latent Space Factorization, Dubrovina et al, 2021 2. Deep Functional Maps Litany et al. 2017 3. Map-based exploration of intrinsic shape differences and variability Rustamov et al. 2013

Questions

please see above.

Rating

4

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

The submission should state the scalability of this approach given its reliance on Laplacian Eigen basis.

Authorsrebuttal2024-08-11

Questions remaining?

Dear 2Npe, Thank you again for your comments and feedback! As the end of the author / reviewer discussion period is fast approaching, we would love to hear your thoughts and see if we can address any remaining questions in the remaining time. Please let us know whether we addressed your comments and questions appropriately! Thank you! Best, the authors

Reviewer FFcP2024-08-09

Thanks for your clarification. I have read the experiments in the attached pdf, and I am clear now that the equivariance error is at least comparable with the existing methods. The reply on the experiments is also convincing. However, the other two concerns are not addressed: the type of equivariance can be modeled and the robustness to partial visibility. The author argued that "Any other transformation can in theory be approximated with a sufficiently expressive latent space". It is imprecise. On the other hand, due to the lack of description of the robustness against partial visibility, and the model does not have any constraint on the equivariance of the input, it is hard to understand why the model is robust or how robust I should expect it to be for practical tasks. The author mentioned that this issue will be investigated in the next step, but since current version of paper does not include any discussion on this issue, I think it is not complete.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC