Learning Infinitesimal Generators of Continuous Symmetries from Data

Exploiting symmetry inherent in data can significantly improve the sample efficiency of a learning procedure and the generalization of learned models. When data clearly reveals underlying symmetry, leveraging this symmetry can naturally inform the design of model architectures or learning strategies. Yet, in numerous real-world scenarios, identifying the specific symmetry within a given data distribution often proves ambiguous. To tackle this, some existing works learn symmetry in a data-driven manner, parameterizing and learning expected symmetry through data. However, these methods often rely on explicit knowledge, such as pre-defined Lie groups, which are typically restricted to linear or affine transformations. In this paper, we propose a novel symmetry learning algorithm based on transformations defined with one-parameter groups, continuously parameterized transformations flowing along the directions of vector fields called infinitesimal generators. Our method is built upon minimal inductive biases, encompassing not only commonly utilized symmetries rooted in Lie groups but also extending to symmetries derived from nonlinear generators. To learn these symmetries, we introduce a notion of a validity score that examine whether the transformed data is still valid for the given task. The validity score is designed to be fully differentiable and easily computable, enabling effective searches for transformations that achieve symmetries innate to the data. We apply our method mainly in two domains: image data and partial differential equations, and demonstrate its advantages. Our codes are available at \url{https://github.com/kogyeonghoon/learning-symmetry-from-scratch.git}.

Paper

References (39)

Scroll for more · 27 remaining

Similar papers

Peer review

Reviewer 4Qv35/10 · confidence 3/52024-06-14

Summary

This paper proposes using neural ODEs to parameterize symmetries by viewing the ODEs flow as an element of a one-parameter group. They show that by learning the parameters of the neural ODEs, they are able to recover ground truth symmetries in image classification and PDE tasks.

Strengths

The paper is easy to read. The proposed ideas are clear, and appear to be mostly novel.

Weaknesses

See questions below.

Questions

1. Line 174: Do you also need "and for all $s\in[-\sigma,\sigma]$" in addition to the "for all $f\in\mathcal{D}$"? 2. Section 4.1: According to your definition of a symmetry $\vartheta$ (there exists $\sigma>0$ such that $\vartheta_s(f)$ is "valid" for all $f\in\mathcal{D}$ and all $s\in[-\sigma,\sigma]$), it seems like solving for $\vartheta^*$ from (4) does not guarantee that $\vartheta^*$ is actually a symmetry, even if the optimal value of (4) is less than $C$. That is, if the optimal value of (4) is less than $C$, then you only know that the validity score is less than the threshold on average, not for all data $f$ and all transformation scales $s$. Can you please clarify this aspect? 3. Line 217: Something is strange with the definition of $\vartheta_{\mathcal{X}}(f)$. In particular, what is $T_{\mathcal{X}}$, and why does $f$ not appear in the right-hand expression? It seems to me like perhaps you meant to define $\vartheta_{\mathcal{X}}(f)(x) = f(\vartheta_{\mathcal{X}}^{-1}(x))$, and hence that you need to assume the transformation $\vartheta_{\mathcal{X}}$ on $\mathcal{X}$ to be invertible. Can you please clarify whether this is a typo, or if I am missing something? Also, it is unclear what $\vartheta_{\mathcal{Y}}(f)$ is doing or why it is needed. Indeed, as written, you are defining two different transformations of $f$ in (8) (resulting in two different functions of $x$). 4. Line 226: How reasonable is it to assume that you know the actual number of possible symmetries for a given task? Is it possible for you to select $N_{\textup{sym}}$ to be too small, and then to miss out on learning some of the most important symmetries of a task? 5. Line 232: Typo ("orthonomral"). 6. Line 230: In practice, are symmetries actually mutually orthogonal (as vector fields) to one another? It seems like this is not always the case per your Table 1, which shows that uniform scaling is not orthogonal to $x_1$-axis scaling. A comment on the validity of the orthogonality assumption underlying the use of this regularizer would be appreciated. 7. Line 243: Typo ("undersirable"). 8. Line 246: Typo ("Lipshictz"). 9. Line 280: "...the two translations are the most invariant one-parameter group..." Can you please explain in further detail how you come to this conclusion? 10. Table 2 and Figure 5: I suggest moving these to be centered with the rest of the text; it looks poorly placed in the right-hand margin as is, and causes Figure 5 values to be too small and hard to read. 11. Line 322: "We found the ground truth symmetries in all the experiments as in Figure 5..." It looks like you did not recover the true symmetries in the cKdV setting, as there are non-negligible components coming from each of the ground truth symmetries that appear in the learned symmetries. 12. Did you verify in your experiments that the learned "symmetries" are actually symmetries according to your definition using the validity score threshold? Overall, your experimental results in Figures 4, 5, and 10 seem to me like you are indeed learning orthogonal vector fields, and that sometimes the vector fields end up being equal to one of the ground truth vector fields, but other times not (turns out to be linear combination of multiple ground truths). This does not seem super convincing that you are reliably learning/recovering the ground truth symmetries. 13. Why did you introduce (8)? I don't see this transformation being used anywhere else in the paper.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

N/A.

Reviewer 7EZa6/10 · confidence 4/52024-07-08

Summary

The paper pertains to the topic of data-driven symmetry discovery. The authors propose a method allowing symmetry discovery beyond pre-defined Lie groups, by learning to transform datapoints, potentially in a non-affine manner, via a learned ODE (referred to as the *one-parameter group*, where the single parameter is the time variable in the ODE), the velocity field of which is typically parametrised by an MLP. Crucially, to optimise this, the authors choose an objective - *validity score*, which is a predefined measure of the extent to which a transformed datapoint is symmetric to the input one (in their examples: for images, they use the cosine similarity between features extracted by a pretrained NN, while for PDEs, they measure the value – error – of the PDE for the transformed datapoint). Additional regularisers are used to ensure the diversity of the symmetries learned (orthogonality between different learned velocity fields) and smoothness (minimisation of an estimate of their local Lipschitz constants). Experimentally, the method is tested on image classification (CIFAR10) and PDE solving (KdV, KS, Burger’s equations) showing that known symmetries are retrieved along with additional approximate symmetries, while the learned symmetries are subsequently used for data augmentation showing competitive results to methods using pre-defined augmentations.

Strengths

**Significance** . The paper studies an important problem (*data-driven symmetry discovery*) in machine learning, but also physical sciences where symmetries are abundant but potentially unknown. Identifying unknown symmetries and incorporating them in downstream ML models (e.g. via augmentation) can improve generalisation, especially in low-data regimes, while additionally, it can potentially provide novel insights about the task at hand. **Novelty/Generality** - The presented methodology has the capacity to recover symmetries arising from *non-affine* data transformations. This is contrary to prior work, where mostly linear/affine transformations are dealt with. - Additionally, this method does not require making assumptions about the structure of the target group. This is common in prior art, where typically a subgroup of a predefined group is learnt. - The authors take advantage of well-established concepts that are underexplored by the ML community (e.g. modelling transformations via the one-parameter group) - this helps to broaden the available toolbox in the field of ML & symmetries/ equivariant ML. **Execution/Implementation** - Although the proposed method has multiple complicated components (NeuralODEs, difficult objective to optimise for), it is nonetheless well-executed yielding competitive results and recovering known symmetries in popular testbeds.

Weaknesses

**Applicability and scope**. Perhaps the biggest limitation of the proposed method is the *reliance on the validity score*. Although the authors claim to be able to learn symmetries by making as few assumptions as possible (see strengths), this seems to be contradicted by the need to manually design a validity score. Moreover, I have the impression that the validity score is not merely a hyperparameter, but it is decisive for the symmetries that will be learnt (basically it is the objective function of the optimisation problem). - For example, in the case of images, the choice seems ad hoc (closeness in the representation space of a pre-trained encoder NN). What leads the authors to believe that the features of equivalent (symmetric) images extracted from the pre-trained NN should be close? Have the authors tried to verify this assumption? I think the empirical validation is insufficient here (section 5.1.), so I am not entirely convinced. - In general, I do not see a generic way to define validity scores and perhaps the authors have slightly overclaimed in that respect. I would like to read the authors' viewpoints on that. For PDEs, the validity scores are indeed reasonable and generic, so perhaps, they would like to put more emphasis on this perspective. Furthermore, the authors introduce the concept of learning symmetries via the one-parameter group, claiming that it is more general than prior parametrisations that can only learn linear/affine groups. However, it is unclear what the present parameterisation can express, e.g. does it allow learning any continuous group or implicit assumptions are made here as well? - Additionally, could the authors discuss if it would be possible to learn finite groups with this method as well and if not, how could those be incorporated? **Related Work/Comparisons**. The work of Forestano et al., MLST’2023 is quite relevant to the present manuscript, with the main difference being that in that work, the transformations are parameterised by an MLP instead of a NeuralODE (the oracle used in this work seems similar to the validity score used here). Since the two works have many similarities, I think that the authors should discuss in more detail their differences and the advantages of their work (e.g. as far as I understand the MLP cannot guarantee that the transformations form a group). Note that modelling transformation via an MLP (or any NN in general) instead of a NeuralODE seems more straightforward and probably easier to train and more computationally friendly. **Experiments**. I believe some additional empirical evidence would strengthen the authors' claims. - Most importantly, an experimental comparison against the type of parameterisation used in Forestano et al. (MLP) should be provided, to verify if NeuralODEs are indeed a more appropriate parameterisation. - Moreover, baselines are mostly missing, e.g. comparing against other methods for data-driven symmetry discovery (I am not super familiar with these works, but if I am not mistaken LieGAN by Yang et al., ICML’23 is a recent example). - The reported results after augmenting with the learned symmetries do not seem to improve significantly compared to known/default augmentations. Can the authors discuss why this might be the case? This is important since it might undermine the necessity of the proposed approach. To be more convincing, perhaps the authors should perform experiments on problems where the symmetries are not known a priori. - Additionally, ablation studies seem to provide important insights but are only discussed in the appendix. I would recommend being more upfront in the main paper and discussing in the rebuttal the following: sensitivity to hyperparameters (multiple are needed: loss coefficients, $\sigma$ and $\tau$), the method for choosing them, the difficulty of optimisation (3 terms are used in the objective) and if all losses are optimised in a balanced manner. Similarly, for the parameter $N_sym$, which is now chosen based on prior knowledge of the number of existing symmetries. **Presentation/Exposition**. (disclaimer - this is a minor weakness) Given that the notions discussed here are probably not widely known across the ML community, I believe that the authors should aim to provide more in-depth explanations to make their work more accessible. For example, - Multiple group theory/symmetry concepts are discussed without definitions (group generators, Lie group, Lie algebra, Lie bracket etc.). Additional examples that are not well-contextualised include in section 2 the one-parameter group discussion, the Lie algebra of the affine group and the discussion on the PDE symmetries (Lie point symmetries etc.). Adding some references here and providing some examples for the mentioned PDE symmetries would help. - In section 5.2., some concepts regarding the experimental details are mentioned without appropriate explanations, while others are only mentioned in the appendix, although it appears that they are crucial for the method. Perhaps the authors should be more upfront and explanatory regarding the aforementioned.

Questions

- How is the weight function in L234 defined, and how important is this for optimisation? Are different weight functions ablated? - It’s unclear why the stop-gradient is needed in L237. I did not find the justification fully convincing. Could the authors elaborate? What happens experimentally if one does not use stop-gradient? - Although intuitive, it’s unclear why the inline Eq. in L212 holds for negative $\alpha$. **Suggestion**. In case the authors do want to present their method as generic, I think a deeper experimental evaluation in data beyond PDEs would help a lot (e.g. testing on other image datasets, ablating different validity scores etc). **Minor**: - What if the chosen validity score is not differentiable? - *Notation*. - The notation $\theta^V_s(x)$ is a bit dense (why use V as a superscript?). Eq (1) is a bit confusing. The inline Eq in line 83 sees clearer to me. - Notation in sec. 4 could be also simplified a bit or made more accessible with examples (e.g. 4.1: give an example for $\mathcal{A}$, i.e. the set of all natural images).

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

Some of the limitations are adequately discussed. An important missing point is, in my opinion, the need to manually design the validity score in domains beyond PDEs. No foreseeable negative societal impact.

Authorsrebuttal2024-08-12

Thank you for your comment. - Q1. We indeed acknowledge that choosing the validity score is critical and not direct. However, we clarify our stance: "symmetry" is not a fixed mathematical concept but a task-dependent jargon. For example, isometries of a metric space are transformations that preserve the metric, while symmetries of a physical system preserve the Hamiltonian or governing equations. Image symmetries are typically not rigorously defined but refer to transformations that maintain invariance to human perception; we replace "human eyes" with "neural network" in this context. - Q2. **The two experiments we have done are exactly that.** We replaced the neural ODE with an MLP that does not involve ODE integration, which now requires taking Lie derivatives (i.e. derivative along the vector field modeled by an MLP) to the validity score. We compared these setups using both image data and a toy experiment searching for isometries in Minkowski space. While using Lie derivatives (i.e., a plain MLP) fails with the image task, our method succeeds in both scenarios.

Reviewer UU8r6/10 · confidence 3/52024-07-14

Summary

This paper proposes a symmetry learning algorithm based on transformations defined via infinitesimal generators. Using Neural ODE, an infinitesimal generator is learned that is capable of producing a sequence of transformed data through ODE integration. Validity score has been defined to check if the transformed data is valid wrt to a given task. For images, the validity score is picked to be cosine similarity while for PDEs the validity score is defined as the numerical errors in the original equations after the transformation. In addition to symmetry loss, two regularizations, orthonormality loss, and Lipschitz loss have been added, to remove trivial solutions. The authors present experiments on CIFAR10 and KdV equation and Burgers' equation in 1D for PDE.

Strengths

1. The paper motivates the need for learning continuous symmetries well. 2. The idea presented in the paper is very neat and shows potential beyond the presented experiments. 3. This approach can learn both affine and non-affine symmetries in image classification and PDE tasks as shown in the experiments section.

Weaknesses

The discussion on compute and model parameters comparisons with baseline missing. No other method was shown as a baseline in either of the experiments. There is some ambiguity in how exactly the validity score is used and in some cases, can be defined if the given task is equivariant.

Questions

1. How does the validity score take into consideration, invariant tasks vs equivariant tasks? 2. Using cosine similarity between the extracted features for validity score, can be a problem for extreme transformations (as cosine values do not linearly change with parameter). How does this affect the learning of symmetry build on validity scores? 3. In Table 2, it is unclear how the validity score players out in learning the symmetries. Especially for default augmentation case. Could you please elaborate on this? 4. Total loss and loss-scale section: How is the true scale learned, if the inner product is normalized? Additionally if all the terms are affected by a data-dependent scale, shouldn't $w_{Lips}$ affect the overall weights? ### Clarifications 1. The figures can be made a little bigger. 2. In line 16, 'whether the transformed data are still valid for the given task'. This phrasing is confusing. 3. In line 172, what does valid mean? simply put; does it comprise invariant transformations and approximately equivariant up to the threshold C? 4. Intuition on how to locate transformed data on grid is missing.

Rating

6

Confidence

3

Soundness

2

Presentation

2

Contribution

3

Limitations

1. The authors do not compare with other methods of learning symmetry, as those methods use datasets that have explicit symmetry (like Rot-MNIST). Similarly for that of the PDE experiments. It would have been useful to have them in the paper. 2. The number of parameters and compute comparison across the proposed method and baseline approach like [9,14] is missing. Also discussion the scale of numerical computation with an increase in dimension missing.

Reviewer 7EZa2024-08-12

Post rebuttal

I thank the authors for their response. - Regarding the validity score, unfortunately, I did not find the response satisfactory. Arguably, choosing the validity score is a central element of this work, therefore, I strongly encourage the authors to be upfront about it and clarify that this is a limitation of the approach and that (perhaps strong) assumptions are needed in domains beyond PDEs, where the choice is natural. - Regarding the comparison with baselines, I was not able to locate an exact response to my concern. Simply put, my point was to replace the NeuralODE with an MLP (no ODE integrator, just direct prediction of the transformation). This is a natural baseline and would empirically support the necessity of employing a NeuralODE. Have the authors performed this experiment? - I also encourage the authors to include in an updated version of their manuscript the discussion regarding the groups that can be learned (continuous and differentiable, but not finite). I keep my recommendation for acceptance, but as per my review, I have certain important reservations.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC