Symmetries in Overparametrized Neural Networks: A Mean-Field View

We develop a Mean-Field (MF) view of the learning dynamics of overparametrized Artificial Neural Networks (NN) under data symmetric in law wrt the action of a general compact group $G$. We consider for this a class of generalized shallow NNs given by an ensemble of $N$ multi-layer units, jointly trained using stochastic gradient descent (SGD) and possibly symmetry-leveraging (SL) techniques, such as Data Augmentation (DA), Feature Averaging (FA) or Equivariant Architectures (EA). We introduce the notions of weakly and strongly invariant laws (WI and SI) on the parameter space of each single unit, corresponding, respectively, to $G$-invariant distributions, and to distributions supported on parameters fixed by the group action (which encode EA). This allows us to define symmetric models compatible with taking $N\to\infty$ and give an interpretation of the asymptotic dynamics of DA, FA and EA in terms of Wasserstein Gradient Flows describing their MF limits. When activations respect the group action, we show that, for symmetric data, DA, FA and freely-trained models obey the exact same MF dynamic, which stays in the space of WI laws and minimizes therein the population risk. We also give a counterexample to the general attainability of an optimum over SI laws. Despite this, quite remarkably, we show that the set of SI laws is also preserved by the MF dynamics even when freely trained. This sharply contrasts the finite-$N$ setting, in which EAs are generally not preserved by unconstrained SGD. We illustrate the validity of our findings as $N$ gets larger in a teacher-student experimental setting, training a student NN to learn from a WI, SI or arbitrary teacher model through various SL schemes. We last deduce a data-driven heuristic to discover the largest subspace of parameters supporting SI distributions for a problem, that could be used for designing EA with minimal generalization error.

Paper

References (60)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer aSNH7/10 · confidence 2/52024-06-21

Summary

This paper studies the Mean-Field limit of generalized shallow neural networks after learning with Wasserstein gradient flow under data augmentation, feature averaging or equivariant architectures as well as standard training. The provided results provide insights into learning with symmetries by covering two notions of invariant laws on the parameter space, varying assumptions on the data distribution and preserved symmetry over the course of training.

Strengths

The paper provides several deep and quite general theoretical results on learning with symmetries in the mean-field limit. Although the results are quite technical, the authors manage to present them in good clarity and embedded in the related literature of mean field limits of shallow models. Considering the examples of Data Augmentation, Feature Averaging and Equivariant Architectures seems natural and provides several interesting insights, for example the equivalence of DA, FA and even free training under symmetric data.

Weaknesses

The paper already seems quite complete and well written and I could not identify significant weaknesses. The following weaknesses are only minor points: - The description of Figure 1 is hard to parse. Markers like (a)-(d) could reduce ambiguity of what is meant by a column. - The markers in Figure 2 should be larger and it is unclear to me which additional insight to draw from the second row. As I am not an expert in the mean-field and Wasserstein gradient flow literature, I did not verify the correctness of all proofs, and did not double-check the claimed novelty. **Typos:** 22: lastly, 65 and 66: b=N?, 75: Definition name in bold for consistency?, 120: […], 275: will, 329: a heuristic

Questions

- How close is your setting to practical neural networks trained with SGD? In future work, are there hopes to extend your mean-field theory results to deep networks or to SGD with large learning rates and edge of stability dynamics? Similarly is the entropy regularization merely theoretically pleasing or also practically feasible? - Can GNNs or Transformers also be encoded in your class of equivariant architectures? If so, these example would significantly broaden the target audience. - In future work, in which ways could Assumption 1 be relaxed?

Rating

7

Confidence

2

Soundness

4

Presentation

3

Contribution

4

Limitations

The authors adequately acknowledge the main limitations of the paper. These limitations lie beyond the scope of this paper, but pose interesting directions for future work: - The provided results are only asymptotic and only consider gradient flow on shallow models as opposed to edge of stability dynamics in finite deep models. - The practical relevance of the findings and proposed algorithm at finite width for practical architectures, datasets and optimization procedures could be empirically verified. In particular, does vanilla training remain SI throughout training beyond the simple student-teacher setting?

Reviewer 4cbm7/10 · confidence 3/52024-06-28

Summary

The paper investigates the learning dynamics of neural networks with several symmetry-leveraging techniques from a mean-field perspective. The main result indicates that optimizations with data augmentation and feature averaging (and the corresponding Wasserstein gradient flows) are equivalent in the mean-field limit under mild assumptions, while equivariant architecture may suffer from limited expressivity. Moreover, for equivariant data, freely-trained models also have the mean-field flow remaining in the set of strongly invariant distributions if initialized there, which is supported by the experiments. Finally, the authors propose a heuristic to discover equivariant architectures by gradually expanding the equivariant subspace.

Strengths

* The math in the paper is rigorous and detailed. Complete proofs and abundant examples and remarks are provided. As more of a practitioner myself, I can understand the main ideas of the paper without much difficulty, though I did not check the proofs carefully. * The experiments are compact and well-designed, clearly supporting the theoretical results.

Weaknesses

* This is not essentially a weakness, but the theoretical results don't bring many interesting insights to me. They mostly match my existing intuitions. E.g. for equivariant data, freely-trained models tend to have the same training dynamics as symmetry-aware models. I will leave it to other reviewers to comment on the theoretical significance of these results. * I suggest the authors provide more detailed explanations about some mathematical terminology, e.g. the pushforward of a probability measure; dt-a.e. standing for almost everywhere wrt dt. This could make the paper more accessible to the general audience.

Questions

* Can you explain what is the connection between jointly equivariant and equivariant? The way I understand it is that we have some group actions on $\mathcal Z$ induced from the definition of joint equivariance, i.e. $\sigma_* (\rho_g x, M_g z) = \hat\rho_g \sigma_*(x,z)$. In other words, $M_g$ is determined by the group action on data and the model architecture. Then, if $M_g z = z$, then a NN model parametrized by $z$ is equivariant. Is this correct? * Proposition 5 states that equivariant architectures can lead to optimal solutions provided some universality conditions. How restrictive are these conditions? * Is the property of data equivariance (defined in L150) equivalent to the marginal distribution on $\mathcal X$ being invariant and the function $f: \mathcal X \to \mathcal Y$ being equivariant, given that the mapping from $\mathcal X$ to $\mathcal Y$ is deterministic? If so, would it be too restrictive to assume the invariance of $\pi_\mathcal X$? * In Section 4.1, the authors show that SI-initialized but freely-updated models remain in $\mathcal E^G$ for large $N$s. This is indeed an interesting phenomenon. I wonder if this is related to the exact WI-initialized teachers. Can you provide, in addition to the arbitrary particles in Appendix D, the WI-initialized particles? Most likely, sampling from the pushed-forward empirical measure would result in a non-symmetric teacher, but there's a chance that the WI teacher might contain two pairs of symmetric parameters (e.g. $\theta_1^*, g\theta_1^*, \theta_2^*, g\theta_2^*$) and another singled out, making it mostly equivariant. After all, if the teacher is not equivariant, then the data is not equivariant, and the assumptions in Theorem 5 do not hold. * Section 4.2 reminds me of another work on discovering symmetry [1], which initially specifies a large candidate group and discovers the subgroup symmetry by inspecting the weights of relaxed group convolution filters. Can you comment on the connection between your method and theirs? ## References [1] Wang, Rui, et al. "Discovering Symmetry Breaking in Physical Systems with Relaxed Group Convolution." Forty-first International Conference on Machine Learning.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have discussed the limitations.

Reviewer 2QUz7/10 · confidence 3/52024-07-12

Summary

This work studies the symmetric structure of the model and data structures with respect to the action of $G$ and studies the Wasserstein gradient flow (WGF) for learning overparameterized neural networks under the mean-field (MF) regime. In particular, the authors consider data augmentation (DA), feature averaging (FA), and equivariant architecture (EA) as symmetry-leveraging techniques. They show that for symmetric data, DA and FA follow the same dynamics by exploiting weakly and strongly invariant laws (WI and SI) on the parameter space. In particular, the following statements are given. - DA and FA are equivalent in terms of optimal values and also equivalent to standard training (free training) for WI measures (Theorem 2). - Furthermore, this equivalence is shown over the space of general measures when the data distribution is equivalent (Corollary 2). - WGF for an invariant functional preserves weak invariance of the measure (Theorem 3 and Corollary 3). Moreover, there is an equivalence of WGF for DA, FA, and free training under suitable conditions.

Strengths

The paper proposes the notion of WI and SI and proves several equivalences between DA, FA, and free training. These results indicate when these techniques (DA and FA) are meaningful.

Weaknesses

- This work mainly focuses on two-layer neural networks; hence, it is unclear if the results can be applied to deep neural networks. - The results, especially for WGF, seem to fully rely on the mean-field limit $N\to\infty$. Hence, it’s unclear whether the theory has some implications for finite-neuron settings. - (Minor) The paper is somewhat hard to follow because of the numerous notations. I suggest including a table that summarizes these notations.

Questions

- In Theorem 4, the noisy SGD (4) involves the projection step of the noise. I think such a projection is non-trivial in general. Are there any practical examples? - Is WGF essential for the results in Section 3.3? Do similar results hold for other dynamics (e.g., underdamped version)?

Rating

7

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

Limitations are addressed in the paper.

Reviewer j4Mu7/10 · confidence 3/52024-07-14

Summary

This paper presents a mean-field analysis on a class of overparameterized neural networks that are expressed as an ensemble of $N$ units, trained with SGD under symmetries in data distribution and possible use of symmetry-leveraging techniques (data augmentation, feature averaging, and equivariant architectures). While the analysis is based on prior work on mean-field analysis on the same class of neural networks e.g. based on Wasserstein gradient flows [1], the contribution of this work is in extending to learning under symmetries. The setup is as follows: - In considered neural networks, a unit computes $\sigma_*:(\mathbf{x},\theta_i)\mapsto\hat{\mathbf{y}}$ and the network computes $\Phi_{\theta}^N:(\mathbf{x},\theta)\mapsto \frac{1}{N}\sum_{i=1}^N\sigma_*(\mathbf{x},\theta_i)$. - Overparameterization means the width $N\to\infty$, yielding a limiting ensemble $\Phi_\mu$ where $\mu$ is understood a probability measure on the space of parameters $\theta_i$. - The parameters $\theta$ are optimized using noisy SGD. The mean-field theory of the neural networks considered above [1] aims to analyze the training dynamics of the limiting ensemble $\Phi_\mu$, given that its optimization using a convex loss function becomes a convex optimization problem (unlike $\Phi_{\theta}^N$ which induces non-convex optimization) and that its global optimum may be approximated by training $\Phi_{\theta}^N$ with a large width $N$. Specifically, it is known that SGD on $\Phi_{\theta}^N$ approximates Wasserstein gradient flow on $\Phi_\mu$ in the scaling limit $N\to\infty$ under certain assumptions, which then, under certain assumptions, converges to the global optima of the convex optimization problem. Given the background, the paper presents a mean-field analysis given that the data generating distribution is possibly under a symmetry described by a compact group $G$, and symmetry-leveraging techniques are possibly used. The authors consider data augmentation, feature averaging (i.e. symmetrization; also called the Reynolds operator), and equivariant architectures, in accordance to literature [2]. The idea for the analysis is to distinguish between two types of symmetries for the measure $\mu$ on parameters: weakly-invariant and strongly-invariant, where weakly-invariant refers to an invariant measure on the subspace containing possibly non-invariant elements, and strongly-invariant refers to a measure on invariant subspace. The authors first show in Section 3.1 that, given that the unit $\sigma_*(\cdot, \cdot)$ is jointly equivariant, $\Phi_\mu$ is equivariant if and only if $\mu$ is weakly-invariant. Then, feature averaging (and data augmentation) can be understood as being related to weakly-invariant distributions obtained through symmetrization, and equivariant architecture is related to strongly-invariant distributions. The authors then, in Section 3.2, begin mean-field analysis on optimality under a set of assumptions including that the unit $\sigma_*(\cdot, \cdot)$ is jointly equivariant (Assumption 1), and show that: - (Proposition 3) Optimizing with respect to an invariant loss function yields a weakly-invariant optimum if it has a unique minimizer, - (Corollary 1) If input data distribution is invariant, optimizing under data augmentation and feature averaging are equivalent and corresponds to approximating the symmetrized version of mean of the labeling function, - (Corollary 2) If data distribution is equivariant, optimizing under data augmentation and feature averaging provide no advantage, - (Proposition 4 and 5) But even when data distribution is equivariant, equivariant architectures (i.e. strongly invariant measures $\mu$) may not achieve the optimum with respect to a given invariant loss function, unless universality properties that are "good enough" for the given data distribution are assumed. Then, in Section 3.3, the authors provide mean-field analysis on training dynamics, and show that: - (Theorem 3, Corollary 3, and Theorem 4) Under an invariant loss function and Assumption 1 (among others), a training dynamics $\mu_{0:t}$ following the Wasserstein gradient flows yields a unique weakly-invariant solution $\mu_t$. If the data distribution is equivariant, from Corollary 2 this result applies to freely-trained neural networks without symmetry leveraging techniques, and data augmentation and feature augmentation again provide no advantages. - (Theorem 5 and 6) Under an invariant loss function and Assumption 1 (among others), a training dynamics $\mu_{0:t}$ starting at a strongly-invariant initial condition $\mu_0$ and following the Wasserstein gradient flows yields strongly-invariant solutions $\mu_t$. Importantly, this means that if the data distribution is equivariant, $\mu_t$ stays strongly-invariant throughout training even though it is not explicitly forced or constrained to be invariant. Furthermore, equivariance of data distribution can be dropped if one of the symmetry leveraging techniques are used, which all yield coinciding solutions. In Section 4, the authors provide synthetic experiments focusing on validating Theorem 3-6, and then propose and demonstrate an empirical method for finding good equivariant architecture for the task at hand (i.e. the most expressive among invariant subspace) by training according to Theorem 5 and 6 until convergence. [1] Lenaic Chizat and Francis Bach. On the global convergence of gradient descent for over-parameterized models using optimal transport. Advances in neural information processing systems, 31, 2018. [2] Clare Lyle, Mark van der Wilk, Marta Kwiatkowska, Yarin Gal, and Benjamin Bloem-Reddy. On the benefits of invariance in neural networks. arXiv preprint arXiv:2005.00178, 2020.

Strengths

- S1. The paper targets an important and challenging problem of understanding training dynamics of overparameterized neural networks under symmetries and their leveraging. The technical approach that applies mean-field analysis of generalized shallow neural networks [1] to learning under symmetries is original as far as I am aware. - S2. The paper considers three types of symmetry-leveraging techniques: data augmentation, feature averaging, and equivariant architectures. These three are sufficiently general to represent currently used techniques [2]. The descriptions as well as strengths and weaknesses of each technique discussed in Section 2.3 are correct. - S3. The authors' approach of imposing distributional symmetries on the measure $\mu$ to describe equivariant architectures as strongly-invariant (and feature averaging as weakly-invariant given that the unit $\sigma_*(\cdot, \cdot)$ is jointly equivariant) is interesting and could be potentially useful in future work. - S4. The results presented in Section 3 and 4 are interesting, in particular Theorems 5 and 6 which shows that overparameterization and equivariant units, combined with equivariant data distribution or symmetry leveraging, leads to the property that strongly-invariant parameter initialization stays strongly-invariant throughout the training even without explicit constraints to do so. (But also see W1.) [1] Lenaic Chizat and Francis Bach. On the global convergence of gradient descent for over-parameterized models using optimal transport. Advances in neural information processing systems, 31, 2018. [2] Clare Lyle, Mark van der Wilk, Marta Kwiatkowska, Yarin Gal, and Benjamin Bloem-Reddy. On the benefits of invariance in neural networks. arXiv preprint arXiv:2005.00178, 2020.

Weaknesses

- W1. I have concerns regarding Assumption 1, which is a key assumption underlying Section 3 and assumes that the unit $\sigma_*:(\mathbf{x},\theta_i)\mapsto\hat{\mathbf{y}}$ is jointly equivariant with respect to the input $\mathbf{x}$, the parameters $\theta_i$, and the output $\hat{\mathbf{y}}$. This assumption states that we have already (partially) baked in the symmetry constraint to our model class. This may be natural for analyzing equivariant architectures, but does not coincide with the practical uses of data augmentation and feature averaging (e.g. please see [2]), as they are generally applied to an arbitrary, non-equivariant function, while in the setup of this work they are applied on already equivariant units. This limits the usefulness of the results regarding data augmentation and feature averaging. Perhaps the setup of this work could be relevant to symmetry breaking in already equivariant architectures [3], but I am not entirely sure. - W2. In Line 92-93, the authors claim that "Note that requiring an infinite and i.i.d. sample from $\pi$ is key when letting later $N\to\infty$. This is not truly a restriction, since one can always sample from the empirical distribution of finite data points". This seems problematic since (1) if we can infinitely sample from $\pi$, then we have access to it, which violates basic assumptions in machine learning (Line 82), and (2) if we infinitely sample from an empirical distribution supported on finite set of data points, the resulting distribution would not be $\pi$, and consequently the SGD scheme in Equation (1) would not be optimizing the objective in Equation (2). (Please correct me if I am wrong.) - W3. The writing of the paper could be in general improved. For example, the paper uses an excessive number of abbreviations (MF, MFL, NN, SL, DA, FA, EA, WI, SI, WGF, RMD, V), which is hurting readability. It would be also helpful to provide an example of a measure and its (non-equivalent) invariant symmetrization and projection in Definition 5; these are important concepts that are used in all later sections, but I was not able to directly grasp them from reading the text. For example, considering a uniform distribution on $[0, 1]\subset\mathbb{R}$, and a multiplicative group {1, -1} acting on $\mathbb{R}$ by multiplication, may be sufficiently informative. - W4. The architecture discovery algorithm proposed in Section 4.2 was a bit questionable to me, since the architecture of the ensemble is already specified (at least up to a significant portion) by the jointly equivariant unit(s) $\sigma_*(\cdot,\cdot)$, and the algorithm searches for their parameters within a strongly-invariant regime. I am not sure why this algorithm can be understood as doing architecture discovery, instead of doing the usual parameter optimization. - W5. Currently, it is not immediate how the study of the ensemble considered in the paper would benefit the development of practical neural networks that leverage symmetry. Showing how some of the currently used symmetry-leveraging deep neural networks can be viewed as such ensembles could be informative in this directions. [2] Clare Lyle, Mark van der Wilk, Marta Kwiatkowska, Yarin Gal, and Benjamin Bloem-Reddy. On the benefits of invariance in neural networks. arXiv preprint arXiv:2005.00178, 2020. [3] Rui Wang, Elyssa Hofgard, Han Gao, Robin Walters, and Tess E. Smidt, Discovering symmetry breaking in physical systems with relaxed group convolution, arXiv preprint arXiv:2310.02299, 2023.

Questions

Please see the Weaknesses section.

Rating

7

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The authors have partially addressed limitations of their work, for example in Section 5. Since the discussions are currently scattered across the main text, I encourage the authors to gather them in a dedicated limitations section.

Reviewer 4cbm2024-08-08

Thank you for your response. My concerns have been addressed, and I'm happy to keep the current recommendation. I have some more comments about Q4&5. For Q4, how many particles are there in the WI-initialized teacher? Originally I thought it was 5, as you mentioned in L1585. But the WI distribution $(\nu_{\theta^*}^{N_*})^G$ has support on 10 (5 pairs of) points. Do you randomly sample 5 particles from this distribution, or do you include all 10 points? If it is the latter case, then the teacher is indeed strictly equivariant. For Q5, I think the difference between "discovering the symmetry of the data" and "optimizing model architectures compatible with data symmetries" is subtle. I feel these two objectives are often connected to each other. In [1], for example, one can tell from the optimized architecture what is the maximal symmetry in data. This is also done in [2] but in a quite different way, by solving the Lie derivative constraint wrt a non-equivariant network. Conversely, one can also start by discovering the data symmetry and design equivariant architectures accordingly [3]. These works may be relevant if you plan to further investigate the idea in Sec 4.2. [1] Discovering Symmetry Breaking in Physical Systems with Relaxed Group Convolution. ICML 2024. [2] LieGG: Studying Learned Lie Group Generators. NeurIPS 2022. [3] Generative Adversarial Symmetry Discovery. ICML 2023.

Authorsrebuttal2024-08-09

Thank you very much for your response, and we are happy as well for your positive recommendation. Regarding your follow-up questions: Regarding L.1585, we agree that the current phrasing can easily lead to confusion, and we will be sure to change it in the final version. We only wanted to refer to the arbitrary teacher having 5 particles and, indeed, as you have noticed, the WI teacher has exactly 10 particles (which are the ones we provide in our original rebuttal). Namely, it is exactly $G$-equivariant (and not just an empirical sample from a $G$-equivariant measure). As we mention in our original rebuttal, the current version “implicitly” mentioned both of these facts, as from our remarks in lines 196-203 and the phrase in line 1587-1588; but we will surely make this clear and explicit in our final version. We do want to stress that the exact equivariance of the WI teacher doesn’t mean the problem is any easier. Since the WI-teacher’s particles don’t live in ${\cal E}^G$, one intuitively would believe that the freely-trained particles of a NN (initialized within ${\cal E}^G$) would prefer to “escape” ${\cal E}^G$ in order to better approximate the teacher. However, as shown by our numerical experiments, this doesn’t happen, and the particles remain (up to numerical error) within ${\cal E}^G$. Regarding Q5, we deeply thank you for the provided references and ideas. We will be sure to consider them for our upcoming works regarding the topic.

Reviewer j4Mu2024-08-08

Thank you for the comprehensive rebuttal. I am trying to technically understand the response on the joint-equivariance of single-hidden-layer NN when the group acts trivially on the hidden layer (regarding W1). Following Line 65-67, let $\mathcal{X}=\mathbb{R}^d$, $\mathcal{X}=\mathbb{R}^c$, and $\mathcal{Z}=\mathbb{R}^{c\times b}\times \mathbb{R}^{d\times b}\times \mathbb{R}^b$, $z = (W, A, B)\in \mathcal{Z}$, $\sigma:\mathbb{R}^b\to\mathbb{R}^b$, and $\sigma_*(x,z)\coloneqq W\sigma(A^\top x+B)$. From Line 170-171, the joint-equivariance is written as $\sigma\_* (\rho\_g \cdot x, M\_g \cdot z) = \hat{\rho}\_g \sigma\_* (x,z)$ for all $g,x,z\in G\times \mathcal{X}\times \mathcal{Z}$. Now let $M\_g$ be a trivial action. Then joint-equivariance is required as $\sigma\_* (\rho\_g \cdot x, z) = \hat{\rho}\_g \sigma\_* (x,z)$, or equivalently $W\sigma(A^\top \rho\_g \cdot x+B) = \hat{\rho}\_gW\sigma(A^\top x+B)$, for all $g,x,z\in G\times \mathcal{X}\times \mathcal{Z}$. As far as I am aware, this is not true in general; let $A=0$, $\sigma=\mathrm{id}$, $W=I$, and let $B\in\mathbb{R}^b$ have distinct entries, then we have $\sigma\_* (\rho\_g \cdot x, z) = B$ which is not equal to $\hat{\rho}\_g \sigma\_* (x,z) = \hat{\rho}\_gB$ for faithful actions $\hat{\rho}_g$. Therefore I do not see how joint-equivariance holds for single-layer NNs if the group acts trivially on the hidden layer. Am I missing something?

Authorsrebuttal2024-08-09

We thank you very much for your response, and we will take the opportunity to provide a clarifying example for the question at hand. As noted in Lines 1012-1016, under the same example you are considering, the “natural” action of $G$ on the parameter space ${\cal Z} = \mathbb{R}^{c \times b} \times \mathbb{R}^{d \times b}\times \mathbb{R}^b$ corresponds to taking, for $g \in G$ and $(z = W, A, B) \in \cal Z$: $$M_g.z = M_g.(W, A, B) := (\hat{\rho}_g.W.\eta_g^T, \rho_g.A.\eta_g^T, \eta_g.B). $$ This is what we refer to as the “intertwining action” on the layer parameters, as it recalls the traditional definition of an intertwinning linear map from representation theory (also, see the discussion on lines 165-180). Notice that, indeed, this involves the action of the group on the corresponding input/output spaces of each layer (e.g. from $\mathbb{R}^d$ to $\mathbb{R}^b$ and from $\mathbb{R}^b$ to $\mathbb{R}^c$). In our original rebuttal, notice that we referred to “a trivial action on the intermediate layer”, and not plainly “a trivial action”. Indeed, we referred to taking $\eta$ to be trivial, rather than $M$ itself; the latter, as you have correctly noticed, doesn’t work in general. Thanks to the orthogonality (and linearity) of our group action, a similar development to yours yields, for any $\sigma: \mathbb{R}^b \to \mathbb{R}^b$, any $g \in G$ and for any $z = (W, A, B) \in \cal Z$: $$\sigma_*(\rho_g.x, M_g.z) = \sigma_*(\rho_g.x, (\hat{\rho}_g.W.\eta_g^T, \rho_g.A.\eta_g^T, \eta_g.B)) =(\hat{\rho}_g.W.\eta_g^T). \sigma((\rho_g.A.\eta_g^T)^T.(\rho_g.x) + \eta_g.B)$$ $$ = \hat{\rho}_g.(W.\eta_g^T. \sigma(\eta_g. A^T.(\rho_g^T.\rho_g).x + \eta_g.B)) = \hat{\rho}_g.(W.\eta_g^T. \sigma(\eta_g. (A^T.x + B)))$$ Namely, it is enough that $\sigma: \mathbb{R}^b \to \mathbb{R}^b$ is $G$-equivariant (w.r.t. the action of $\eta$) in order to get $\sigma_*$ to be jointly equivariant. Indeed, this would yield: $$\sigma_*(\rho_g.x, M_g.z) = \hat{\rho}_g.(W.\eta_g^T.\eta_g. \sigma(A^T.x + B)) = \hat{\rho}_g.(W.\sigma(A^T.x + B)), $$ as required. This justifies our claims about the pointwise activations and/or the use of a Norm-ReLU activation to ensure the joint-equivariance of $\sigma_*$ for different levels of complexity of $G$ (and its action on the intermediate layer). More specifically, if this action on the intermediate layer $\eta$ is taken to be trivial, any function $\sigma: \mathbb{R}^b \to \mathbb{R}^b$ is $G$-equivariant w.r.t. $\eta$ (indeed, $\sigma(\eta_g.x) = \sigma(x) = \eta_g.\sigma(x)$); i.e. no matter the inner activation function $\sigma$, under this action $M$, $\sigma_*$ is jointly-equivariant. In the end, a trivial $\eta$ assumes “no action of the group on the latent space”; and so we only act on the parameters on the dimensions defined by the data: e.g. $\rho$, that acts on $\cal X$, will act only on $A$ on the left; and $\hat{\rho}$, that acts on $\cal Y$, will act only on $W$ on the left as well. This readily generalizes to the case where $\sigma_*$ represents a whole multilayer unit, for which we take the action on parameters to be trivial on the hidden layers, but not on the input and output layers (where it acts according to the data). Indeed, no matter the inner activations, such $\sigma_*$ would be jointly equivariant with respect to such group action. As we also mention in our original rebuttal, a “trade-off” follows from considering such a “boring” group action $M$: the space ${\cal E}^G$ might not represent very “interesting” equivariant architectures. This doesn't seem too problematic since it wouldn’t be the use case if $\sigma$ didn’t respect any sort of equivariance to begin with. It does, however, allow to state that DA and FA are tightly related, even for unconstrained architectures. This also makes sense when we consider that DA and FA only interact with NN models through the input and output layers. We hope to have clarified your question with this example, and we’re open to answer further questions if required.

Reviewer j4Mu2024-08-10

Thank you for the thorough explanation. Together with the original response, my major concerns have been addressed. I will adjust my score accordingly.

Authorsrebuttal2024-08-12

Thank you

We thank the reviewer very much for their positive comment and for adjustment of their score.

Reviewer aSNH2024-08-09

Thank you

I thank the authors for their thoughtful response and thorough work. I am satisfied with the changes promised in the global response and I will keep my positive evaluation.

Authorsrebuttal2024-08-09

We thank the reviewer for their comment, and are very happy for their positive evaluation. Thanks again!

Reviewer 2QUz2024-08-12

Thank you for the detailed response. The authors have adequately addressed my concerns. Given the strength of the paper, some limitations seem negligible. I will raise my evaluation. The discussion provided in the general comments will significantly help readers understand the importance and applicability of the theory.

Authorsrebuttal2024-08-12

Thank you

We thank you for pointing out your concerns, as they have undoubtedly helped us provide a clearer exposition of our central ideas and contributions in the final version of our paper. We are also thankful for your positive comments and the evaluation adjustment.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC