acceptodds
Under review as a conference paper at ICLR 2027

Permutation Equivariant Neural Networks for Antisymmetric Tensors

Abstract

Antisymmetric tensors change sign under index swaps, yet learning methods that exploit this structure remain largely unexplored. We characterise all linear permutation equivariant maps between antisymmetric power spaces of , showing that they are determined by a class of objects that we term antispherical bipartitions. As a consequence, we show that the dimension of the space of such maps is at most two, independently of the tensor orders and of . We further introduce IO patterns, which replace large weight matrices with a representation whose storage is independent of . This framework is well-suited to graph-structured data with signed or directed edges, where antisymmetry encodes the orientation of edges. We illustrate our approach on the Wiki-RfA dataset, showing that the characterised layers perform at least as well as general permutation equivariant maps at a lower parameter cost, and that IO patterns remain practical at sizes where the dense weight matrix exhausts available memory.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
equivariance, deep learning architectures, symmetric group, antisymmetric tensors, graphs

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_dQ3EOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies linear permutation equivariant maps between antisymmetric tensors of arbitrary order. It gives a diagrammatic construction of a complete basis for these maps and proves that the dimension of this map space is at most two. In addition, it introduces an IO pattern, so that applying these maps does not require storing large weight matrices. Experiments on the Wiki-RfA dataset show that the method can achieve performance comparable or better than existing general permutation equivariant models while using fewer parameters. The paper is well-written and clearly organized.

However, I think the novelty of this work is relatively weak compared with existing mathematical results. The dimension formula and the conclusion that the dimension is at most 2 can both be derived directly from classical representation theory results [1]. In addition, reference [2] has already given a closely related diagrammatic basis for the case where the input and output have the same tensor order. The authors should clearly compare the main results of this paper with these existing results, and clarify which parts of the proposed characterization are new contributions.

[1] William Fulton , Joe Harris, Representation Theory: A First Course, 1991

[2] Alexander Wilson. The Multiset Partition Algebra: Diagram-Like Bases and Representations, 2023.

Critical strengths

The authors provide a complete basis for equivariant linear maps with arbitrary input and output orders, and prove that the dimension of the corresponding mapping space is at most 2. This result lays a clear mathematical foundation for building equivariant networks that exploit antisymmetric structures.

Critical weaknesses

I think the main weakness of this work is the lack of clear motivation. Although this paper gives a theoretical characterization of permutation equivariant linear maps between antisymmetric tensors of arbitrary order, the practical motivation for considering higher order tensors is still unclear. The experimental section focuses on a standard directed graph dataset, where pairwise relations can already be represented by second order antisymmetric tensors (i.e., adjacency matrices). The architecture in the experiments only involve 1-order and 2-order tensors. So it is not clear what additional advantages this higher order antisymmetric representation actually offer in this setting. The authors should clarify what meaningful machine learning tasks actually benefit from such higher order representations, a theoretical justification of their benefits would also be helpful.

Questions for authors

Can the authors identify representative machine-learning tasks in which higher-order antisymmetric tensor representations offer advantages that are difficult to achieve with lower-order representations? Could the practical value of the proposed framework be further demonstrated on a task that genuinely requires higher-order antisymmetric tensors?

3 comments

We thank the reviewer for their review, and we are glad that they found the paper to be well written and that our results provide “a clear mathematical foundation for building equivariant networks that exploit antisymmetric structures.” We answer their points below.

Wilson, 2023. Wilson’s work concerns symmetric rather than antisymmetric tensors, so we will cite it in our discussion of symmetric tensors in the Related Work section. The multiset partition algebra that was introduced in their paper is the algebra of SnS_n-equivariant maps on a space of degree-rr polynomials in commuting variables, which decomposes into tensor products of symmetric powers of Rn\mathbb{R}^n, so its diagram-like basis describes equivariant maps on symmetric tensors. Our work considers equivariant maps on antisymmetric tensors, and our diagrams, which come from antispherical bipartitions and are constructed differently from Wilson’s, must account for the signs that arise from reordering indices. This is the key technical difficulty in our setting: we resolve it with our sign rules (Figure 2), which also explain why certain diagrams vanish (Lemma 4.13). We also characterise maps between different orders k≠lk \neq l for all nn, whereas Wilson only considers maps from a space to itself and needs n≥2rn \geq 2r.

Fulton & Harris. We agree that the dimension in Corollary 4.16 can also be recovered classically: writing Rn≅1⊕V\mathbb{R}^n \cong \mathbf{1} \oplus V, where V={x:∑ixi=0}V = \lbrace x : \sum_i x_i = 0 \rbrace is the standard representation, gives Λk(Rn)≅ΛkV⊕Λk−1V\Lambda^k(\mathbb{R}^n) \cong \Lambda^k V \oplus \Lambda^{k-1} V (cf. Eq. (85)). The ΛjV\Lambda^j V are irreducible (Fulton & Harris, Prop. 3.12) and are pairwise non-isomorphic, since they correspond to distinct hook partitions (n−j,1j)(n-j, 1^j), and Schur’s lemma then gives the dimension count. We will add this remark and the citation.

The dimension, however, is a by-product rather than our main contribution. Schur’s lemma gives the number of maps but not the maps themselves. By contrast, our construction gives an explicit basis of weight matrices in the coordinates in which antisymmetric tensors are stored, uniformly in nn, and directly leads to the IO patterns that make these layers practical at scale (see Appendix J). This mirrors the characterisation of permutation equivariant layers on general tensors by Maron et al. (2019), whose Bell-number dimension likewise follows from the classical theory of the partition algebra (Jones, 1994): in both cases the value for equivariant learning lies in turning the abstract count into layers that can be implemented.

Novelty. In summary, our new contributions are 1) the explicit basis of weight matrices for every pair of orders (k,l)(k, l) and every nn, given by antispherical bipartitions and their sign rules; 2) the IO patterns that implement these maps with O(1)O(1) storage of each map in nn; and 3) the first explicit characterisation of these layers for equivariant learning.

What the experiment tests. Our experiment is not designed to compare higher-order with lower-order representations. It compares our layers with general permutation equivariant layers (Maron et al., 2019) at the same tensor orders, with everything else held fixed. Even at orders one and two these layers had not been characterised before, and in our updated results (see our response to Reviewer drUC) they achieve a higher AUC than the general layers at every width, with fewer parameters.

Where higher orders help. Higher orders help when what we want to model involves three or more entities together. For example, consider pairwise comparison data, such as match results between teams or human preferences between language models. Let TijT_{ij} be the log-odds that ii beats jj. The 2→32 \to 3 map in Table 2 then gives Tij+Tjk+TkiT_{ij} + T_{jk} + T_{ki} for each triple. If the three comparisons fit together, then this is zero, but if they go round in a circle (i.e. ii beats jj, jj beats kk and kk beats ii), then it is large. Such cycles matter for both tasks, because no single ranking can explain them. Whether there is such a cycle is a question about the triple, and is not about any one pair, so an order-three tensor is the natural place to store the answer. Order-two layers, which can only combine values along rows and columns, cannot compute it directly.

A higher-order experiment. Such a task would be an interesting application in its own right, but it would not change the main contributions of the paper. The characterisation, the explicit basis, and the IO patterns are established for every order, so they do not depend on an experiment at any particular order. The role of our experiment is to show that layers that are built from the characterisation work well in practice, and our updated results show exactly that.

Thank you for the detailed response. The additional explanations help clarify the mathematical contributions and their relationship to existing work. However, my main concern regarding the motivation and practical significance of higher-order antisymmetric tensor representations remains insufficiently addressed.

Overall, the response does not substantially change my assessment of the paper. I therefore maintain my original score.

We thank the reviewer for their reply. We are glad that our response clarified our mathematical contributions and their relationship to existing work. We are grateful for their time and feedback.

Reviewer_drUCOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper characterizes permutation equivariant linear maps between totally antisymmetric tensors of arbitrary order. They prove the surprising result that these maps have dimension 0, 1, or 2, unlike the general case where the dimension follows the Bell numbers (Maron et al 2019). The authors then develop an implementation of these linear maps that does not require instantiating the full n^k x n^l matrix which allows them to build efficient models. Finally, they apply their model on a small example, comparing it against a few baseline models.

Critical strengths

S1: The paper is well motivated, addressing a problem that hasn't seen much research.

S2: The theoretical results are well-developed, comprehensive, and convincing. The authors give a complete characterization of the permutation-equivariant linear maps between tensors of any order, covering all the cases one might care about. The dimension result is surprising, another signal of its relevance to researchers in this area.

S3: The efficient implementation is another nice result, enabling this method's adoption more readily by practitioners.

S4: The authors explain the method thoroughly and include plenty of examples to help the reader wade through all the notation and indices required.

Critical weaknesses

W1: The numerical experiments are weak, even with the caveat that they are not claiming state-of-the-art performance. The most problematic part is that the authors method performs worse than a simple MLP. I would find this paper more convincing if:

  1. the authors method with comparable number of parameters performs better than an MLP, or

  2. the authors method performs comparably with an MLP, but has far fewer parameters, or obviously the best case

  3. the authors method does better and has fewer parameters

So to answer this question, I really want to know how your method performs when you give it more channels/parameters. Doing worse than the Signed Message Passing GNN is ultimately not a problem, although I am curious if that is also due to the huge disparity in number of parameters. Also, it is possible that I am misunderstanding the size comparison of the two methods, i.e. they are of comparable size since your method only requires <= 2 parameters per linear map which may be a very large n^k x n^l linear map. If ultimately after trying many things, the MLP still does better than your method implementing the symmetries, it makes me wonder if the dataset/problem is hard enough or doesn't rely on the symmetries enough, in which case another dataset would be better.

W2: After stating Lemma 4.13 and Theorem 4.14 you should link to where you prove them in the appendices so that the reader can find those easily.

Questions for authors

Q1: Did you explore nonlinearities other than tanh? Requiring an odd nonlinearity is a necessary, but unfortunately significant restriction. I have not had success with tanh in the past, which makes me wonder if you tried any others.

Q2: How does the performance change when you increase the number of channels? 1 and 4 channels seems quite small.

Q3: In physics there are odd-parity tensors, such as angular velocity, vorticity, etc., which undergo a sign flip under the reflections of O(d). Is there a connection to your antisymmetric tensors? See for example https://openreview.net/pdf?id=1FCZ4f8dAY.

Q4: It took me a little to understand that the sentence on line 430 is saying that your method has complexity O(n^{l+1]}), while the naive implementation has O(n^{k+l}). Consider rewording this to make it clearer.

3 comments

We thank the reviewer for their detailed and positive review, and we are glad that they found both the theory to be comprehensive and the dimension result to be surprising.

W1 and Q2 (experiments and channels). Having reflected on the reviewer’s question, we found an issue in our original setup, which resulted in every invariant model (including the signed GNN) being unfairly penalised relative to the MLP.

For every subgraph that was constructed, the candidate was placed at node 0. This meant that the MLP could identify the candidate by position, whereas the invariant models, which pooled over all nodes, could not tell which of the ten users was the candidate, and hence whose votes mattered most. In effect, in our original setup, we incorrectly made the invariant models invariant to all of S10S_{10}, whereas the task is actually only invariant to S9S_9 (i.e. any relabelling of the non-candidate nodes is allowed).

We have now corrected this for every invariant model by also passing the output at the candidate node to the readout. The layers themselves are unchanged, and each model is now invariant to exactly the permutations that fix the candidate. This means that the results below now give a fair comparison between our models and the baselines. We reran all models using the same training and evaluation details that were given in the paper (Appendix L). We have also added the channel sweep for Q2. The results are given as mean ± std over three splits (the MLP scores are unaffected):

ChannelsOurs: params / acc / AUCGeneral: params / acc / AUC
1138 / 0.786 ± 0.014 / 0.879 ± 0.012163 / 0.787 ± 0.011 / 0.860 ± 0.018
481 / 0.786 ± 0.008 / 0.881 ± 0.011181 / 0.785 ± 0.008 / 0.856 ± 0.019
8193 / 0.783 ± 0.006 / 0.878 ± 0.008393 / 0.781 ± 0.010 / 0.854 ± 0.008
16513 / 0.781 ± 0.006 / 0.878 ± 0.007913 / 0.778 ± 0.010 / 0.857 ± 0.003
321,537 / 0.777 ± 0.013 / 0.872 ± 0.0142,337 / 0.778 ± 0.010 / 0.850 ± 0.008

MLP: 10,689 / 0.786 ± 0.012 / 0.861 ± 0.014.

Signed GNN: 9,729 / 0.797 ± 0.012 / 0.893 ± 0.011.

We see that our models now match the MLP in accuracy and exceed it in AUC with up to 130 times fewer parameters. This corresponds to the reviewer's point 2 in W1 and, in AUC, to the reviewer’s point 3 in W1. Our models also achieve a higher AUC than the general permutation equivariant layers across each of the channels.

In answer to Q2, the results show that the performance barely changes with the number of channels. Across the channels, our models have AUC between 0.872 and 0.881 and accuracy between 0.777 and 0.786, and the general permutation equivariant layers behave in a similar way. Looking at our logs, the models with more channels achieved their best validation loss sooner (about 1,200 epochs with 1 channel versus 270 epochs with 32 channels). This suggests that the extra capacity mostly leads to earlier overfitting.

Finally, the signed GNN stays slightly ahead (0.893 against 0.881 in AUC) with about 120 times more parameters than our four-channel model. Since making our models larger did not improve them, we doubt that extra parameters explain this small lead.

We will replace Table 1 with these results and describe the correction in Section 6.

W2 (links to proofs). We will add references to the proofs in Appendix B directly after Lemma 4.13 and Theorem 4.14.

Q1 (nonlinearities). We did not explore other nonlinearities, and only used tanh throughout our equivariant layers. We agree that requiring odd nonlinearities is a genuine restriction, but it only applies for tensor orders that are above one: for orders zero and one the antisymmetry does not impose any condition hence any nonlinearity can be used there. Comparing tanh with other odd choices is something that we will explore in future work. We will clarify this in Remark 4.19.

Q3 (odd-parity tensors). We thank the reviewer for pointing us to this work. The answer is yes, there is a direct connection. Permutation matrices are orthogonal, so Sn⊆O(n)S_n \subseteq O(n), and every transposition is a reflection. Our action of SnS_n on Λk(Rn)\Lambda^k(\mathbb{R}^n) is therefore the restriction of the standard O(n)O(n) action, so the sign that our diagrams track is exactly the odd-parity behaviour of Λk\Lambda^k under the reflections that are contained in SnS_n. In three dimensions, Λ2(R3)\Lambda^2(\mathbb{R}^3) is the space of pseudovectors, which is why angular velocity and vorticity can be written as antisymmetric matrices. The groups differ, with O(d)O(d) acting on space versus SnS_n acting on labels, but the sign arises in the same way. We think that combining the two is a natural direction for future work. We will add this to the paper.

Q4 (line 430). We will rephrase this as: “Applying a layer through its IO patterns costs O(nl+1)O(n^{l+1}) time, compared with O(nk+l)O(n^{k+l}) when using a dense weight matrix.”

Reviewer_drUCOpenReview ↗

Response to the Rebuttal

I thank the authors for addressing my questions and running additional experiments. I am glad they are able to identify the issue, and it led to improved results for the invariant models. The original setup compared to the fixed setup is also an interesting ablation that shows when it makes sense to use an invariant model and when it does not. The authors could consider adding it to the Appendix.

Although the authors model now performs how I would expect it to, I believe the experiments are still the weakest part of the paper. It is true that your model does better under the AUC metric, but being tied with the MLP on accuracy still suggests to me that this problem is not hard enough. I think it would improve the paper to have additional experiments on a harder problem to really show the strength of your method. With that said, I believe the major contributions of this paper are in the theory, and those results are excellent. Great theory and now decent experiments means I will raise my score from a 3 to a 4.

We thank the reviewer for their engagement throughout the discussion, and for raising their score. As suggested, we will add the comparison between the original and corrected setups to the Appendix as an ablation, since it shows the cost of imposing more symmetry than the task actually has. We agree that a harder task would further strengthen the experiments, and we see this as an important direction for future work. We are also grateful for the reviewer’s recognition of our theoretical contributions.

Reviewer_GmiKOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper explores the design of layers which are equivariant to the permutation action on antisymmetric tensors. An antisymmetric tensor is a tensor Ti1…ikT_{i_1… i_k} such that permuting components with a permutation σ∈Sk\sigma \in S_k, Tσ(i1)…σ(ik)T_{\sigma(i_1) \dots \sigma(i_k)}, results in the same tensor up to a sign sgn(σ)∈{−1,1}\text{sgn}(\sigma) \in \{-1,1\}. That is Tσ(i1)…σ(ik)=sgnTi1…ikT_{\sigma(i_1) \dots \sigma(i_k)} = \text{sgn}T_{i_1… i_k}. These types of tensors appear in a number of settings including physics, geometry, and graph applications. The paper’s goal is to characterize linear layers that are equivariant to this action. Rather than doing this directly with matrix constraints, the paper converts the problem into one involving diagrammatic algebra. Here, basis elements in the space of all homomorphisms between ∧k(Rn)\wedge^k(\mathbb{R}^n) and ∧l(Rn)\wedge^l(\mathbb{R}^n) are represented as patterns of lines and bubbles. Specified operations and equivalences can be performed on these diagrams. The paper then uses this diagrammatic calculus to answer some basic questions about the space of linear equivariant maps. The central result is a classification of the dimension of the space of all such maps (their dimension is at most 2) as well as an approach to constructing them. The paper ends by describing an efficient implementation for computing with these matrices without explicitly manifesting them and by describing an experiment which applies very small networks to a dataset called Wiki-RfA dataset.

Critical strengths
  • The diagrammatic approach to describing weight spaces is relatively underexplored but promising: Diagrammatic approaches to representing and computing algebraic structure are now well-established in representation theory. Using diagrammatics often simplifies computation and allows a researcher to take advantage of human’s sophisticated visual intuition. There are also a few works now within machine learning that use diagrammatic approaches to understand the design of equivariant layers. The reviewer found this to be a nice example of using diagrammatics to simplify the solution to a question about equivariant architectures.
  • Clarity: The paper is easy to follow despite having to traverse a circuitous route; moving from diagrams to spherical bipartitions to linear equivariant maps. This is not easy to do in 9 pages. The reviewer believes this work will be accessible to most ICLR participants despite involving serious representation theory. Further, the approach presented here is one that might be helpful in a range of deep learning problems that relate to symmetry.
Critical weaknesses
  • The need for networks that are equivariant on antisymmetric tensors is not convincingly conveyed: The reviewer left the paper unsure of the value of the results. Is the impact of this work primarily another item in the classification of linear layers equivariant to a specific symmetry? Is there a need beyond this? Can we characterize what kind of problem would benefit from these results? Answering these questions will help us assess the value of this work.
  • The extent to which these models can actually be realized is unclear: The paper presents a memory-efficient implementation of these layers and calculates complexity. Are there other obstacles to actually using these architectures? Are we limited by the growth in dimension that occurs as kk and nn grow? The reviewer also would have appreciated some commentary on what the central result about dimension of the Hom-space says about architecture design? The paper notes that having Hom-space dimension capped at 2 is very different than the case of generic tensors where growth follows the Bell numbers. What does this say about these types of problems from an architecture design perspective? Will one necessarily have far less expressive architectures?

Nitpicks:

  • It might be helpful to say what the Wiki-RfA dataset is in the introduction, as well as including a citation.
Questions for authors
  • The experiments were very small (the biggest model had 147 parameters). Why was this? Is there an issue to scaling up to small but not tiny size?
  • Can the relationship between Wiki-RfA and antisymmetric tensors be spelled out explicitly? To the reviewer’s memory, this wasn’t described directly.
  • How does the central result of this paper change the way we should think about this family of equivariant architectures relative to analogous settings where Hom-space dimension grows?
3 comments

We thank the reviewer for their positive review. We are glad that they found the paper easy to follow and the diagrammatic approach promising for other problems.

W1 (impact and need). We see the classification as a means rather than an end. Until now, a practitioner working with antisymmetric data could either use general permutation equivariant layers (Maron et al., 2019), which do not preserve antisymmetry and spend parameters on maps that are irrelevant to such data, or discard the orientation, for example by symmetrising a directed graph. We determine the layers that respect both symmetries, and we provide IO patterns that make them practical at scale (Appendix J). Our updated experiments (see our response to drUC) show that these layers work in practice. They achieve a higher AUC than the general layers at every channel width, and match the accuracy of an MLP with up to 130 times fewer parameters. We discuss under W2 how the dimension result changes how these networks should be designed.

W1 (which problems benefit). Our networks are useful for any problem where the input is an antisymmetric tensor over interchangeable entities, such as users, teams or language models, and where relabelling those entities should simply relabel the predictions. Our layers build in both properties, so neither has to be learned from data. In our experiment on the Wiki-RfA dataset, the input is the net endorsements between Wikipedia users. Another example is pairwise comparison data, where the input is the log-odds that one competitor beats another. Higher orders are useful when the structure involves several entities at once. For instance, the 2→32 \to 3 map in Table 2 lets a model of pairwise comparisons represent whether they are consistent around a cycle i→j→k→ii \to j \to k \to i.

W2 (obstacles and growth in kk, ll). The only extra constraint is that pointwise nonlinearities must be odd for tensor orders above one (Remark 4.19). The number of parameters in a layer does not grow with kk, ll or nn, and with our IO patterns, a layer needs only constant storage in nn and O(nl+1)O(n^{l+1}) compute. The remaining cost is the input and output tensors themselves, which any method must store.

W2 (implications for architecture design). We get two concrete design rules from our results. 1) A layer from order kk to order ll is non-zero only when ∣k−l∣≤1|k - l| \le 1, so an architecture moves through tensor orders one step at a time: for example, to reach order three from order one, it must pass through order two. 2) Because there are at most two basis maps for any pair of tensor orders, any extra capacity must come from width and depth rather than from richer layers. We see this as a strength, since general equivariant layers spend many of their B(k+l)B(k+l) parameters on maps that either vanish on antisymmetric data or break antisymmetry.

W2 (expressivity). The architectures need not be less expressive. At the level of linear layers nothing is lost, because our maps are all of the equivariant linear maps between antisymmetric tensors. A general equivariant layer that is applied to an antisymmetric input and then projected back onto antisymmetric tensors will always be a combination of our basis maps. A universality result for full networks, in the spirit of Deep Sets, is interesting future work. We will add this discussion after Corollary 4.16.

Q1 (model size). The model size reflects the design of the experiment. We wanted the most direct comparison possible, so we used a single channel and changed only the family of layers. Each of our layers has at most two parameters per channel, so with a single channel the six-layer backbone has nine, and almost all of the remaining parameters are in the common readout (Appendix L). There is no obstacle to scaling up. In our updated experiments, models with up to 32 channels (1,537 parameters) trained without difficulty. IO patterns also remain practical up to n=7500n = 7500 on a single 10 GB GPU (Appendix J).

Q2 (Wiki-RfA). Each vote is a directed, signed edge: Aij=+1A_{ij} = +1 if user ii supported user jj's request for adminship and −1-1 if they opposed it. We set T=A−A⊤T = A - A^\top, so TijT_{ij} is the net endorsement from ii to jj and TT is antisymmetric by design. Relabelling the users permutes the rows and columns of TT together, which is exactly the action of SnS_n on Λ2(Rn)\Lambda^2(\mathbb{R}^n). We will state this in Section 6 and, as suggested, describe and cite the dataset in the introduction.

Q3 (relation to growing Hom-spaces). Where the Hom-space grows, each layer has many basis maps, so higher orders quickly become expensive in parameters, which in practice limits the orders that can be used. Here there are at most two maps for any pair of orders, so higher-order layers cost no more parameters than low-order ones. The cost of higher orders lies only in the size of the tensors, and flexibility comes from the design choices that were described under W2.

Reviewer_GmiKOpenReview ↗

Reviewer response

We thank the author for their responses. Regarding W1, the author says that "Our networks are useful for any problem where the input is an antisymmetric tensor over interchangeable entities, such as users, teams or language models, and where relabelling those entities should simply relabel the predictions." Is there a more specific, concrete application area? Maybe I am not understanding this correctly, but it feels like the response is, "This type of network will be valuable anywhere where a network with these properties would be valuable".

We thank the reviewer for the follow-up question. To make the pairwise comparison example from our earlier response concrete, we take sport as an example. Given the results so far in a season between nn teams, suppose that the task is to predict the result between every pair of teams. Let TijT_{ij} be the total goal difference of team ii against team jj, so that Tji=−TijT_{ji} = -T_{ij} and TT is an antisymmetric matrix. Our network takes TT as input and outputs an antisymmetric matrix of predicted results, one for every pair of teams. By construction, the predicted result of jj against ii is exactly the negative of that of ii against jj, and relabelling the teams relabels the predictions in the same way. Neither an MLP nor a general permutation equivariant network guarantees the first property, and an MLP does not guarantee the second. We will add this example to the introduction. We hope that this answers the reviewer’s question, and we would be happy to clarify anything further.

Comments on the paper

1 comment

Dear reviewers, the authors responded to your reviews. Could you please go over the responses, see if you have any further questions, and if appropriate, reassess your evaluation of the paper? Note that the deadline for all reviewer-author discussions is November 18th. Thanks, AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.