acceptodds
Under review as a conference paper at ICLR 2027

Homogeneous Autoencoders: Magnitude-Preserving Representation Learning

Abstract

Standard nonlinear dimension-reduction methods do not in general guarantee preservation of observation magnitude, limiting their use in magnitude-sensitive tasks such as extreme-value analysis. We introduce the Homogeneous Autoencoder (HAE), whose encoder preserves a prescribed magnitude exactly while learning a map from ambient to latent directions. The encoder is positively homogeneous, so rescaling an observation rescales its latent point by the same factor. The decoder preserves the same magnitude while accommodating curved geometry. We identify conditions for homogeneous embeddings and provide a sufficient latent dimension bound for existence. For regularly varying (heavy-tailed) data, we characterise how encoding and decoding transform the limiting distribution of extremes and show that the radial tail index is unchanged. Common threshold-based models of extremes use approximate independence of magnitude and direction to extrapolate from moderate to larger events. HAE preserves this independence whenever it holds in the data, and we prove that, among magnitude-preserving encoders, homogeneity above the threshold is necessary and sufficient to preserve it for every input distribution. Experiments on a synthetic heavy-tailed curved surface and ERA5 precipitation fields show that HAE preserves latent magnitude exactly at reconstruction accuracy comparable to parameter-matched autoencoders, with the resulting latent representation supporting more accurate generation of held-out exceedances than a standard autoencoder.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
representation learning, autoencoders, scale equivariance, extreme value theory, regular variation, manifold learning

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_4zpVOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes the Homogeneous Autoencoder (HAE), a dimension-reduction architecture whose encoder has the form f(x) = r(x)·e(θ(x)). Here r is a chosen degree-one homogeneous radial functional and e is a learnt map between unit shells. As a result, the encoder is positively homogeneous and preserves the radius exactly. The decoder also preserves the radius exactly, and adds a radius-dependent angular correction that vanishes uniformly at large radii. This lets it represent non-conic bulk geometry while remaining asymptotically homogeneous.

Critical strengths
  1. Clear, well-motivated problem. Standard autoencoders give no guarantees on latent magnitude, and this is a real obstacle to doing extreme-value analysis in a learnt latent space. The paper states this cleanly, and Prop. 5 makes the non-identifiability point precise.
  2. Simple and principled construction. Factorising the encoder through the unit shell is elegant. The guarantees are architectural, so they hold independently of the loss and of optimisation. The vanishing decoder correction, built via the (1+r)⁻¹ reparameterisation with the boundary value subtracted, is a sensible way to reconcile curved bulk geometry with conic tail geometry.
  3. Prop. 11 is the most interesting result. It explains why exact radius preservation alone is insufficient. The AE + radius ablation isolates this effect empirically, and Table 5, with radii permuted to enforce independence, is a clean and convincing confirmation: HAE is the only neural model that never rejects independence.
  4. Careful, honest writing. The limitations section, the discussion of heterogeneous variables (App. I.4), and the explicit statement that the guarantees "need not reduce reconstruction error" are appreciated. Experimental details are thorough, and code is provided.
Critical weaknesses
  1. Downstream benefit on real data is not demonstrated. The paper's main motivation is extreme-value analysis in latent space, but the generation advantage appears only on the synthetic surface. On ERA5 precipitation:

    • In Table 4, AE + penalty attains lower generation error than HAE under both the fitted law and the empirical law.
    • In Table 5, PCA outperforms all neural models, including HAE, under both laws.

    The authors attribute this to the vMF fit. That may be correct, but it also means the real-data experiments do not show that HAE's better-preserved independence translates into better downstream analysis.

  2. The synthetic benchmark is favourable to HAE by design. The surface is constructed so that, above the 0.85 quantile, direction is nearly independent of radius and the tail is close to conic, which are exactly HAE's assumptions. A setting in which these assumptions hold only approximately, or in which the tail support is genuinely curved, would be more informative.

  3. Many reported metrics are tautological for HAE. Radial error, rank correlation and Hill difference are exactly 0 or 1 by construction. Moreover, AE + penalty already achieves rank correlation 1.000 and small Hill differences everywhere. On radial metrics alone, a soft penalty seems to suffice in practice, so the case for HAE rests mainly on independence preservation and equivariance. This should be stated more directly.

Questions for authors
  1. How does HAE compare with a "θ-autoencoder + raw radius" baseline, i.e. an AE trained on θ(x) whose output is rescaled by r(x)?
  2. On ERA5, is there any setting in which HAE's lower distance correlation leads to measurably better generation than AE + penalty?
1 comment
AuthorsOpenReview ↗

Response to Review by Reviewer 4zpV

We thank the reviewer for a clear and constructive review. The reviewer reads the presentation of the model, experiments and limitations as we intended.

Regarding the downstream benefit on real data, the experimental results are intended to show that HAE provides guarantees on the latent space yet is able to match the practical performance of competitors, retaining its use for representation learning tasks whilst providing additional desirable structure, not to support a claim of uniform superior reconstructive or generative performance.

As well as clarifying the work's intended contribution, we hope that the general comment and revision further address the reviewer's main concern (and question 2) regarding performance on real data by providing a wind field example. In this experiment, HAE's preserved independence does indeed align with the lowest generation error for a factorised tail generator fit in latent spaces of the competing dimension-reducing models (Table 6). To provide further interpretation of Table 5, Appendix J identifies the fitted angular law as the key limitation on the neural models' generation and Section 7 recognises that a linear approach may be sufficient where the support is near a linear subspace.

The synthetic benchmark being compatible with HAE by construction is one reason why experiments are also conducted on real world data. As well as this, the synthetic benchmark satisfies the radial-angular independence assumption of the downstream generator only approximately. Its tail is correctly identified by the reviewer as close to, but not exactly, conic at finite scales. For generation, the precipitation exceedances provide the contrasting setting in which the factorised latent generative model's assumptions are not met and the generation advantage enabled by HAE is expectedly not observed. The wind exceedances experiment added to the revision provides a real data scenario in which independence holds approximately, as the reviewer suggests (Appendix K).

We agree that the radial metrics are exact for HAE by construction, as the daggers in Tables 1 and 3 indicate. These metrics are intended as diagnostics to show how HAEs architectural constraints contrast to the properties of competing models and not independent measures of learned performance. The experiments serve to illustrate that the competing methods fail to provide the same guarantees as HAE, whilst demonstrating that HAE can provide these properties while simultaneously achieving similar reconstructive performance and thus presents a minimal tradeoff.

If our interpretation of the reviewer's suggested θ\theta-autoencoder + radius model is correct, its angular encoder and decoder act only on θ(x)\theta(x) with the decoded direction rescaled by r(x)r(x). Therefore the reconstruction map is positively homogeneous and it has a conic image. Under the same unit-shell latent parameterisation, it is architecturally equivalent to HAE with the decoder correction turned off and is thus less expressive than HAE. At the minimal latent dimension, this restricts the decoder's ability to represent manifolds with radius-dependent angular geometry (as on the curved surface example). Avoiding this restriction is what the HAE correction allows (Figures C.2 and C.3).

We thank the reviewer again for their careful reading and constructive comments. We hope these clarifications and additional real-data evidence address the reviewer's concerns.

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

The paper introduces homogeneous autoencoders (HAEs) for nonlinear dimension reduction that exactly preserve a prescribed magnitude. The encoder learns a representation of the input direction and rescales it by the original radius, yielding exact magnitude preservation and positive-scale equivariance. A magnitude-preserving decoder includes a radius-dependent angular correction that vanishes asymptotically, accommodating curved geometry at finite radii. The paper establishes sufficient conditions for compatible embeddings, transport of regularly varying tails, and a characterization of homogeneous encoders through universal preservation of radial–angular independence. Experiments on a synthetic curved surface and ERA5 fields compare reconstruction, radial distortion, dependence, and angular generation against PCA and several autoencoder controls.

Critical strengths
  1. This paper presents an architectural contribution with meaningful theoretical justification. The encoder enforces magnitude preservation and scale equivariance independently of the training objective, while the decoder correction permits finite-radius curvature without abandoning asymptotic homogeneity. Beyond the direct architectural identities, Proposition 8 establishes compatibility with embeddings under explicit geometric and dimensional conditions, and Proposition 11 characterizes homogeneity by preserving radial–angular independence for every input law. These results explain both the feasibility of the constraint and why it can matter beyond simply retaining the radius. The paper also appropriately distinguishes existence results from guarantees about learned representations.

  2. The experimental design provides useful evidence for the contribution of homogeneity. Specifically, the AE + radius baseline preserves magnitude exactly while allowing latent directions to depend on radius, helping distinguish the effect of homogeneity from magnitude preservation alone. The parameter-matched comparisons, penalty baseline, and repeated fits further support the evaluation. Moreover, comparing fitted and empirical angular laws helps separate the effects of the learned representation from those of the angular generative model. The radius-permuted precipitation experiment and threshold sensitivity analysis also provide useful controls. These experiments support the proposed mechanism, even though the generation results on real data are less consistent.

Critical weaknesses
  1. The practical benefits on real data are not fully established. Table 4 shows that AE + penalty achieves lower angular generation error than HAE under both the fitted law (0.0204 versus 0.0238) and the empirical law (0.0137 versus 0.0151). In the radius-permuted experiment, PCA also outperforms HAE under both generation settings (Table 5). Although HAE preserves magnitude exactly and limits the introduction of radial–angular dependence, these properties do not consistently improve generation quality on the real dataset. Consequently, it remains unclear when the proposed constraints provide a practical advantage over the simpler penalty baseline. Additional evaluation on a concrete magnitude-sensitive downstream task would help establish the benefit of these guarantees beyond the radial and equivariance metrics.

  2. Limited evaluation beyond the training magnitude range. In Section 6.1, the autoencoder is trained on all training exceedances, while only the angular generative model is fitted to the lower half. Appendix F also uses radii from held-out observations when constructing generated samples. Although this setup isolates the transfer of the angular law from moderate to larger exceedances, it does not directly evaluate the autoencoder at magnitudes outside its training range. Given the emphasis on extreme-value analysis and asymptotically homogeneous decoding, it remains unclear whether the learned angular reconstruction remains accurate under such extrapolation. Training the autoencoder on a restricted magnitude range and evaluating it at progressively larger held-out magnitudes would strengthen the evidence for this use case.

Questions for authors
  1. Clarification of practical benefit: How should the Table 4 result favoring AE + penalty be interpreted when deciding whether to use HAE? Can you identify a concrete magnitude-sensitive analysis for which the exact structural guarantees provide an advantage despite the higher angular generation error? Any existing results on such an analysis would help establish the practical importance of the proposed constraint; a completely new application is not required to clarify the intended claim.

  2. Additional validation that would strengthen the assessment: Do you have results in which the autoencoder itself is trained only on moderate magnitudes and evaluated at progressively larger held-out magnitudes? Angular reconstruction or generation errors in this setting, particularly relative to AE + radius, would clarify whether the asymptotic decoder structure improves the extrapolation of actual magnitudes. Please distinguish such a test using supplied target radii from evaluation of a complete generator with an estimated radial tail model; the latter would be a broader extension rather than a prerequisite for this representation-focused contribution.

2 comments
AuthorsOpenReview ↗

Response to Review by Reviewer Vqyw

We thank the reviewer for their careful and supportive review. The summary shows an understanding of the paper as we intended and the reviewer correctly distinguishes the architectural identities, the embedding results and the factorisation theorem.

To address the reviewer's concern about HAE's practical benefit and Table 4, we thought it helpful to clarify that our intended contribution is a is a learned representation with specified structural guarantees at comparable reconstruction accuracy to competitors, not a promise that every downstream generative model improves. This goal is mentioned in Section 1 and is encountered in adjacent work imposing similar architectural restrictions instead of loss-based incentives (Cohen and Welling, 2016, Richter-Powell et al., 2022, Su et al., 2022).

As well as this clarification of intended contribution, we hope that the wind experiment set out in the posted general comment and revised draft provides the concrete downstream modelling evidence the reviewer is seeking. In the new experiment, the independence assumed by the factorised tail generator holds approximately in the data, a feature which was lacking in the previous real precipitation data example. HAE is the only neural model that preserves this independence and we show that HAE leads to more accurate generation than the competitors (Table 6). Table 4 remains in the revision as the contrasting case, in which independence is already rejected in the data and no advantage is expected. We hope that this clarifies the situations in which a practical advantage may be expected from employing HAE as a choice of dimension reduction model before latent space generative modelling.

HAE should be preferred to AE + penalty when a representation is required to preserve magnitude exactly, such as for generation of precipitation fields for a flood model. HAE achieves this exactly and AE + penalty only approximately (Section 3). It should also be preferred when a downstream model leverages radial homogeneity in the data (radial-angular independence), which HAE is constructed to preserve (Proposition 11) and other models may invalidate (Table 6).

Regarding evaluation beyond the training magnitudes, no claim is made about the reconstructive ability of HAE beyond its training range and the generation experiment of Section 6.1 is not intended to test extrapolation. Lemma 6 and Proposition 10 establish limiting decoder behaviour and tail transport, but not that the learnt limiting directions match the true ones, which Section 3 states is needed. Since the image of the HAE decoder is naturally conic at large radii (Corollary B.1), accurate extrapolation benefits when the data manifold aligns itself with this conic image. Extrapolative ability therefore depends on the dataset and a general study of extrapolative performance would require analysis across numerous applications and datasets, which we leave to future work.

We thank the reviewer again for their time, attention and constructive review.

References: Cohen and Welling, Group Equivariant Convolutional Networks, ICML 2016. Richter-Powell, Lipman and Chen, Neural Conservation Laws: A Divergence-Free Perspective, NeurIPS 2022. Su, Byeon and Huang, Scaling-up Diverse Orthogonal Convolutional Networks by a Paraunitary Framework, ICML 2022.

Thank you for the clarification. I agree that the main contribution is a representation with exact structural guarantees, rather than a general improvement in downstream generation. The reported wind experiment directly addresses my question about when preserving radial–angular independence can provide a practical benefit. I also appreciate the distinction between asymptotic decoder properties and the ability to extrapolate accurately beyond the training range. I think accurate extrapolation beyond the training magnitude range is an open empirical question rather than a requirement for the present contribution. I maintain my weak acceptance recommendation.

Reviewer_Xks7OpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes Homogeneous Autoencoders (HAEs), which preserve the input radius exactly while learning a lower-dimensional angular representation. The architecture combines a positively homogeneous encoder with a radius-preserving decoder that accommodates curved geometry. The analysis establishes embedding conditions, radial tail-index preservation for regularly varying data, and a characterization of encoders that preserve radial–angular independence. Experiments on a synthetic heavy-tailed surface and ERA5 precipitation fields show reconstruction accuracy comparable to parameter-matched autoencoders and improved angular generation of held-out exceedances relative to a standard autoencoder.

Critical strengths

The paper clearly connects autoencoder design with extreme-value theory. The architecture and its assumptions are well explained, and the inclusion of the AE + radius baseline helps assess what homogeneity adds beyond exact radial preservation.

Critical weaknesses

1. Insufficient justification for exact radial preservation.

Although the paper motivates radial preservation through interpretability and extreme-value analysis, it does not sufficiently establish when exact equality between the ambient radius r(x)r(x) and the latent radius r(z)r(z) offers meaningful advantages over less restrictive alternatives. The embedding results demonstrate that this constraint is compatible with representation learning under sufficient dimensional conditions, but do not establish its benefit for downstream tasks.

For heavy-tailed data, radial tail-index preservation follows directly from retaining each observation’s radius; this guarantee alone does not establish more accurate modeling of angular tail structure or improved downstream performance. Empirically, AE + penalty achieves lower generation errors than HAE on the original ERA5 exceedances under both fitted and empirical angular laws. Thus, further justification is needed to clarify when the benefits of exact radial preservation outweigh its restrictions on the representation.

2. Unclear practical advantage over existing radial–angular approaches.

HAE provides representation guarantees distinct from those of generative models such as ExtVAE [1], particularly through its homogeneous encoder and characterization of radial–angular independence preservation. However, the practical value of these guarantees for extreme-value generation remains insufficiently established. A comparison with ExtVAE on a shared conditional angular-generation task, using matched radial definitions and evaluation settings, would help clarify when HAE’s representation constraints improve generation quality.

[1] Lafon, Naveau, and Fablet. A VAE Approach to Sample Multivariate Extremes. Journal of Statistical Computation and Simulation, 2026.

Questions for authors
  1. Given the comparable reconstruction accuracy and the lack of a consistent generation advantage on ERA5, could the authors provide a more detailed interpretation of the results, clarifying the practical benefits and limitations of exact radial preservation?

  2. How does the effect of exact radial preservation change with latent dimension? Under stronger compression, requiring the latent radius to match the input radius may leave less flexibility to represent angular structure. Could this constraint adversely affect reconstruction or the modeling of extreme dependence? How do the embedding and extreme-value results inform this potential trade-off?

  3. Could the authors compare with ExtVAE under a matched evaluation protocol, including generation from a radial model fitted only to training data? What explains HAE’s inconsistent advantage over simpler baselines on ERA5?

2 comments
AuthorsOpenReview ↗

Response to Review by Reviewer Xks7

We thank the reviewer for their detailed assessment and constructive questions. We are pleased that the reviewer finds the architecture and assumptions well explained and identifies the strength and purpose of the AE + radius baseline. We agree and are glad that the reviewer acknowledges that HAE "provides representation guarantees distinct from those of generative models such as ExtVAE". We do not claim that HAE's guarantees uniformly improve latent generation. Our general comment and revision explain the precipitation results and detail a wind experiment demonstrating HAE's practical advantage in a contrasting setting.

Regarding the justification for exact radial preservation, Section 1 motivates interpreting latent magnitude on the ambient scale, with Proposition 5 serving to show why approaches targeting reconstructive performance alone cannot provide this latent space interpretability. We agree that tail index preservation follows directly from preserving the radius, as intended by our construction. Exact radial preservation is not intended to necessarily give more accurate angular modelling. Combined with homogeneity, radial preservation also keeps radial-angular independence whenever the data possess it. Retaining the radius alone does not guarantee this property (Proposition 11). Akin to established work (Cohen and Welling, 2016, Richter-Powell et al., 2022, Su et al., 2022), these guarantees are a property of the architecture rather than loss based and thus desirably hold independently of the training procedure.

Providing further interpretation of the results and contribution, we emphasise that the results show that the guarantees of the HAE architecture, which Section 1 states "need not reduce reconstruction error", come at little practical cost, which is demonstrated by comparable reconstruction accuracy to the neural competitors. Beyond preserving the independence assumptions required for factorised tail models or other approaches leveraging homogeneity in the data, they also allow fields and analysis to be generated in latent space at a specific radius, such as wind and precipitation fields of a prescribed magnitude for a storm impact model, with an appropriate radial functional. Section 8 discusses situations in which preserving magnitude may be inappropriate.

Concerning HAE at lower latent dimensions, exact radial preservation and latent homogeneity may indeed limit angular reconstruction under stronger compression. Section 4 and Appendix D.3 provide a detailed analysis of these restrictions. The presented extreme value results (Propositions 9 to 11) apply to positive latent dimensions for which the stated maps satisfy their hypotheses and compression affects the recovered tail measure via the limiting angular reconstruction (Appendix E.5). The practical consequences of stronger compression are not observed to be restrictive in the experimental settings considered. This is demonstrated by both the synthetic surface experiment, operating at the known minimal latent dimension, and the newly presented wind experiment, where HAE compresses 195 dimensions to 2 with reconstruction error comparable to the other neural competitors, all of whom are ahead of PCA (Table 6).

On comparison with ExtVAE, we would like to reiterate that HAE is a dimension-reducing representation learning architecture rather than a generative framework, with no claim of uniform superior generative performance. The deliberately simple standard von Mises-Fisher model is used to demonstrate Proposition 11 empirically, not a claimed contribution of a superior generative framework. Using the same downstream generative model across latent representations allows us to isolate the effects of the architectural constraints. A latent generative model suitable for comparison with ExtVAE is beyond the intended scope of this work but would certainly be an interesting avenue for future research (Section 8). HAE's latent radii are identical to the ambient radii, so a radial model fitted only to training data would transfer unchanged between the latent and ambient spaces. The precipitation results exist to show that HAE need not provide an advantage for generation if the assumptions the downstream model relies upon do not hold in the first place. For the radius-permuted precipitation exceedances, the fitted angular law is identified in Appendix J as the key limitation on the neural models' generation.

We hope these clarifications and additional evidence address the reviewer's concerns behind the low soundness score, and would welcome clarification of any remaining issues regarding the theoretical guarantees or experimental evidence.

References: Cohen and Welling, Group Equivariant Convolutional Networks, ICML 2016. Richter-Powell, Lipman and Chen, Neural Conservation Laws: A Divergence-Free Perspective, NeurIPS 2022. Su, Byeon and Huang, Scaling-up Diverse Orthogonal Convolutional Networks by a Paraunitary Framework, ICML 2022.

Thank you for the clarifications and the additional experiments. These have addressed several of my original concerns. I now place less emphasis on uniform generative gains and more on HAE’s representation guarantees. The additional wind experiment provides useful evidence that these guarantees can be beneficial in an appropriate setting.

I remain somewhat cautious about the novelty of the combined design and the range of real-data settings in which the decoder correction is important. However, these concerns relate primarily to the significance and scope of the contribution, rather than specific issues with its soundness.

Taking the theoretical analysis and the additional empirical evidence together, I am raising my recommendation to weak accept. I have also lowered my confidence to reflect limitations in my initial understanding of the contribution.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Provision of further practical evidence with an additional experiment on wind data

We thank the reviewers for their time, careful reading and constructive comments. The reviews note that on the ERA5 precipitation exceedances HAE fails to uniformly outperform the benchmarked competitors. In particular, AE + penalty has lower generation error than HAE (Table 4) and PCA outperforms all neural models once the radii are permuted (Table 5).

The exceedance experiments use a deliberately simple latent generative model to demonstrate Proposition 11 empirically, that the radial-angular independence assumption required to implement factorised radial-angular tail models is preserved through HAE's nonlinear dimension reduction whenever they hold in the data. In the precipitation data, however, radial-angular independence is already rejected by an ambient-space permutation test at each threshold examined (Appendix J). The factorised generator's assumption is therefore violated before encoding, and since HAE does not claim to remove existing dependence, no generation advantage is expected there. The experiment is intended to serve to demonstrate that HAE need not increase generative performance for latent space models when their assumptions are incorrect in ambient space. In the radius-permuted precipitation experiment, where independence is restored, HAE indeed preserves it.

We recognise the reviewers' concern that it would be more compelling to see a latent generative model under HAE's encoding outperform the competitors on natural real world data, rather than only on synthetic or permuted data and agree that the addition of such an experiment would increase the demonstrated strength and applicability of the contribution.

To address this, the uploaded revision includes the addition of a contrasting setting, in which the radial-angular independence assumption does hold. An experiment is run on the ERA5 u10 zonal 10m wind field under the same setup and evaluation procedure as for the original precipitation data (detailed in Appendix J for the original precipitation experiments and a new Appendix K for the wind experiment).

HAE is the only neural model that preserves the independence upon which the latent angular generator relies, resulting in the HAE latent model producing the lowest generation error under both angular laws. HAE expectedly has reconstruction error comparable to the neural competitors (Table 6).

Section 6.2 now presents precipitation and wind alongside each other as contrasting cases. We have made minor changes to the abstract, introduction and discussion sections to match. Existing tables 1 to 5 have kept their numbering. As well as this, we discovered and fixed a small sampling bug in the independence test. The correction only resulted in some minor changes in dependence statistics and does not alter any of the conclusions presented and is included in the reuploaded submission. Please see below for the results from the the new wind experiment, means over 25 fits. Full results with standard deviations are in Table 6.

AmbientPCAAEAE + penaltyAE + radiusHAE
Generation error, fitted law0.02040.04690.02980.01710.0138
Generation error, empirical law0.009460.01860.04670.02890.01770.0135
Distance correlation0.05360.04670.1650.1040.08690.0540
Fits rejecting independence0/10/125/2520/2519/250/25
Reconstruction RMSE0.05470.04880.04850.04840.0491

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.