acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Discounting of Compression Error Feedback for Decentralized Learning

Abstract

Decentralized learning enables collaborative model training across multiple nodes without relying on a central server, where each node performs local updates using its own data and exchanges compressed updates with neighboring nodes. Although numerous methods have been proposed for decentralized non-convex optimization under communication constraints, effectively correcting compressed error and mitigating the impact of network topology, particularly in sparse networks, remains an open challenge. To address this issue, we propose A-DEFEAT-GT (adaptively discounted error feedback algorithm for decentralized learning with gradient tracking), which integrates error-feedback with an adaptively tuned discount coefficient to dynamically control the influence of accumulated compression errors, together with gradient tracking technique to deal with heterogeneous data, especially in sparse networks, without relying on momentum propagation or consensus model update that require additional computation. Experiments demonstrate that A-DEFEAT-GT consistently outperform existing decentralized methods by achieving higher test accuracy under the same communication budget, particularly under sparse network conditions. These results highlight both the effectiveness and practical efficiency of the proposed method.

Primary area:
optimization
Keywords:
Decentralized learning, adpative discouting, network topology, compression, error feedback

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_vZ3nOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper designs two algorithms for decentralized training with compression. They focus on the algorithm design with discounted error feedback, which allows balancing the error term into the next iterations through a multiplier. Such an idea was primarily studied in a centralized setting, so this paper extends the study to a decentralized setting.

Critical strengths
  • The dependency on the network topology is improved for the proposed algorithms compared to prior work. This is important, especially in the case of sparse communication networks.

  • The analysis is performed with and without the bounded gradient divergence assumption. To mitigate the use of this assumption, the authors incorporate standard gradient tracking into the algorithm design. This is an important step since it allows theoretical convergence without heterogeneity bounds, an important feature in federated learning applications.

  • The experiments demonstrate improved performance of the proposed algorithms in practice over existing baselines.

  • I went over the proofs and did not find any problems

Critical weaknesses
  • The motivation behind the use of discounted error feedback is not clear enough. It is unclear to me why we should focus on it. It is one of the ways that worked in the centralized setting, but there are other existing ideas. I encourage the authors to provide more discussion around it and explain why they chose to focus on it. This would make the motivation of the work clearer.

  • Figure 1 is not understandable. The text provides insufficient information to understand the idea behind it. It is not clear what is being tested there, what is changing, and so on. The authors should provide a clearer description of Figure 1, which serves as some sort of motivation that the naive idea of discounted error feedback might fail.

  • The influence of compression on the convergence is hidden in the main part. Since it is a main focus of the work, it must be shown explicitly in the main part.

    • In particular, for A-DEFEAT-GT, the compression factor 1−β1-\beta affect the deterministic 1/T1/T term LD(1−β)2ρT\frac{LD}{(1-\beta)^2\rho T} and the middle term (LDσ)2/3(1−β)4/3ρ2/3T2/3\frac{(LD\sigma)^{2/3}}{(1-\beta)^{4/3}\rho^{2/3}T^{2/3}}. This means that the improvement in the topology ρ\rho comes with deterioration in the compression, since the optimal deterministic term is LD(1−β)ρT\frac{LD}{(1-\beta)\rho T}.
    • Moreover, it seems that σ2\sigma^2 is also present in 1/T1/T term through E∣G0∣2\mathbb{E}\\|G_0\\|^2, which goes against lower bounds. This constant can significantly slow down the convergence (take f1(x)=L/2x2−MX,f2(x)=L/2x2+Mxf_1(x) = L/2x^2 - MX, f_2(x) = L/2x^2+Mx; I think the 1/T1/T term grows with both LL and MM, while it should only with LL)
    • For A-DEFEAT, the deterministic term becomes optimal, but at the cost of using a gradient divergence assumption.
    • Therefore, the authors should discuss these limitations of the analysis, whether it's an artifact of the analysis or a property of the algorithms. I suggest adding controllable experiments that satisfy the assumptions and demonstrate that these limitations of the analysis are not critical.
  • The convergence is limited to the general non-convex setting. EControl's analysis, where the discounted error feedback was proposed, demonstrates convergence in three cases: non-convex, convex, and strongly convex. This is not the case for an alternative EF21-SGDM algorithm. The authors should discuss whether the analysis can be extended to convex and strongly convex cases to demonstrate this flexibility also in the decentralized setting.

  • The choice of hyperparameters in practice might violate the theoretical restrictions; I believe there is not enough information to confirm this. Therefore, I encourage the authors to discuss whether there is indeed a gap between theory and practice.

Questions for authors
  • Can the authors clarify what min⁡ν,m,τ=Ω(1)\min\\{\nu,m,\tau\\}=\Omega(1)? What should be taken to ∞\infty here? Is it always satisfied, or only for some particular cases?
7 comments

We thank the reviewer for the careful reading and provide so many instructive comments, including the check of all the proofs. We would like to address the comments as follows.

W1. This is a very good comment. Our original goal was to design an efficient local error-feedback scheme for decentralized learning, where residual correction can be adjusted directly without necessarily attenuating the entire consensus correction. The reason we are using error-feedback is that it is an effective approach to compensate the error due to compression and it has been used in other papers. Centralized methods such as EControl provide a useful starting point by introducing a coefficient to regulate feedback strength. However, unlike EControl's shared-model, server-based setting, decentralized nodes maintain different local models and exchange information only with neighbors. Compression residuals therefore influence both local optimization and subsequent consensus dynamics, making their control a distinct decentralized challenge. This motivates introducing a dedicated coefficient on the residual entering each compressed model update. Its role differs from the consensus gains in CHOCO, BEER, and MoTEF, which scale the consensus correction. Furthermore, because the relative magnitudes of the residual and the current tracked update vary across nodes and training stages, we adapt this coefficient locally using their clipped norm ratio. The residual accumulates past compression errors, whereas the current tracked-gradient step reflects the present optimization scale. Comparing these quantities lets us adjust the feedback strength when accumulated errors become large or small relative to the current update. Our objective is to develop locally adaptive residual correction with a convergence guarantee under explicit conditions. This combination of direct residual control, local adaptation, and theoretical analysis explains our focus on adaptive discounted error feedback.

W2. Figure 1 is a toy motivating example that isolates the effect of the feedback-coefficient rule on a four-node quadratic optimization problem. It compares two fixed coefficients, γ=0.1\gamma=0.1 and γ=1\gamma=1, with an adaptive feedback rule, while keeping the data, initialization, sample sequence, and other settings identical. The example uses a ring network, learning rate η=0.1\eta=0.1, the dense contractive map C(b)=0.05b\mathcal{C}(b)=0.05b, and averaging weights of 1/31/3 for each node and its two neighbors. Each panel shows node 1's trajectory over 40 iterations; the blue arrow indicates its final distance to the global minimizer. The purpose is to illustrate that choosing a fixed discount coefficient can substantially affect the trajectory, and that adapting the correction to the relative magnitudes of the local gradient and accumulated residual can be beneficial. The adaptive choice achieves a smaller final distance in this example, providing motivation for our design. We agree that the original description can be improved and will expand the caption and Appendix A with these settings, the precise adaptive rule and update order, and clear definitions of the plotted markers.

W3. Explicit compression dependence. We would like to thank Reviewer vZ3n for this comment. We will include the compression parameter explicit in all the theorems in the main paper. Currently, the explicit compression dependence is shown in Appendix F, Corollaries F.3 and F.5, equations (75) and (78), and the limitations of A-DEFEAT-GT and A-DEFEAT are discussed in Appendix H. All explicit dependence on β\beta and ρ\rho for the convergence upper bound is shown in Appendix G.

The appendix corollaries use consensus gain (67) and their respective prescribed stepsizes. For A-DEFEAT-GT, Corollary F.3 additionally chooses τ=1/2\tau=1/2; this is a sufficient choice for the explicit corollary, while the main Theorem 4.1 allows general τ\tau satisfying its stated range and stability conditions. Under the assumptions stated in Theorem 4.1 and 4.2, the resulting rates are

RTGT=O(σLDnT+(LDσ)2/3(1−β)4/3ρ2/3T2/3+LD(1−β)2ρT+∥F0Π∥F2nρT+E∥G0∥F2nT),R_T^{\mathrm{GT}}=\mathcal O\Bigg(\frac{\sigma\sqrt{LD}}{\sqrt{nT}}+\frac{(LD\sigma)^{2/3}}{(1-\beta)^{4/3}\rho^{2/3}T^{2/3}}+\frac{LD}{(1-\beta)^2\rho T}+\frac{\lVert F_0\Pi\rVert_F^2}{n\rho T}+\frac{\mathbb E\lVert G_0\rVert_F^2}{nT}\Bigg),

and

RTA=O(σLDnT+(LD)2/3(1−β)2/3T2/3(σ2ρ+ϵ2ρ2)1/3+LD(1−β)ρT),R_T^{\mathrm{A}}=\mathcal O\left(\frac{\sigma\sqrt{LD}}{\sqrt{nT}}+\frac{(LD)^{2/3}}{(1-\beta)^{2/3}T^{2/3}}\left(\frac{\sigma^2}{\rho}+\frac{\epsilon^2}{\rho^2}\right)^{1/3}+\frac{LD}{(1-\beta)\rho T}\right),

where RT=T−1∑t=0T−1E∥∇f(xˉt)∥2R_T=T^{-1}\sum_{t=0}^{T-1}\mathbb E\lVert\nabla f(\bar x_t)\rVert^2 and D>0D>0 bounds the initial objective gap. The A-DEFEAT rate additionally assumes bounded gradient divergence ϵ2\epsilon^2.

W3(a). Thank you for highlighting this distinction. We agree that the explicit A-DEFEAT-GT bound has weaker compression dependence: the deterministic term scales as (1−β)−2(1-\beta)^{-2}, and the T−2/3T^{-2/3} term scales as (1−β)−4/3(1-\beta)^{-4/3}. Our result therefore does not establish optimal compression dependence or a uniformly better joint dependence on compression and topology. Our contribution concerns the improved spectral-gap exponents under the stated assumptions. For example, compared with MoTEF, the topology factors improve from ρ−3\rho^{-3} to ρ−1\rho^{-1} in the deterministic term and from ρ−5/3\rho^{-5/3} to ρ−2/3\rho^{-2/3} in the intermediate stochastic term, while the compression factors worsen. Thus, the comparison depends on both network connectivity and compression quality. Under the bounded-gradient-divergence assumption, A-DEFEAT already achieves the same ρ−2/3\rho^{-2/3} dependence in the T−2/3T^{-2/3} and ρ−1\rho^{-1} dependence in the T−1T^{-1}. In our current analysis, controlling the compressed model and tracker updates together leads to larger compression penalties, which is (1−β)−2(1-\beta)^{-2} in the objective-gap term and (1−β)−4/3(1-\beta)^{-4/3} in the T−2/3T^{-2/3} term. The additional compression factors arise in our sufficient stability analysis of the coupled compressed model and tracker recursions. We do not establish that this deterioration is unavoidable for the algorithm. Obtaining sharper joint dependence remains an open direction. We discussed the dependence on β\beta in Appendix H, but we will make it clear in the final manuscript.

W3(b). We agree that the initialization term can be large and represents a limitation of our finite-time guarantee. We will add this limitation in Appendix H in the final manuscript. Nevertheless, we believe this is a limitation of our analytical approach for gradient tracking and would like to point out that similar initialization dependence appears in the gradient tracking related methods. The stochastic contribution arises because the tracker is initialized with stochastic gradients. For example, initial Lyapunov terms also appear in BEER (Theorem 2), while MoTEF (Remark 2) retains its initial Lyapunov value under single-sample initialization.

We also would like to point out that, for a fixed problem with positive stochastic variance, this contribution decays as T−1T^{-1}, faster than the leading T−1/2T^{-1/2} stochastic term, and therefore does not alter the asymptotic convergence rate or linear speedup. In the deterministic setting, the guarantee remains O(1/T)O(1/T), with an initialization-dependent coefficient. Our guarantee avoids a uniform bounded-heterogeneity assumption while retaining the initial-gradient contribution. Hence, we believe this is a limitation of our analysis but it is not a critical weakness.

W3(c). Yes. This is exactly the benefit and the limitation of A-DEFEAT. It can improve the dependence on the compression factor but at the cost of the requirement of the bounded gradient divergence assumption. Removing the bounded gradient divergence assumption is a principal motivation for introducing gradient tracking. This tradeoff is also shown in the experiments in Appendix I. However, we do not believe this is a weakness of our paper rather than a limitation of A-DEFEAT algorithm.

W3(d). We agree that the distinction between limitations of the analysis and limitations of the algorithms should be clarified. Based on our above comments, we will explicitly discuss the compression–connectivity tradeoff, the dependence on initial local-gradient norms, and the limitation of A-DEFEAT in the final manuscript. These factors arise in sufficient upper bounds, and we do not claim that their precise powers are necessary.

Our existing experiments provide evidence of useful performance across the evaluated network and heterogeneity settings. For example, on FashionMNIST with 5% top-kk compression and Dirichlet parameter 0.050.05, A-DEFEAT-GT maintains approximately 81% accuracy across rings with 10–200 nodes. These results support practical effectiveness in those settings. They do not isolate the individual theoretical factors or establish that all sufficient theorem conditions hold. Accordingly, our conclusions are limited to improved proved network dependence under the stated conditions and empirical effectiveness in the evaluated settings. We do not claim that the remaining dependencies are tight or practically negligible throughout the admissible problem class.

We will add controllable experiments (synthetic problems, possibly extension of the motivating example) in the final manuscript.

W4. Our main focus is smooth nonconvex optimization, motivated by objectives that arise in decentralized neural-network training. The cited analyses of DCD, DeepSqueeze, CHOCO-SGD, BEER, and MoTEF all cover smooth nonconvex objectives. Only CHOCO-SGD and MoTEF also cover the strong convex objectives. The smooth nonconvex setting therefore provides a common basis for comparing the methods. Our derivations can be extended to the convex cases with the different convergence bound and rate. We agree that deriving the rates that exploit convexity or strong convexity would complement the current analysis, and we leave this to future work due to the current length of our paper.

W5. Thank you for raising this point. We agree that the relationship between the theoretical restrictions and the experimental hyperparameters needs clarification. Our convergence guarantees require explicit stability conditions and learning-rate bounds, whereas the experiments select learning rates and tracker stepsizes empirically and use consensus gain ω=1\omega=1. Other hyperparameters are chosen based on the theoretical conditions as discussed in the response to Reviewer 63EM. Thus, there is a gap between the parameter regime covered by our guarantees and the settings used in these experiments.

As we mentioned before, these conditions are sufficient, and we do not establish their necessity. Their violation therefore means that our theorems do not certify the corresponding runs; it does not imply that the algorithms must diverge. The experiments demonstrate practical performance beyond the certified regime, but they do not validate the convergence bounds in that regime. Our constructive consensus-gain choice provides an admissible regime for any fixed compression level, subject to the remaining assumptions and stepsize restrictions, but that choice was not used in the reported experiments.

We will make this distinction explicit in the experimental discussion and limitations, report the selected hyperparameters alongside the applicable theoretical restrictions, and identify conditions that fail or remain unverified. Extending the guarantees to the empirically effective parameter regime remains an open problem.

Q1. Here, Ω(1)\Omega(1) means uniformly bounded away from zero; no quantity is required to tend to infinity. We will replace the notation with the explicit condition

min⁡{ν,ϑ,m,τ}≥c>0,\min\lbrace\nu,\vartheta,m,\tau\rbrace\ge c>0,

where cc is a fixed numerical constant across the network and compression settings under consideration. This is a sufficient condition in Theorem 4.1, ensuring uniform control of the stability constants in the convergence guarantee.

Reviewer_63EMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies decentralized non-convex optimization with compressed communication. The authors propose an adaptive error-feedback mechanism that adjusts the discount coefficient based on the current update and residual norms. They develop two algorithms, A-DEFEAT-GT with gradient tracking and A-DEFEAT without gradient tracking to reduce communication cost. Theoretical analysis establishes convergence guarantees with improved dependence on the network spectral gap under the stated conditions. Experiments on different datasets and network topologies demonstrate the effectiveness of the proposed algorithms.

Critical strengths
  1. The paper clearly explains the motivation and challenges of compressed decentralized learning. The overall presentation is well organized and easy to follow.
  2. The proposed adaptive discounting idea is simple yet effective, and its effectiveness is supported by the experimental results.
  3. The paper establishes rigorous convergence guarantees for both algorithms, achieving improved dependence on the network spectral gap compared with existing works.
Critical weaknesses

None.

Questions for authors

Could the authors clarify how the experimental settings align with the theoretical conditions required by Theorems 4.1 and 4.2?

1 comment

We thank the reviewer for the positive assessment and the concrete suggestions. We would like to address them as follows.

Q1. First, the neural-network experiments use the model, residual, and tracker recursions of Algorithms 1 and 2, with the adaptive norm ratio and clipping choices described in the experimental protocol. Hyperparameters are selected from the ranges in Appendix B. The connectivity matrices in the code satisfy Assumption 3.1. The unbiased mini-batch gradients of Assumption 3.4 are guaranteed by the data sampling, which draws indices uniformly at random from the local data. The tracker step in Theorem 4.1 requires 0<τ≤min⁡{1,1/μ}0 < \tau \leq \min\lbrace 1, 1/\mu \rbrace. For the ring networks mainly used in experiments, μ=4/3\mu=4/3 and thus 0<τ≤3/40<\tau\leq3/4, which contains our search grid τ∈{0.02,…,0.5}\tau\in \lbrace 0.02, \ldots, 0.5 \rbrace.

Second, for other assumptions, our neural-network experiments can not certify the global smoothness, variance and gradient divergence bounds (for A-DEFEAT), and the resulting stepsize restrictions (34) and (64). In addition, the parameter conditions stated in these two theorems are sufficient for the convergence guarantees but may not be necessary. In practice, as in the literature, we select tunable hyperparameters for our methods and the baselines by grid search over the ranges, determined by the assumptions in these two theorems, in Appendix B.

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

This paper studies decentralized optimization under limited communication and proposes A-DEFEAT-GT, which combines adaptive residual feedback with gradient tracking. The theoretical analysis preserves the difference structure of consecutive compression residuals and stochastic gradient noise, and improves the dependence of the convergence bound on the network spectral gap under explicit stability conditions. The experiments cover four image classification datasets, multiple topologies, network sizes, and levels of data heterogeneity. They compare performance under both a fixed number of aggregation rounds and a fixed communication budget.

Critical strengths

1.Analyzing the network filtering gain through its difference structure is a major theoretical contribution. To the best of my knowledge, this is the first work to do so.

2.A-DEFAT-GT and A-DEFEAT are, in fact, two theoretically complementary structures, and the paper discusses their relationship clearly.

3.The experiments are comprehensive.

Critical weaknesses
  1. The theoretical presentation could be clearer. For example, Assumption 1 could be stated much more clearly, but is currently formulated in a complicated way.

  2. The paper uses the test set for hyperparameter selection. I find this practice unusual; hyperparameters should be selected using a validation set. Have I misunderstood the protocol? Please correct me if so.

  3. The paper could consider a wider range of network topologies.

Questions for authors

See the weaknesses.

3 comments

We thank the reviewer for the positive assessment and the concrete suggestions. We would like to address as follows.

W1. We agree. Assumption 3.1 mixes the assumption itself with definitions of derived spectral quantities. We will simplify Assumption 3.1 and introduce the spectral quantities separately. Assumption 3.1 will be "For n≥2n\ge 2, let J=1n1n⊤/nJ=\boldsymbol{1}_n\boldsymbol{1}_n^\top/n, where 1n\boldsymbol{1}_n is the all-ones vector. The mixing matrix WW is symmetric and entry-wise nonnegative, with W1n=1nW\boldsymbol{1}_n=\boldsymbol{1}_n and ∥W−J∥2<1\|W-J\|_2 < 1." Then, the main text will say "The row-sum condition gives the consensus eigenvalue λ1(W)=1\lambda_1(W)=1. We order the remaining eigenvalues by decreasing absolute value and define the spectral gap immediately after the assumption as ρ:=1−∥W−J∥2=1−∣λ2(W)∣\rho := 1-\lVert W-J \rVert_2 = 1-\lvert \lambda_2(W) \rvert. We will define λ−\lambda_{-}, μ\mu, and ν\nu where we introduce the stability margins".

W2. We would like to thank Reviewer WzMV for this meaningful question. When we wrote our paper, we observed that some related works use the test set for hyperparameter selection. For example, EControl explicitly selects stepsizes on the test set (Section 7.3 in EControl), and CHOCO's released code includes a utility that ranks configurations by test accuracy. This motivated us to use the test set for the hyperparameter search. In addition, we observed numerical instability when using training loss for some of the baseline algorithms. To obtain the best performance of the baseline algorithms, we believe using the test set for hyperparameter search is a reasonable choice. Note that for the proposed A-DEFEAT and A-DEFEAT-GT, we do not need to search for the discounted error feedback factor, while baseline algorithms such as MoTEF and BEER may need to. Hence, we believe using the test set for hyperparameter search improves the performance of the baseline algorithms, not ours.

Nevertheless, we would like to clarify that even if we use a validation set for hyperparameter search, the proposed algorithms still achieve superior performance. We therefore additionally report comparisons with hyperparameters selected using a separate validation set. We hold out 10% of the training set (600 images per class) before the data is partitioned among nodes. We use the parameter grids from Appendix B and select the setting with the highest mean validation accuracy over the last 50 rounds. The A-DEFEAT and A-DEFEAT-GT rows below use the fixed clipping interval [1−1/2,1+1/2][1-1/\sqrt{2}, 1+1/\sqrt{2}]. The selected setting is then run with three new seeds and evaluated on the test set. The setting is the one of Table 2 (50-node ring, FashionMNIST, α=0.05\alpha = 0.05, top-5%, 1000 rounds). We report test accuracy (%), with the mean and population standard deviation over the three seeds.

MethodTable 2 (test-selected)Validation-selected
DCD62.1 ± 0.368.8 ± 0.7
CHOCO68.1 ± 0.477.5 ± 0.6
DeepSqueeze68.7 ± 0.576.9 ± 0.8
BEER73.8 ± 0.379.0 ± 0.4
MoTEF74.9 ± 0.957.3 ± 19.5
A-DEFEAT76.2 ± 0.380.5 ± 0.2
A-DEFEAT-GT81.0 ± 0.481.7 ± 0.2

A-DEFEAT-GT retains the highest reported accuracy under validation-based selection. MoTEF shows substantial seed-to-seed variability, although all three runs completed 1000 rounds. A-DEFEAT achieves comparable accuracy with one rather than two compressed messages per round. The new runs use different splits and seeds, so the absolute accuracies are not a controlled comparison with Table 2. These validation-based results cover the FashionMNIST settings reported here and in W3.

W3. We agree that more network topologies would strengthen the comparison and improve our paper. We would also like to emphasize that the paper already considers a number of network topologies: Appendix I includes four topologies (ring, cyclic-4, random-4, and random-9) with 10 to 200 nodes. To make our results more convincing, we add a 100-node two-dimensional torus, compared with cyclic-4 and random-4 graphs of the same size and degree under the validation protocol, so that network size, degree, and protocol are held fixed while the topology varies. The table below uses 100 nodes, degree 4, FashionMNIST, α=0.05\alpha=0.05, top-5%, 1000 rounds, and three fresh seeds, and reports validation-selected test accuracy (%).

MethodTorus (new)Cyclic-4Random-4
DCD71.0 ± 0.870.2 ± 0.771.2 ± 0.8
CHOCO79.9 ± 0.378.8 ± 1.180.1 ± 0.3
DeepSqueeze79.8 ± 0.478.4 ± 0.979.6 ± 0.5
BEER80.7 ± 0.979.7 ± 0.681.2 ± 0.2
MoTEF76.8 ± 8.381.0 ± 0.780.1 ± 3.7
A-DEFEAT82.5 ± 0.181.7 ± 0.282.5 ± 0.1
A-DEFEAT-GT83.5 ± 0.183.0 ± 0.283.4 ± 0.1

I believe this paper has reached an acceptable level, so I will maintain my positive score.

Dear Reviewer WzMV,

Thanks a lot for your response!! We appreciate all your efforts in the review process and your positive score. Please let us know if you have any additional questions or comments.

Best regards, Authors of Submission 36576

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.