acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Aware Slow-Fast Learning for Decentralized Federated Learning

Abstract

Data heterogeneity in decentralized federated learning is inherently non-uniform: different model parameters exhibit different levels of cross-client stability and adaptation speed, while different clients vary in their compatibility for knowledge exchange. However, most existing approaches still rely on largely uniform optimization and consensus rules, overlooking this fine-grained structure of heterogeneous collaboration. We therefore argue that effective decentralized learning should jointly account for two complementary dimensions: parameter-wise adaptation timescales and client-wise collaboration geometry. To this end, we propose Geometry-aware Slow-Fast Learning (GeoSF), a decentralized learning framework that explicitly models both forms of heterogeneity. At the parameter level, GeoSF organizes model parameters into two dynamically evolving groups: slow parameters that remain stable across clients and preserve transferable shared knowledge, and fast parameters that respond more strongly to local data and support client-specific adaptation. GeoSF then matches their distinct roles with different optimization and communication timescales, using smaller learning rates and more frequent synchronization for slow parameters, while allowing fast parameters to evolve more rapidly and synchronize less frequently. At the client level, GeoSF adapts peer-to-peer consensus according to the geometric compatibility of neighboring models, promoting beneficial information exchange while suppressing interference from heterogeneous peers. Together, these mechanisms establish a unified slow-fast learning paradigm that jointly coordinates parameter dynamics and client interactions under decentralized heterogeneity. Extensive experiments across feature-skew and label-skew settings, model architectures, communication topologies, and client scales demonstrate that GeoSF consistently outperforms representative federated and decentralized baselines.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Federated Learning, Slow-Fast Learning, Heterogeneous Data

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_aFHcOpenReview ↗
Rating
3
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeoSF, a decentralized federated learning framework that dynamically separates parameters into slow and fast groups with different learning rates and synchronization frequencies. It further adjusts neighbor aggregation weights based on model similarity. Experiments across heterogeneous data distributions, model architectures, and communication topologies demonstrate the effectiveness of the proposed approach.

Critical strengths

-GeoSF considers both differences in parameter stability and compatibility between clients. Combining adaptive parameter partitioning with different update and synchronization frequencies provides a flexible way to balance shared knowledge and local personalization.

-The experiments cover feature skew and label skew, image and text classification, different model architectures, multiple communication topologies, and settings with up to 100 clients. The results show strong improvements across diverse configurations.

-The paper evaluates different combinations of the update, synchronization, and weighting mechanisms, compares score-based parameter selection against random selection, and examines sensitivity to major hyperparameters.

Critical weaknesses

-The paper would benefit from a more thorough discussion of what these forms of heterogeneity represent, how they arise, and how they affect optimization and collaboration in federated learning.

-Datasets such as CIFAR-10, CIFAR-100, and Office10 provide limited evidence of performance under more complex data heterogeneity. Including a more challenging benchmark, such as DomainNet, would strengthen the empirical validation.

-The tasks considered may not be sufficiently challenging. The high accuracy achieved in some classification settings raises the question of whether these tasks adequately test the capabilities of the proposed method. I would encourage the authors to explore more demanding vision or embodied tasks to demonstrate its broader applicability.

Questions for authors

Could the authors clarify how model parameters are selected for computing distances between client models? How do the computational and communication costs scale as the number of parameters increases, and does the approach remain practical for large models?

4 comments

We sincerely thank Reviewer aFHc for carefully evaluating our work and providing constructive feedback that helped us improve the manuscript. Below, we provide point-by-point responses to these comments.

Weaknesses:

W1: Discussion of heterogeneity and its effects on federated learning.

A1: Thank you for your valuable suggestions. Heterogeneity refers to the divergence in the direction (sign consistency) and magnitude (variance) of model parameters arising from differences in data distributions across clients, as illustrated in Fig. 1(b) of the manuscript. This heterogeneity complicates both optimization and collaboration. Specifically, direct model aggregation may combine conflicting parameters, thereby degrading model performance on individual clients. Based on the degree of parameter divergence, we divide model parameters into “slow” and “fast” components: the former support federated collaboration through aggregation, while the latter facilitate local adaptation. We have expanded the discussion in the revised manuscript to clarify the impact of heterogeneity.

W2: Evaluation on the more challenging benchmark (DomainNet).

A2: Thank you very much for this helpful suggestion. We agree that more challenging benchmarks would provide a more thorough evaluation of GeoSF under complex data heterogeneity. Accordingly, we have conducted additional experiments on DomainNet. As shown below, GeoSF achieves the best average performance, with a C.avg of 23.31±0.43 and a D.avg of 29.05±0.44, further supporting its effectiveness under more challenging domain heterogeneity. We have included these experiments and the corresponding discussion in the revised manuscript.

MethodClipartIgraphPaintQdrawRealSketchC.avgD.avg
SOLO11.443.045.8942.1222.296.3015.18±0.1621.63±0.17
FedAvg22.424.9013.0141.9225.2411.0019.75±0.2124.94±0.29
Ditto20.774.6712.0140.6624.3015.1919.60±0.1324.51±0.12
D-PSGD21.304.8912.6842.4325.7015.0620.34±0.2225.57±0.08
DFedAvgM20.564.7012.2042.7925.8214.4620.09±0.1025.50±0.18
DisPFL22.705.0213.2941.8925.0614.8320.47±0.1725.40±0.21
DFedSMDC23.285.3213.6541.8626.4315.7621.05±0.1026.02±0.10
DFedPGP22.654.8512.9641.9423.916.5718.81±0.3524.04±0.41
GeoSF26.055.4413.2647.0030.0218.0723.31±0.4329.05±0.44

W3: Evaluation on more challenging tasks.

A3: Thank you for your valuable suggestion. We agree that evaluation on more challenging tasks would better demonstrate GeoSF’s broader applicability. To this end, we have extended our experiments beyond image classification to multi-center prostate MRI segmentation on the ProstateMRI dataset, which contains MRI scans and segmentation annotations from six medical centers. This task requires precise pixel-level delineation of anatomical structures while handling data heterogeneity across centers, providing a complementary and demanding evaluation of GeoSF. The results are summarized in the table below.

GeoSF achieves a Dice score of 85.04±0.87 and an IoU of 75.53±1.09, outperforming all evaluated baselines. These results demonstrate that GeoSF’s benefits extend beyond classification to dense prediction under client data heterogeneity. We have added these experiments and discussions to the appendix of the revised manuscript.

MetricSOLOFedAvgDittoD-PSGDDFedAvgMDisPFLDFedSMDCDFedPGPGeoSF
Dice80.54±0.4882.82±0.4382.02±0.7983.01±1.4083.63±0.3683.08±0.6382.48±0.8481.15±1.3185.04±0.87
IoU69.91±0.6072.79±0.4671.75±0.9873.27±1.6673.97±0.4473.39±0.8472.19±1.1170.82±1.3675.53±1.09

Questions:

Q1: Parameter selection for distance computation.

A1: Thank you for raising this insightful question. In GeoSF, distances between client models are computed using a selected subset of shared parameters. Specifically, each client evaluates directional consistency and magnitude variance across parameter coordinates in its neighbors’ models, then selects the top-ranked coordinates according to a predefined shared-parameter ratio. The resulting binary mask identifies the coordinates used for distance computation. When comparing client mm’s model with a neighbor’s model, the mask constructed locally by client mm is applied to both models to ensure that the distance is computed over corresponding parameter coordinates.

Q2: Scalability of computational and communication costs.

A2: Thank you for raising this important question. To evaluate how GeoSF’s computational and communication costs scale with model size, we tested ResNet18, ResNet50, ViT-Base, ViT-Large, and ViT-Huge, as reported below. Total computation time increases from 15.40 s to 76.96 s, with local training consistently accounting for most of the computation. For ViT-Huge, computation beyond local training takes only 3.68 s, approximately 5.0% of the local training time. This indicates that GeoSF’s additional computational overhead remains modest even for large models under the tested settings.

Furthermore, GeoSF’s communication cost scales linearly with the number of transmitted parameters, ranging from 89.45 MB to 5.05 GB per node per round. Since GeoSF transmits no additional statistical information, its communication cost matches that of standard decentralized methods, such as D-PSGD and DFedAvgM, under the same model and communication settings. These results support GeoSF’s computational practicality for large models, provided that device resources are sufficient to support model training.

MetricResNet18ResNet50ViT-BViT-LViT-H
Local training12.94 s23.65 s28.72 s47.29 s73.28 s
Total Computation15.40 s25.82 s32.32 s50.29 s76.96 s
Communication89.45 MB188.65 MB687.68 MB2.43 GB5.05 GB

Thank you for the author's response. Most of my concerns has been solved. I especially like the extra experiments of extra application. I will maintain my original score. Good luck.

Thank you again for your time and valuable suggestions, which have further helped us improve the completeness of our work.

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

This paper proposes a geometry-aware slow-fast learning framework for DFL by accounting for differences in cross-client stability and adaptation speed among model parameters. At the parameter level, GeoSF dynamically partitions parameters into slow and fast groups with distinct optimization and synchronization timescales. Slow parameters use a smaller learning rate and synchronize more frequently, while fast parameters adapt more rapidly and synchronize less frequently. At the client level, GeoSF adjusts consensus weights according to the geometric compatibility between neighboring models. Extensive comparative experiments demonstrate the effectiveness and robustness of the proposed framework.

Critical strengths
  1. The proposed slow-fast learning framework offers a novel approach to handling data heterogeneity. By partitioning parameters according to their cross-client stability and applying distinct optimization and aggregation strategies to each group, GeoSF effectively balances shared knowledge and local adaptation.
  2. The paper provides theoretical analysis and detailed supplementary materials, strengthening the technical foundation of the method and facilitating verification and reproducibility.
  3. Comprehensive experiments across feature-skew and label-skew settings, model architectures, communication topologies, and client scales demonstrate the effectiveness and robustness of GeoSF.
Critical weaknesses
  1. The definition of α\alpha in Figure 2(d) appears to be unclear. The authors should clarify what the horizontal axis represents and explain how changes in the fast and slow learning rates affect model performance. 2.There appear to be a few typos in the paper. For the Ring topology, the main text reports 42.21%, whereas Table 4 reports 49.21%. Additionally, the learning rate for the slow parameters is listed as 0.01 in the appendix, though it may have been intended to be 0.05. Please check and reconcile these values.
  2. Does partitioning parameters into slow and fast groups based on their cross-client stability introduce any additional computational overhead in GeoSF.
  3. When parameter masks differ across clients, the mixing matrix constructed using Equation (9) may not be symmetric. The authors should discuss the potential implications of this asymmetry.
Questions for authors
  1. The authors should clarify how the aggregated parameters are assigned and used to initialize local optimization in Algorithm 1, making the procedure easier to follow.
  2. Including a GeoSF variant with a uniform learning rate would help further illustrate the benefits of using different learning rates for the slow and fast parameters.
  3. It is suggested that the authors discuss any additional computational or communication overhead relative to standard DFL baselines to further assess the practical efficiency of GeoSF.
2 comments

We sincerely thank Reviewer KoCF for carefully evaluating our work, recognizing its contributions, and offering constructive feedback. We respond to each comment and question below.

Weaknesses:

W1: Clarification of the learning rate ratio α\alpha in Fig. 1(d).

A1: Thank you very much for your question. In Fig. 1(d), aa denotes the learning rate ratio. For the personalized parameter curve shown in blue, α=ηpersonalized/ηshared\alpha=\eta_{\mathrm{personalized}}/\eta_{\mathrm{shared}}, with ηshared\eta_{\mathrm{shared}} fixed. For the shared parameter curve, α=ηshared/ηpersonalized\alpha=\eta_{\mathrm{shared}}/\eta_{\mathrm{personalized}}, with ηpersonalized\eta_{\mathrm{personalized}} fixed. The vertical axis reports the corresponding accuracy.

The results generally favor a smaller learning rate for shared parameters than for personalized parameters. This is consistent with slower updates supporting stable consensus and faster updates facilitating local adaptation. We have further clarified this definition in the revised manuscript.

W2: Typographical errors in the manuscript.

A2: Thank you for pointing out these inconsistencies. The accuracy for GeoSF under the Ring topology is 49.21%, as reported in Tab. 3, and the learning rate for the slow parameters is 0.05. We apologize for these typographical errors. We have corrected them in the revised manuscript and checked for consistency throughout.

W3 & Q3: Discussion on the additional computational or communication overhead.

A3: Thank you for your valuable suggestions. Decomposing parameters into “fast” and “slow” components in GeoSF does indeed introduce some additional overhead. To evaluate GeoSF’s practical efficiency, we compared its computational and communication costs with those of standard decentralized federated learning (DFL) baselines, as reported below.

Regarding communication, GeoSF requires only the exchange of model parameters, consistent with D-PSGD and DFedAvgM, without transmitting additional statistical information. Each node transmits an average of 89.45 MB per round, matching both baselines. In terms of computation, compatibility calculation and mask construction introduce modest overhead relative to local training. Taken together, these results demonstrate that, under identical experimental settings, GeoSF achieves higher accuracy with only a modest increase in computational cost and no additional communication burden. Accordingly, we have also included a relevant discussion in the appendix of the revised manuscript.

CompatibilityAggregationComputing MasksLocal Training
0.49 s0.34 s0.70 s12.94 s
MetricD-PSGDDFedAvgMDFedSMDCDFedPGPGeoSF
Communication89.45 MB89.45 MB89.41 MB89.41 MB89.45 MB
Computation14.02 s14.34 s19.45 s14.90 s15.40 s
Accuracy75.55±0.8575.58±0.5675.26±0.1969.64±0.5382.00±0.69

W4: Mixing matrix asymmetry with client-specific masks.

A4: Thank you for raising this important question. In our implementation, we assign identical compatibility values to every pair of clients. Consequently, under Eq. (9), the mixing matrix constructed locally by each client is symmetric, even when clients use different masks. Although these masks may produce different local mixing matrices, each client’s aggregation weights remain nonnegative and sum to one.

Alternatively, using all model parameters to compute distances yields a common symmetric mixing matrix under a shared undirected topology. GeoSF retains strong performance with this alternative, achieving 85.01±0.36 and 81.55±0.54 on Office10. These results suggest that its effectiveness does not depend exclusively on mask-based distance computation.

Questions:

Q1: Initialization of local optimization in Alg. 1.

A1: Thank you very much for your question. Before local optimization, client mm initializes its slow parameters with the aggregated slow parameters, θms,t←θm,gs,t\theta_m^{s,t}\leftarrow\theta_{m,g}^{s,t}. On communication rounds satisfying mod⁡(t,s)=0\operatorname{mod}(t,s)=0, its fast parameters are also initialized with the aggregated fast parameters, θmf,t←θm,gf,t\theta_m^{f,t}\leftarrow\theta_{m,g}^{f,t}. Otherwise, the fast parameters retain their locally trained values from the previous round. Local optimization then proceeds from this initialized model, with θm,gs,t\theta_{m,g}^{s,t} held fixed as the reference for the slow-parameter regularization term.

Q2: GeoSF with a uniform learning rate.

A2: Thank you for your valuable suggestion. We agree that a GeoSF variant with a uniform learning rate would help clarify the benefits of using distinct learning rates for “slow” and “fast” parameters. This variant is evaluated in Fig. 4, where α=1\alpha=1 corresponds to a uniform learning rate and achieves accuracies of 78.39% on Office10 and 72.32% on Cifar10.

Thank you to the author for the response, my main questions have been addressed.

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

The submission presents a decentralized federated learning (DFL) framework named GeoSF (Geometry-aware Slow-Fast Learning), which aims to mitigate performance degradation caused by data heterogeneity across clients. Unlike standard approaches that treat all parameters uniformly, GeoSF identifies that different model parameters exhibit varying levels of cross-client stability and adaptation speeds. At the parameter level, GeoSF dynamically splits model parameters into two categories based on direction consistency and variance: (1) slow parameters, which carry stable, shared global knowledge and receive smaller learning rates and more frequent synchronization, and (2) fast parameters, which capture client-specific distributions and adapt quickly with larger learning rates and periodic synchronization. At the client level, GeoSF discards uniform averaging in favor of a geometry-aware weighted consensus mechanism. By leveraging proximity in the parameter space as a proxy for similarity, it dynamically assigns higher weights to structurally compatible neighboring models, thereby minimizing destructive interference and conflicts during parameter aggregation.

Critical strengths
  1. This submission shifts away from uniform decentralized federated learning rules by jointly optimizing two complementary dimensions: parameter-wise adaptation timescales (slow vs. fast) and client-wise collaboration geometry (graph state compatibility). Instead of relying on static, predefined parameter splits or frozen classification heads, GeoSF introduces a dynamic scoring mechanism that fuses cross-client directional agreement (sign consistency) and magnitude dispersion (variance) to evaluate parameter shareability on the fly.

  2. The authors present solid mathematical backing by establishing convergence guarantee for non-convex objectives over connected networks. The proof links the parameter selection strategy to a tighter convergence bound by bounding the slow-gradient diversity.

  3. The evaluation is thorough, testing the framework against eight strong baselines across both feature-skew and label-skew scenarios, multiple network topologies (Ring, Grid, Exp, Full), and diverse architectures ranging from small CNNs to Vision Transformers (ViT).

Critical weaknesses
  1. The framework introduces a large array of sensitive hyperparameters, including the learning rates, shrinkage coefficient, communication interval, temperature parameter, and the shared-parameter ratio. As illustrated in the sensitivity analysis, choosing sub-optimal parameters may create immediate performance trade-offs.

  2. Calculating peer-wise compatibility matrices and parameter-level masks requires clients to track, communicate, or process neighbor parameters to compute variance and run top-k indices operations over millions of weights. This may introduce an unquantified computational overhead at the edge device level.

  3. A central goal of FL is to protect data privacy. While GeoSF operates in a peer-to-peer fashion, sharing full model parameters to continuously evaluate raw directional sign consistency and magnitude variances across neighbors might introduce security risks, exposing local distribution signatures to inference or reconstruction attacks without defensive mechanisms (like differential privacy).

  4. The theoretical analysis assumes a static, doubly stochastic, and well-connected graph structure. Real-world decentralized peer-to-peer systems are heavily plagued by dynamic topologies, intermittent client dropouts, asymmetric directional links, and message loss, which are not accounted for.

  5. While the authors evaluate scalability up to 100 clients, true decentralized settings often target cross-device scenarios involving thousands of edge components. The empirical scaling limits of calculating large pairwise adjacency matrix columns under huge node counts remain unaddressed.

  6. The submission is in general presented well. But still there are some presentation issues. E.g., page 7, the title of Section 5.2 should remove the full stop; the references are not presented in a consistent way (pay attention on conference and journal names). Also the literature discussion needs to include most recent research papers given that this is a very active research area.

Questions for authors

Please check the details in the Critical Weaknesses section and clarify them.

3 comments

We sincerely thank Reviewer xiAA for carefully reviewing our work and for providing valuable comments and constructive suggestions. We greatly appreciate the opportunity to clarify these points and provide our point-by-point responses below.

Weaknesses:

W1: Hyperparameter sensitivity and performance trade-offs.

A1: Thank you very much for this important question. We agree that inappropriate hyperparameter choices can affect GeoSF’s performance. Our sensitivity analysis quantifies these effects and provides practical guidance for parameter selection. Specifically, the learning rates control the update speeds of fast and slow parameters, while the shrinkage coefficient balances local adaptation and consensus. The communication interval and shared-parameter ratio determine the frequency and extent of aggregation, respectively, and the temperature controls how aggregation weights respond to model differences.

In fact, when set within suitable ranges, these hyperparameters ensure stable performance improvements. For instance, these ranges include μ∈[0.001,0.1]\mu \in [0.001,0.1], a communication interval of s∈[3,30]s \in [3,30], a learning rate ratio of α∈[1,5]\alpha \in [1,5], τ∈[1,20]\tau \in [1,20], and ρ∈[0.1,0.9]\rho \in [0.1,0.9]. Although the optimal configuration may vary depending on the specific scenario, these results demonstrate that GeoSF’s advantages can be realized as long as the hyperparameters remain within reasonable ranges.

W2: Computational overhead of compatibility and mask construction.

A2: Thank you for raising the key point. We agree that computing peer-wise compatibility matrices and parameter-level masks introduces additional client-side overhead. To quantify this overhead, we measured the per-round runtime of individual operations using ResNet50, ViT-Base, ViT-Large, and ViT-Huge, as reported below. Mask construction takes 0.97–1.33 seconds, compared with 23.65–73.28 seconds for local training. Together, mask construction and compatibility computation account for only 3.1%–7.0% of the local training time. These results indicate that GeoSF’s additional computational cost is modest relative to local training under the tested settings. We have included a discussion of GeoSF’s additional computational overhead in the revised manuscript.

MetricResnet50ViT-BViT-LViT-H
Local Training23.65 s28.72 s47.29 s73.28 s
Computing Masks0.97 s1.15 s1.24 s1.33 s
Computing Compatibility0.55 s0.86 s0.73 s0.91 s
Total Runtime25.82 s32.32 s50.29 s76.96 s

W3: Privacy risks associated with model parameter sharing.

A3: Thank you very much for raising this concern. In fact, GeoSF primarily addresses differences in parameter evolution rates and parameter incompatibility caused by data heterogeneity. Like other parameter-exchange methods, GeoSF may indeed introduce potential privacy leakage risks. Compared with FedAvg, it exchanges only model parameters and computes the required statistics locally, without transmitting additional information that could introduce further privacy risks. GeoSF can also incorporate existing privacy-preserving mechanisms. To strengthen privacy protection, we integrated differential privacy (DP) and conducted additional experiments. As shown in the table below, GeoSF achieves an accuracy of 79.22% under the same privacy noise intensity, demonstrating its ability to maintain strong performance under DP protection.

MetricD-PSGDDFedAvgMDFedSMDCDFedPGPGeoSF
DP-Acc68.40±0.9169.32±0.7371.23±0.3365.81±0.3179.22±0.35
Original Acc75.55±0.8575.58±0.5675.26±0.1969.64±0.5382.00±0.69

W4: Theoretical applicability to dynamic and unreliable networks.

A4: Thank you very much for pointing out this important practical issue. We apologize for not explaining the theorem clearly enough before. In fact, since our theoretical analysis assumes λ=max⁡Wt∈W∥Wt−P∥2<1\lambda=\max_{W^t\in\mathcal{W}}\|W^t-P\|_2<1, it encompasses scenarios where the bidirectionally connected topology varies over time. Consequently, the theorem is not restricted to static graphs and remains applicable under intermittent client disconnections or message loss, provided that the stated conditions continue to hold. The theorem remains applicable to asymmetric directed communication links when the associated mixing matrices are doubly stochastic. If double stochasticity fails, the analysis could be extended by incorporating a suitable bias-correction mechanism such as Push-Sum.

W5: Scalability to thousands of clients.

A5: Thank you for raising this valuable question. We agree that constructing and processing dense adjacency matrices can introduce computational and memory burdens as the network grows. In fact, the compatibility scores within the adjacency matrix only need to be computed for connected neighbors. With KK neighbors and dd model parameters, the per-client compatibility computation costs approximately O(Kd)\mathcal{O}(Kd), scaling linearly with the number of neighbors. To quantify this cost, we evaluated per-client adjacency-weight computation using ResNet18 with 50, 100, 500, and 1,000 nodes at a connectivity density of 0.5. As shown below, computation time increases from 0.0939 s to 1.5340 s, while the storage required for neighbor weights in FP32 ranges from 0.10 KB to 2.00 KB per client.

MetricM=50M=100M=500M=1000
Time0.0939 s0.1080 s0.8233 s1.5340 s
Memory0.10 KB0.20 KB1.00 KB2.00 KB

W6: Presentation, reference formatting, and recent literature.

A6: Thank you for your positive feedback and valuable suggestions. In the revised manuscript, we have removed the full stop from the title of Sec. 5.2 and standardized conference and journal names throughout the references. We have also expanded the Related Work section to cover recent advances in decentralized federated learning, including Decentralized Federated Learning with Distributed Aggregation Weight Optimization, Mitigating the Privacy–Utility Trade-off in Decentralized Federated Learning via f-Differential Privacy, and DFLMoE: Decentralized Federated Learning via Mixture of Experts for Medical Data Analysis. These additions provide a more timely discussion and better position GeoSF within the existing literature.

Thanks a lot the authors for the responses to the points raised. I changed my rating of the technical soundness and also corrected my confidence (I am quite familiar with the topic). Good luck.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.