Research map / Federated learning and privacy / Communication-efficient FL
Group Knowledge Transfer: Federated Learning of Large CNNs at the Edge
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace FedAvg with FedGKT using ResNet-8 edges and a ResNet-56 server | federated image classification · CIFAR-10, CIFAR-100, and CINIC-10, I.I.D. and non-I.I.D. | top-1 accuracy: 92.88 -> 92.97; 86.60 -> 86.59; 68.09 -> 69.57; 63.78 -> 63.76; 81.62 -> 81.51; 77.85 -> 77.80 | Mixed |
| replace FedAvg with FedGKT using ResNet-8 edges and a ResNet-110 server | federated image classification · CIFAR-10, CIFAR-100, and CINIC-10, I.I.D. and non-I.I.D. | top-1 accuracy: 93.49 -> 93.47; 87.20 -> 87.18; 68.58 -> 69.87; 64.35 -> 64.31; 82.10 -> 81.98; 78.43 -> 78.39 | Mixed |
| replace FedAvg's large edge CNN with compact ResNet-8 and transfer knowledge between edge and server | edge training · CIFAR-100 | edge FLOPs: 9 times less than that of ResNet-56 and 17 times less than that of ResNet-110 | Helped |
| replace FedAvg's large edge CNN with compact ResNet-8 | edge training · CIFAR-100 | edge model parameters: 54 times less than that of ResNet-56 and 105 times less than that of ResNet-110 | Helped |
| remove knowledge-distillation losses from both client and server (ablation) | FedGKT · CIFAR-10, CIFAR-100, and CINIC-10 | top-1 accuracy: -/diverge | Hurt |
| add server-to-edge knowledge transfer (ablation) | FedGKT · CIFAR-10 and CINIC-10 | top-1 accuracy: 92.97; 81.51 | Helped |
| add bidirectional server-edge knowledge transfer (ablation) | FedGKT · CIFAR-100 | top-1 accuracy: 68.44 -> 69.57 | Helped |
| replace synchronous with asynchronous training (ablation) | FedGKT · CIFAR-10, CIFAR-100, and CINIC-10 | top-1 accuracy: 92.97 -> 92.92; 69.57 -> 69.65; 81.51 -> 81.43 | No clear effect |
| increase edge count from 16 to 8, 64, or 128 (ablation) | FedGKT · CIFAR-100 | top-1 accuracy: 69.57 -> 69.51; 69.57 -> 69.65; 69.57 -> 69.59 | No clear effect |
| replace ResNet-8 edge model with a smaller ResNet-4 or ResNet-6 (ablation) | FedGKT · CIFAR-10 | test accuracy: 92.97 -> 88.86; 92.97 -> 90.32 | Hurt |
Closest papers by meaning
- Device-Wise Federated Network Pruning (CVPR 2024 · similarity 0.86)
- FedLPS: Heterogeneous Federated Learning for Multiple Tasks with Local Parameter Sharing (AAAI 2024 · similarity 0.86)
- Beyond Local Sharpness: Communication-Efficient Global Sharpness-aware Minimization for Federated Learning (CVPR 2025 · similarity 0.85)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.85)
- Widening the Network Mitigates the Impact of Data Heterogeneity on FedAvg (ICML 2025 · similarity 0.85)
- Federated Learning for Feature Generalization with Convex Constraints (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.