Research map / Federated learning and privacy / Personalized and heterogeneous FL
Improving Model Fusion by Training-Time Neuron Alignment With Fixed Neuron Anchors
Zexi Li, Zhiqi Li, Jie Lin, Tao Shen, Jun Xiao, Yike Guo, Tao Lin, Chao Wu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add TNA-PFN | vanilla training · CIFAR-10 CNN 5-model fusion | interpolated accuracy: 21.3 -> 48.3 | Improved |
| add TNA-PFN | vanilla training · MNIST MLP_h5_w200 5-model fusion | accuracy barrier: 0.643 -> 0.0974 | Improved |
| add TNA-PFN | pretrained ResNet18 · CIFAR-10 LMC | interpolated accuracy: 11.9 -> 54.9 | Improved |
| add TNA-PFN | semi-trained CNN · CIFAR-10 LMC | accuracy barrier: 0.181 -> 0.0413 | Improved |
| add TNA-PFN | greedy soup of CLIP ViT-B/32 · CIFAR-100 different seeds | accuracy: 67.7 -> 73.3 | Improved |
| add TNA-PFN | uniform soup of CLIP ViT-B/32 · CIFAR-10 different learning rates | accuracy: 91.6 -> 91.2 | Hurt |
| add TNA-PFN | mixed model soups · CIFAR-10 / CIFAR-100 | accuracy: CIFAR-100 greedy 80.9 -> 82.6; CIFAR-10 uniform 91.6 -> 91.2 | Mixed |
| add TNA-PFN to FedAvg (ColD fusion) | FedAvg · GLUE COLA+RTE+MRPC Sum | accuracy: 77.66 -> 81.43 | Improved |
| use FedPFN | FedAvg · CIFAR-10 CNN dir 0.1 | top-1 accuracy: 57.5 -> 62.2 | Improved |
| use FedPNU | FedAvg · CIFAR-100 ResNet20 dir 0.1 | top-1 accuracy: 28.5 -> 35.3 | Improved |
| add FedPFN (resample mask each round) instead of fixing the mask (ablation) | FedPFN · CIFAR-10 CNN dir 0.3 | top-1 accuracy: 64.9 -> 65.7 | Improved |
| use gradient-mask FedPFN instead of random pruning at init (ablation) | FedPruning · CIFAR-10 CNN dir 0.3 | top-1 accuracy: 63.7 -> 65.7 | Improved |
Closest papers by meaning
- Model Fusion via Retrofitting (ICML 2026 · similarity 0.88)
- Deep Neural Network Fusion via Graph Matching with Applications to Model Ensemble and Federated Learning (ICML 2022 · similarity 0.86)
- Towards Meta-Pruning via Optimal Transport (ICLR 2024 · similarity 0.84)
- Early Neuron Alignment in Two-layer ReLU Networks with Small Initialization (ICLR 2024 · similarity 0.83)
- Beyond the Permutation Symmetry of Transformers: The Role of Rotation for Model Fusion (ICML 2025 · similarity 0.83)
- Transformer Fusion with Optimal Transport (ICLR 2024 · similarity 0.83)
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.
