Research map / Federated learning and privacy / Communication-efficient FL
Bayesian Coreset Optimization for Personalized Federated Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CoreSet-PFedBayes, which learns per-client Bayesian coreset weights to train personalized federated Bayesian neural networks on small weighted subsets, with theory on generalization error and experiments on MNIST, FashionMNIST, CIFAR-10 and medical datasets.
Problem
Federated learning clients must train on their full local datasets, which is computationally expensive; the paper asks how to select a small, representative weighted coreset per client without losing personalized accuracy.
Approach
Assigns each client a sparse coreset weight vector optimized (via accelerated iterative thresholding) to match both the client's variational posterior and the coreset-weighted likelihood to the full-data ones, and plugs these weighted subsets into a PFedBayes-style federated Bayesian training loop.
Key results
- Reported gains of CoreSet-PFedBayes over PFedBayes with 50% random sampling of +4.87% on MNIST, +8.61% on FashionMNIST and +9.71% on CIFAR (as stated in the text).
- On COVID-19 medical data the method is reported at +9.74% over submodular-selection baselines (per the abstract), though Table 2 shows mixed per-class results.
- Theory claims the generalization error of the coreset-weighted objective converges at rate O(n_k^{-2β/(2β+d)} log^{2δ'} n_k), described as minimax optimal up to logarithms.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use personalized coreset selection instead of random 50% training-data sampling | PFedBayes · MNIST, FashionMNIST, and CIFAR | accuracy: +4.87\% on MNIST; +8.61\% on FashionMNIST; +9.71\% on CIFAR | Mixed |
| Use coreset-selected 50% of client data instead of the full dataset | Federated learning methods · MNIST, FashionMNIST, and CIFAR | accuracy | Mixed |
| Use model-centric coreset selection rather than submodular subset selection | FedAvg · COVID-19 Radiography Database, APTOS 2019 Blindness Detection, and OCTMNIST | classwise accuracy: +9.74\% under COVID-19 dataset | Mixed |
Surprising findings
- In the main CIFAR table, CoreSet-PFedBayes (50%) personal accuracy (69.66) is below full-data PFedBayes (83.46), and the global accuracy on CIFAR (71.5) is below FedAvg full (79.05).
- Table 1 reports dashes for several baselines due to very slow convergence, and the random-subset baseline is at times better than the proposed method on FashionMNIST global accuracy (90.75 vs 92.7 reported differently across rows).
Limitations
- Theoretical results rely on strong assumptions (equal-width networks, bounded parameters, Hölder-smooth functions) and the drift terms are argued heuristically with some sign assumptions.
- Experiments are small-scale (10 clients, 2 clients for medical), with the claimed gains sometimes inconsistent with the tables and the method acknowledged as computationally expensive.
- Baseline comparisons use different settings (full vs 50%) and several baselines lack reported results.
Experimental setup
Datasets: MNIST FashionMNIST CIFAR-10 COVID-19 Radiography Database APTOS 2019 Blindness Detection OCTMNIST Housing Prices 2018
Benchmarks: MNIST FashionMNIST CIFAR-10 COVID-19 Radiography Database APTOS 2019 Blindness Detection OCTMNIST
Metrics: personal model accuracy global model accuracy classwise accuracy KL divergence over epochs
Baselines: FedAvg BNFed pFedMe perFedAvg PFedBayes RandomSubset LogDet FedAvg DispSum FedAvg DispMin FedAvg Independent Learning
Builds on: PFedBayes Bayesian coresets accelerated iterative hard thresholding FedAvg
Closest papers by meaning
- FedAS: Bridging Inconsistency in Personalized Federated Learning (CVPR 2024 · similarity 0.87)
- Understanding the Statistical Accuracy-Communication Trade-off in Personalized Federated Learning with Minimax Guarantees (ICML 2025 · similarity 0.86)
- Improved Coresets for Vertical Federated Learning: Regularized Linear and Logistic Regressions (ICML 2025 · similarity 0.86)
- pFedClub: Controllable Heterogeneous Model Aggregation for Personalized Federated Learning (NeurIPS 2024 · similarity 0.86)
- PFedCS: A Personalized Federated Learning Method for Enhancing Collaboration among Similar Classifiers (AAAI 2025 · similarity 0.86)
- Harnessing Heterogeneous Statistical Strength for Personalized Federated Learning via Hierarchical Bayesian Inference (ICML 2025 · similarity 0.86)
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.