Основні поняття
Proposing FedFixer to address heterogeneous label noise challenges in Federated Learning.
Анотація
Abstract:
FL relies on label quality but faces noisy and heterogeneous labels.
FedFixer introduces a personalized model to select clean client-specific samples effectively.
Introduction:
FL struggles with heterogeneous label noise, impacting generalization performance.
Existing methods face challenges due to limited sample sizes and struggle with noisy data.
Dual Model Structure Approach:
FedFixer uses global and personalized models for effective learning in the presence of label noise.
Regularizers like Confidence Regularizer and Distance Regularizer mitigate overfitting issues.
Experiments:
Conducted experiments on benchmark datasets showcasing FedFixer's effectiveness in filtering noisy labels.
Comparison with State-of-the-Art Methods:
FedFixer outperforms other methods across different settings, especially in highly noisy scenarios.
Ablation Study:
Components like Confidence Regularizer, Distance Regularizer, Alternate Updates, and Personalized Model contribute to performance improvement.
Статистика
ベンチマークデータセットでの実験により、FedFixerはノイズラベルを効果的にフィルタリングすることが示されました。
Цитати
"FedFixer can perform well in filtering noisy label samples on different clients."
"Existing methods designed for the issue of label noise in FL can be broadly categorized into coarse-grained and fine-grained methods."