核心概念
提案されたFedU2フレームワークは、FUSLの表現を向上させるために柔軟な均一正則化子(FUR)と効率的な統一アグリゲーター(EUA)から構成されています。
摘要
この記事では、非IIDデータでのFUSLにおける表現の改善を目指すFedU2フレームワークが提案されています。以下は記事の内容の詳細です。
Abstract:
- Federated learning aims to model decentralized data effectively.
- Existing FUSL methods suffer from insufficient representations due to representation collapse and inconsistent representation spaces among local models.
Introduction:
- Privacy regulations drive the need for federated learning in academia and industry.
- The article focuses on federated unsupervised learning (FUSL) with non-IID data, aiming to model unified representation among imbalanced, unlabeled, and decentralized data.
Method:
- FedU2 consists of FUR to prevent representation collapse and EUA to maintain consistent client model updating.
- FUR minimizes unbalanced optimal transport divergence between client data and random samples.
Experiments:
- Performance comparison shows that FedU2 outperforms existing FUSL methods in both cross-silo and cross-device settings on CIFAR10 and CIFAR100 datasets.
- Sensitivity analysis reveals that λU = 0.1 achieves the highest performance.
Conclusion:
- FedU2 enhances uniform and unified representations in FUSL, addressing challenges like representation collapse entanglement.
統計資料
この記事には重要な数字やメトリクスは含まれていません。