Towards Robust Federated Learning under Label Scarcity

252 / BBB

2026-06-02

David Jin

15:45

Machine learning models are only as robust as the quality and quantity of their training data. In practice, data is often distributed across devices or institutions and cannot be shared directly. Federated learning enables collaborative training without centralizing raw data, allowing models to benefit from diverse decentralized datasets.

Labeled data, however, often remain limited and expensive to obtain, making label-efficient methods essential. Active learning addresses this by iteratively selecting only the most informative samples for annotation.

Combining active learning with federated learning introduces new challenges. Data is heterogeneous, confidential, and distributed across clients, making it difficult to identify informative samples at a global level. Active learning methods must therefore be carefully designed to account for these constraints.

In this presentation, I will explain how active learning can be integrated into the federated learning pipeline. I will present empirical results comparing local client-based and coordinated federated active learning strategies and explain the trade-offs between them. Building on these insights, I will discuss how combining both approaches can lead to more label efficiency across diverse federated learning settings.