From Selection to Scheduling: Federated Geometry-Aware Correction Makes Exemplar Replay Work Better under Continual Dynamic Heterogeneity

Researchers propose FEAT, a federated learning method that improves continual learning by addressing class imbalance and representation collapse across distributed clients. The approach combines geometric alignment and energy-based correction to better utilize exemplar samples while maintaining performance under dynamic heterogeneity.

FEAT addresses a fundamental challenge in federated continual learning where models must learn new tasks without forgetting previous ones while managing heterogeneous data distributions across decentralized clients. The research tackles an overlooked problem: while existing methods focus on selecting which samples to replay, they neglect how to effectively use those samples when data is imbalanced and non-independent across clients.

The solution employs two complementary mechanisms. The Geometric Structure Alignment module uses Equiangular Tight Frame prototypes as class-discriminative references, ensuring consistent feature geometry across tasks and clients. This prevents minority class features from being pulled toward majority classes during distributed training. The Energy-based Geometric Correction module then refines embeddings by removing task-irrelevant directional components, reducing majority-class bias while improving minority-class sensitivity.

This work advances federated learning infrastructure, which has growing applications in healthcare, finance, and IoT systems where data privacy and decentralization are critical. Improved continual learning mechanisms directly benefit these sectors by enabling systems to adapt to new patterns while retaining learned knowledge across distributed networks.

The importance lies in bridging theory and practice: federated continual learning remains largely unsolved for real-world scenarios with class imbalance. FEAT's geometric approach offers a principled way to handle representation collapse, a pervasive problem in distributed machine learning. Future work should validate performance on larger-scale federated settings and compare against emerging federated-learning frameworks. The method's reliance on shared prototypes across clients also raises questions about privacy implications in sensitive applications.

• →FEAT mitigates catastrophic forgetting in federated continual learning by addressing class imbalance and representation collapse.
• →Geometric Structure Alignment module uses shared Equiangular Tight Frame prototypes to maintain consistent feature geometry across distributed clients.
• →Energy-based Geometric Correction removes task-irrelevant components to reduce bias toward majority classes and improve minority-class sensitivity.
• →The approach improves upon prior exemplar replay methods by focusing on how to use selected samples rather than just selecting them.
• →Research advances federated learning applications in privacy-sensitive domains like healthcare, finance, and IoT systems.