SUM: Server-side geometric surgery improves federated class incremental learning by 22%
Federated learning meets continual learning without extra client overhead – boosting accuracy by 22%.
Federated Class Incremental Learning (FCIL) combines federated learning and continual learning to allow distributed AI systems to adapt to new tasks over time while respecting data privacy. However, this combination introduces two coupled sources of interference: spatial interference from heterogeneous client data distributions and temporal interference from sequential tasks. This leads to Spatial-Temporal Catastrophic Forgetting (ST-CF). Existing approaches handle these separately, often requiring extra client-side resources.
In a new paper accepted to ECCV 2026, researchers Jaeik Kim and Jaeyoung Do propose SUM (Unified Geometric Surgery on Spatio-Temporal Adaptation Vectors), a server-only solution. SUM reinterprets FCIL as a unified multi-task learning problem, representing client and task updates as adaptation vectors. During aggregation, it applies spatial geometric surgery to reduce client-level interference within each round and causal online temporal surgery to remove cross-task interference over time. This requires no extra client computation, communication, or memory. Empirically, SUM achieves up to 22% accuracy improvement over prior FCIL methods on diverse vision and language benchmarks, and remains robust to unreliable clients.
- SUM is a purely server-side method that eliminates spatial and temporal interference in federated continual learning without extra client overhead.
- It achieves up to 22% accuracy improvement over prior FCIL methods on vision and language benchmarks.
- Accepted to ECCV 2026, the method handles unreliable clients while maintaining computational efficiency.
Why It Matters
Enables scalable distributed AI that adapts to new tasks without forgetting, crucial for edge and privacy-preserving systems.