SUM: Unified Geometric Surgery on Spatio-Temporal Adaptation Vectors for Federated Class Incremental Learning

Short summary

SUM proposes a server-side framework for Federated Class Incremental Learning that addresses spatial-temporal catastrophic forgetting by performing geometric surgery on adaptation vectors during aggregation. It unifies client and task updates as vectors in shared parameter space, mitigating interference without extra client-side computation. Empirically achieves up to 22% improvement over prior FCIL methods across vision and language benchmarks.

• •SUM tackles spatial-temporal catastrophic forgetting in federated incremental learning via server-side geometric surgery
• •No additional client-side computation, communication, or memory beyond standard federated training
• •Up to 22% improvement over prior FCIL methods on diverse benchmarks

Generated with AI, which can make mistakes.