FeDMRA: Federated Incremental Learning with Dynamic Memory Replay Allocation

Dimensions overall score 7.0

GitHub Code Pulse

No public code linked for this paper yet.

• FeDMRA, a dynamic memory allocation strategy for exemplar storage, improves model performance in Federated Class-Incremental Learning (FCIL) compared to fixed allocation methods.method 90%
• FeDMRA leverages data heterogeneity among clients to allocate memory, rather than uniformly distributing it across all classes.method 95%
• The method achieves significant performance improvements over existing baseline models on three medical image datasets.result 92%
• FeDMRA introduces a regularization loss term (L_KL) to prevent local models from overfitting and causing global objective drift.technical 88%
• The server in FeDMRA performs weighted aggregation of dual-channel importance indices (data distribution and performance contribution) to determine final client significance for memory allocation.method 85%
• Traditional continual learning methods are inapplicable in federated settings due to non-independent and identically distributed (non-IID) data characteristics.limitation 80%
• FeDMRA's final objective function combines cross-entropy loss, KL divergence loss, and a gradient norm minimization loss (L_MG) for stable training.technical 87%
• Class imbalance between old and new classes is a key challenge faced by exemplar replay methods in Federated Continual Learning (FCL).limitation 75%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card