Federated continual learning with joint diffusion-based generative replay

doi:10.1016/j.asoc.2026.115928

Federated continual learning with joint diffusion-based generative replay

2026-07-09

0

PRE

AI

DOI：10.1016/j.asoc.2026.115928

Original

Original request for help

Save

Abstract

En 中文

• FedJoin addresses catastrophic forgetting by jointly modeling image generation and classification. • Dual-state replay mechanism generates both long-term and short-term samples to handle imbalanced task distributions. • Designed feature consistency constraints align new task features with historical class embeddings to preserve prior knowledge. • Compared to existing replay methods, the average forgetting is reduced by 7.78%.

Federated continual learning with joint diffusion-based generative replay

Abstract

Journal

Organization