Towards Long-Term Remembering in Federated Continual Learning

Federated Continual Learning (FCL) involves learning from distributed data on edge devices with incremental knowledge. However, current FCL methods struggle to retain long-term memories on the server.

In this paper, we introduce a method called Fisher INformation Accumulation Learning (FINAL) to address catastrophic forgetting in FCL. First, we accumulate a global Fisher with a federated Fisher information matrix formed from clients task by task to remember long-term knowledge. Second, we present a novel multi-node collaborative integration strategy to assemble the federated Fisher, which reveals the task-specific co-importance of parameters among clients. Finally, we raise a Fisher balancing method to combine the global Fisher and federated Fisher, avoiding neglecting new learning or causing catastrophic forgetting.

We conducted evaluations on four FCL datasets, and the findings demonstrate that the proposed FINAL effectively maintains long-term knowledge on the server.

The exceptional performance of this method indicates its significant value for future FCL research.