Federated continual learning (FCL) allows each client to continually update its knowledge from task streams.
FCL needs to address spatial data heterogeneity between clients and temporal data heterogeneity between tasks.
The proposed Federated Tail Anchor (FedTA) overcomes parameter-forgetting and output-forgetting using trainable Tail Anchor and frozen output features.
FedTA also includes Input Enhancement, Selective Input Knowledge Fusion, and Best Global Prototype Selection for improved performance in downstream tasks.