Task-Agnostic Federated Continual Learning via Replay-Free Gradient Projection

The introduction of FedProTIP marks a significant step forward in federated continual learning (FCL), a field that faces challenges like catastrophic forgetting and data heterogeneity. By utilizing a novel projection technique, FedProTIP mitigates the interference of new tasks on previously learned ones, thereby preserving model performance. This framework not only addresses the pressing issue of task-agnostic inference but also incorporates a mechanism for predicting task identities, allowing for dynamic adjustments to the global model's outputs. Extensive experiments across standard FCL benchmarks have demonstrated that FedProTIP outperforms state-of-the-art methods in average accuracy, particularly in scenarios where task identities are not predefined. This advancement is vital for the future of decentralized machine learning, as it enhances the ability of distributed systems to learn from evolving data streams without compromising on privacy or communication efficiency.