Federated Averaging with Replay-Based Incremental Learning for Distributed Neural Network Training

Abstract

Federated Learning (FL) enables collaborative training of neural networks across decentralized clients without exposing raw data, while Continual (incremental) Learning aims to let a deployed model absorb new knowledge without discarding what it has already learned. In real deployments these two requirements co-occur: a global model produced by Federated Averaging (FedAvg) must later be updated as new, non-stationary data streams arrive at the server, which exposes it to catastrophic forgetting. This paper presents and evaluates a lightweight framework that combines (i) a FedAvg-based federated training module, (ii) a replay-memory incremental learner that mixes new-task samples with a bounded buffer of past samples, and (iii) a robust continuous-learning controller that monitors validation accuracy on a held-out reference set and rolls back weight updates that degrade previously acquired knowledge beyond a safety threshold. We describe the system architecture, implementation, and an empirical study on a three-task synthetic benchmark that mimics heterogeneous client data and sequential concept drift. The results indicate that replay-based updating substantially reduces the accuracy degradation on previously learned tasks relative to naive sequential fine-tuning, at a modest cost in plasticity toward more complex new tasks. We discuss these findings in relation to the wider literature on Federated Continual Learning (FCL) and outline directions for scaling the approach to non-synthetic, non-IID, cross-device settings.