Reproduce “Federated Learning for Data Streams” on OpenResearch

About this paper

Othmane Marfoq, Giovanni Neglia, Laetitia Kameni, Richard Vidal

This research provides a formalization and theoretical analysis for federated learning in scenarios where client devices continuously generate data streams under memory constraints. It introduces a meta-algorithm with adaptive sample weighting, demonstrating that a balanced approach to using historical and new data is critical for achieving optimal model accuracy and convergence in dynamic environments.

Key insights

• An inherent 'bias-optimization trade-off' exists in streaming FL, where strategies minimizing generalization error often increase optimization error, and vice-versa.
• The optimal strategy for sample weighting (balancing old vs. new data) is highly dependent on intrinsic problem characteristics, quantified by a ratio of constants.
• Consistently, a 'Fresh' strategy (only using newly arrived data) is suboptimal, emphasizing the importance of judiciously retaining and reusing some older samples.

Summary by alphaXiv.