AdaptiveFL: Communication-Adaptive Federated Learning Under Dynamic Bandwidth.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Federated learning (FL) enables collaborative model training across heterogeneous devices.
• Communication bottlenecks are a major challenge in FL, with existing solutions like HeteroFL and LotteryFL using gradient sparsification.
• Current methods do not account for the dynamic, constantly changing bandwidth of individual clients during training.

Purpose of the Study:

• To introduce AdaptiveFL, a novel communication-adaptive federated learning framework.
• To address the challenge of dynamic bandwidth constraints in federated learning environments.
• To enhance the efficiency and performance of federated learning under variable network conditions.

Main Methods:

• AdaptiveFL selects a best-fit sub-model for communication based on the currently available bandwidth of each device in each round.
• It employs a local training method allowing devices to train "tailorable" local models adaptable to any sparsity level with competitive accuracy.
• This approach ensures sub-model performance is maintained despite dynamic bandwidth limitations.

Main Results:

• AdaptiveFL demonstrates superior performance compared to existing state-of-the-art (SOTA) communication-efficient methods.
• The framework effectively manages communication costs by adapting to dynamic bandwidth.
• Sub-models communicated under AdaptiveFL maintain competitive accuracy even with varying sparsity.

Conclusions:

• AdaptiveFL presents a robust solution for communication-efficient federated learning in dynamic bandwidth environments.
• The proposed framework significantly outperforms existing baselines, offering a more practical approach to FL.
• AdaptiveFL paves the way for more efficient and reliable distributed machine learning applications.