Model Pruning Enables Efficient Federated Learning on Edge Devices.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Federated learning (FL) enables private model training on decentralized data from edge devices.
• Edge devices possess limited computational and communication resources, posing challenges for FL.
• Existing FL methods struggle with the resource constraints of edge devices.

Purpose of the Study:

• To introduce PruneFL, a novel federated learning approach using adaptive and distributed parameter pruning.
• To reduce communication and computation overhead in FL by dynamically adapting model size.
• To minimize overall FL training time without compromising model accuracy.

Main Methods:

• PruneFL employs initial pruning on a selected client and further pruning within the FL process.
• Model size is adapted by maximizing the approximate empirical risk reduction per FL round time.
• Experiments were conducted on various datasets using edge devices like Raspberry Pi.

Main Results:

• PruneFL significantly reduced training time compared to conventional FL and other pruning methods.
• The automatically determined pruned model size achieved accuracy comparable to the original model.
• The pruned model demonstrated characteristics of a 'lottery ticket' of the original model.

Conclusions:

• PruneFL effectively addresses the resource limitations of edge devices in federated learning.
• The adaptive pruning strategy optimizes model size for reduced training time and overhead.
• PruneFL offers an efficient and accurate solution for privacy-preserving machine learning on edge devices.