Energy-Adaptive Multi-Dimensional Learning Control for Federated Learning in Energy-Harvesting AIoT Systems.

Area of Science:

• Artificial Intelligence of Things (AIoT)
• Machine Learning
• Energy Harvesting Systems

Background:

• Energy-harvesting AIoT systems face challenges with unstable energy availability, leading to device blackouts and unreliable federated learning.
• Time-varying energy can disrupt the training process, impacting the overall performance and efficiency of AIoT devices.

Purpose of the Study:

• To develop an energy-adaptive framework for efficient federated learning in AIoT systems with fluctuating energy supplies.
• To ensure stable learning performance and prevent device blackouts by intelligently managing resources.

Main Methods:

• Proposing an energy-adaptive multi-dimensional learning control framework.
• Jointly determining model complexity and training intensity based on real-time energy status.
• Integrating model pruning, quantization, knowledge distillation, and adaptive local training into a unified decision mechanism.

Main Results:

• The framework significantly reduces device blackouts in solar-energy-harvesting AIoT environments.
• Maintained competitive model accuracy compared to energy-unconstrained scenarios.
• Demonstrated the effectiveness of joint control over multiple learning-cost factors.

Conclusions:

• Joint control of learning-cost factors is crucial for stable and efficient federated learning in energy-harvesting AIoT.
• The proposed framework enables effective federated learning despite energy constraints.
• The method ensures learning effectiveness while preventing energy depletion on devices.