Energy-efficient management of Artificial Intelligence applications for smart eyewears with tensor quantization.

Smart eyewear (SEW) has evolved beyond medical applications, integrating artificial intelligence (AI) for enhanced functionality. However, deploying deep neural networks (DNNs) on SEW is challenging due to hardware constraints such as limited memory, processing power, and battery life. While task offloading to edge and cloud resources alleviates computational burdens, data transfer overhead remains a major issue, consuming in some cases over 50% of total energy. This paper introduces a Reinforcement Learning (RL)-based tensor quantization strategy to reduce data transfer size, improving both energy efficiency and execution time. A Deep Q-Network (DQN) agent dynamically adjusts quantization levels based on system conditions, balancing accuracy with energy consumption. Experimental results show a 55% reduction in energy consumption while maintaining execution time violations below 1.1%, with only 7.2% of accuracy loss, significantly outperforming non-quantized approaches. These findings make tensor quantization a promising approach for optimizing AI applications on SEW.