Neural Architecture Search for Energy-Efficient Deep Learning Model Deployment

Abstract

With the increasing adoption of deep learning models across various domains, optimizing their deployment for energy efficiency has become a critical challenge. Neural Architecture Search (NAS) has emerged as a powerful technique to automate the design of deep learning architectures while optimizing trade-offs between accuracy, latency, and energy consumption. This study presents an energy-aware NAS framework that integrates hardwareaware search strategies to generate deep learning models optimized for power efficiency without compromising performance. Our approach utilizes reinforcement learning-based NAS and multi-objective evolutionary algorithms to explore architectures that minimize computational overhead and dynamic power consumption on heterogeneous hardware platforms, including edge devices, cloud infrastructure, and specialized accelerators. Experimental results demonstrate that our optimized deep learning models achieve up to 38% reduction in energy consumption while maintaining comparable accuracy to state-of-the-art architectures. The findings highlight the potential of NAS in facilitating sustainable AI development and energyefficient inference in real-world applications.