Neural Architecture Search For Efficient Model Deployment On Mobile Devices

Abstract

Neural Architecture Search (NAS) has transformed the automation of deep learning model design, enabling the discovery of optimized architectures tailored for specific hardware constraints. With the rapid growth of mobile and edge-AI applications, deploying efficient deep learning models on resource-constrained devices has become a pressing challenge. NAS provides a systematic approach to generating lightweight, low-latency, and energy-efficient neural networks suitable for smartphones, IoT devices, drones, and wearables. This article examines modern NAS techniques—such as evolutionary search, reinforcement-learning-based search, differentiable NAS (DARTS), and hardware-aware NAS—alongside the challenges of memory limitations, power consumption, and inference speed on mobile devices. Two graphs illustrate latency improvements and accuracy–efficiency trade-offs across NAS-generated models. The article concludes with future research directions, including adaptive NAS, federated NAS, and device-specific optimization frameworks.