Energy-Aware Integrated Neural Architecture Search and Partitioning for Distributed Internet of Things (IoT)

Internet of Things (IoT) are one of the key enablers of personalized health. However, IoT devices often have stringent constraints in terms of resources, e.g., energy budget, and, therefore, limited possibilities to exploit the state-of-the-art Deep Neural Networks (DNNs). Energy-aware Neural Architecture Search (NAS) is proposed to tackle this challenge, by exploring lightweight DNN (DNN) architectures on a single IoT device, but not leveraging the inherently distributed nature of IoT systems. As a result, the joint optimization of DNN architectures and DNN computation partitioning/offloading has not been addressed to date. In this paper, we propose an energy-aware NAS framework for distributed IoT, aiming to search for distributed Deep Neural Networks (DNNs) to maximize prediction performance subjected to Flash Memory (Flash), Random Access Memory (RAM), and energy constraints. Our framework searches for lightweight DNN architecture with optimized prediction performance and its corresponding optimal computation partitioning to offload the partial DNN from edge to fog in a joint optimization. We evaluate our framework in the context of two common health applications, namely, seizure detection and arrhythmia classification, and demonstrate the effectiveness of our proposed joint optimization framework compared to NAS benchmarks.