On-Device NAS Optimizes Tiny Neural Architectures

Researchers have proposed an innovative method for near-sensor computing, enabling Neural Architecture Search (NAS) to be executed directly on the deployment device. This on-device NAS aims to identify the most efficient and compact neural architecture specifically tailored for analyzing real-time sensor data. The approach is particularly beneficial for applications like human-machine interfaces, where neural networks can be dynamically redesigned for individual users, effectively mitigating inter-individual data variations. A new lightweight NAS was designed and rigorously validated using the Italian Sign Language (ISL) dataset, which comprises surface electromyography (sEMG) signals, and the Case Western Reserve University (CWRU) dataset for intelligent fault diagnosis. When tested on a Raspberry Pi 4, the proposed NAS demonstrated superior performance, achieving 0.63 times less RAM occupancy and 5.96 percentage points higher accuracy for the ISL dataset, and 0.44 times less RAM and 0.2 percentage points higher accuracy for the CWRU dataset, surpassing existing state-of-the-art solutions.