On-Device NAS cuts RAM by 37% while boosting accuracy 6 points

A team led by Andrea Mattia Garavagno at the University of Genoa has developed a novel approach to near-sensor computing that performs lightweight Neural Architecture Search (NAS) directly on the deployment device. Their work, published on arXiv, addresses a critical challenge in human-machine interfaces: biometric data varies significantly between users, reducing accuracy of static neural networks. By running NAS on-device, the network can be re-optimized for each new user after a brief guided data collection, fighting inter-individual variability without sending data to the cloud.

The proposed NAS was validated on two embedded systems using the Italian Sign Language (ISL) dataset of surface EMG signals and the Case Western Reserve University (CWRU) fault diagnosis benchmark. On a Raspberry Pi 4, the method discovered tiny architectures that outperformed state-of-the-art: 0.63x lower RAM occupancy (37% reduction) with 5.96 percentage points higher accuracy on ISL, and 0.44x lower RAM (56% reduction) with 0.2 points higher accuracy on CWRU. The approach opens possibilities for adaptive, privacy-preserving edge AI in prosthetics, sign language translation, and industrial monitoring.