Running hardware-aware neural architecture search on embedded devices under 512MB of RAM

The article presents a novel hardware-aware neural architecture search (HW NAS) method designed for embedded devices with less than 512MB of RAM, enabling the generation of compact convolutional neural networks (CNNs) suitable for low-end microcontroller units (MCUs). This approach achieves state-of-the-art performance in human-recognition tasks on the Visual Wake Word dataset, a benchmark for TinyML applications, allowing for on-device model customization while ensuring data privacy. This development is significant for practitioners focusing on deploying efficient AI solutions in resource-constrained environments typical of IoT and wearable robotics.