PrototypeNAS: Rapid Design of Deep Neural Networks for Microcontroller Units

PrototypeNAS is a zero-shot neural architecture search (NAS) method designed to optimize deep neural networks (DNNs) for microcontroller units (MCUs) without the need for extensive training of multiple models. It employs a three-step search strategy that integrates structural optimization of various architectures, utilizes an ensemble of zero-shot proxies for optimization, and applies Hypervolume subset selection for effective model distillation. This approach enables rapid identification of compact DNN models that maintain high accuracy, making it significant for practitioners aiming to deploy efficient AI solutions on resource-constrained edge devices.