PrototypeNAS: Rapid Design of Deep Neural Networks for Microcontroller Units
The episode discusses PrototypeNAS, a zero-shot Neural Architecture Search framework for rapidly designing Deep Neural Networks for resource-constrained microcontroller units (MCUs). The hosts detail its methodology, which involves creating a unified search space combining architecture and optimization settings, using an ensemble of competing zero-shot proxies in multi-objective optimization, and employing Hypervolume Subset Selection to distill the results into a small set of deployable models.