Abstract
Neural architecture search (NAS) is an important research topic of automated machine learning, which aims to automatically search for neural network architectures that can efficiently learn for a given task. Furthermore, there is an increasing demand for deploying computationally efficient NAS systems that can be used on edge devices that have limited computational powers. In particular, federated learning is an online distributed machine learning scheme that requires online and federated search of neural architectures involving edge devices. However, most exiting NAS methods are naturally not well suited for distributed real-time systems, due to high computational and communication costs. This chapter provides a brief introduction to methods for reducing computational cost in NAS, followed by a presentation of two evolutionary frameworks we recently developed for real-time federated NAS.