A Review of Evolutionary Deep Neural Architecture Search with Performance Predictor

Abstract

Automated design of deep neural networks using performance predictor has become a hot topic in current research. Neural architecture search (NAS) methods can be used to enable automatic design of neural network structures by defining different search spaces, search strategies, or optimization strategies. Evolutionary computation by many researchers as the search strategy for NAS, which is called evolutionary NAS (ENAS) . However, ENAS is time-consuming in evaluating the performance of network structures, which hinders the development of ENAS. Therefore, predicting network architecture performance using performance predictor can improve ENAS search speed and save computational resources. This paper summarizes several ENAS methods that utilize performance predictor, and discusses an outlook on search space, search strategies, and the future directions of ENAS assisted by performance predictor.