A Survey on One-shot Neural Architecture Search
Xiao Lan Yao, Yahui Qiu, Xiaofan Li · IOP Conference Series Materials Science and Engineering · 2020
Abstract Complex deep neural network architecture such as AlexNet has great success in image classification, natural language processing and other applications. The choice of network architecture has proven to be critical. And researchers begin to design more and more complex networks in order to obtain better performance. However, designing complex networks manually often consumes a lot of computing resources and time, so the automated Neural Architecture Search(NAS) has attracted more and more attention in recent years. Currently, NAS have shown great potential in designing new architectures with high performance and high efficiency. Thus, first of all, this paper briefly introduces the development context and different directions of NAS. But compared to other NAS surveys, this paper focuses on One-shot-NAS methods. Among different methods of NAS, One-shot-NAS can achieve better performance in less time. Through comprehensive comparison and analysis, we apply a novel categorization on One-shot-NAS methods that are based on DAG, Hypernetwork, and network transformation respectively.