Improving Efficient Neural Architecture Search Using Out-net

Abstract

∗Over the past years, there are many achievements in neural networks architecture design. The artificial neural architecture search (NAS) becomes a new way to find good architecture. Architecture searching with parameters sharing proposed by Google greatly decrease training time. However, it brings other problems like overfitting and unfair performance evaluation introduced by parameters sharing. To solve these problems, we propose a mechanism that using out-net to help training parameters, and select the best model from several candidate models produced by the controller. Experiments show that our method has a better performance when searching a small network, which got 77.3% accuracy on cifar100 with a lower latency.