An evolution-based approach for efficient differentiable architecture search
Masayuki Kobayashi, Tomoharu Nagao · 2020
We propose a hybrid neural architecture search method of evolution-based and gradient-based approaches. In this method, the architecture variables are optimized by evolutionary algorithm and exploited by gradient-based optimization. This hybridity allows an efficient architecture search and also maintains the extensibility of the evolution-based search. On the CIFAR-10 dataset, our method discovers architectures that achieve competitive performance with state-of-the-art models.