NAS-Bench-301 and AlphaNet

The dataset used in the paper is NAS-Bench-301 and AlphaNet. NAS-Bench-301 is a surrogate NAS benchmark built via deep ensembles and modeling uncertainty, which provides close-to-real predicted performances (i.e., accuracy and training time) of 10,017 architectures on CIFAR-10. AlphaNet is a new family of architectures on ImageNet discovered by applying a generalized α-divergence to supernet training.