Differentiable Architecture Search Using Stable Rank of Convolutional Layers

Abstract

We propose Neural Architecture Search method MSR-DARTS, which uses the stable rank of a convolutional layer for selecting the optimal network structure. We evaluate MSR-DARTS on CIFAR10 and ImageNet dataset. It achieves an error rate of 2.80% on CIFAR10 and a top-1 error rate of 24.0% of ImageNet.