1: \begin{abstract}

2:    In neural architecture search (NAS), differentiable architecture search (DARTS) has recently attracted much attention due to its high efficiency.

3:    It defines an over-parameterized network with mixed edges, each of which represents all operator candidates, and jointly optimizes the weights of the network and its architecture in an alternating manner.

4:    However, this method finds a model with the weights converging faster than the others, and such a model with fastest convergence often leads to overfitting.

5:    Accordingly, the resulting model cannot always be well-generalized.

6:    To overcome this problem, we propose a method called minimum stable rank DARTS (MSR-DARTS), for finding a model with the {\it best generalization error} by replacing architecture optimization with the selection process using the minimum stable rank criterion.

7:    Specifically, a convolution operator is represented by a matrix, and MSR-DARTS selects the one with the smallest stable rank. 

8:    We evaluated MSR-DARTS on CIFAR-10 and ImageNet datasets. 

9:    It achieves an error rate of $2.54\%$ with $4.0$M parameters within $0.3$ GPU-days on CIFAR-10, and a top-1 error rate of $23.9\%$ on ImageNet. 

10:    The official code is available at \url{https://github.com/mtaecchhi/msrdarts.git}.

11: \end{abstract}

12: