EC-DARTS: Inducing Equalized and Consistent Optimization into DARTS

In the following, we provide more details about the Kendall Tau [5] used in our manuscript. In addition, we show the samples of architectures that searched on CI-FAR10/100, Tiny-ImageNet-200, and ImageNet by EC-DARTS. Kendall Tau for Correlation Evaluations To conduct the the correlation evaluation, we apply Kendall Tau [5] to measure correlations between different ranks. Specifically, Kendall Tau between two ranks R 1 and R 2 is denoted as ) can be as high as 1 if R 1 and R 2 have a strong correlation, or as low as -1 if R 1 and R 2 are negatively correlated. Note that as τ (R 1 , R 2 ) approaches 0, there tends to be no correlation between R 1 and R 2 . In our manuscript, we use Kendall Tau to evaluate the correlations of 3 paired ranks: • The search accuracies and the corresponding retraining accuracies of 10 architectures are randomly selected from a single search. It should be noted that it is normal for different epochs to output the same intermediate architectures. • The ranks of the operation weights and retraining accuracy of 10 architectures generated from one search result. Specifically, these 10 architectures are generated by replacing the 2 pairwise edges, that are connected with the last intermediate node in the searched architecture, with different possible combinations. Each possible edge connection only retains the operation corresponding to the largest operation weight. • The search accuracies and the corresponding retraining accuracies of 10 independently searched architectures.