Appendix for Evaluating Efficient Performance Estimators of Neural Architectures

As the train ing goes on, the pa ram e ter sizes of top - ranked ar chi tec tures by the OSEs be come larger . We visualize the Pareto frontiers discovered by OSEs in the right subplots of the two figures. The blue lines with square markers show the one-shot scores of the GT Pareto frontier, while the orange/green/red lines show the GT scores of the OS Pareto frontier.