learning on memory-constrained edge devices, which is fundamentally different from existing NAS for efficient

We thank all reviewers for their comments and thank R2 and R3 for suggesting the literature. The presentation will be polished and more discussions will be included in the broader impact section. R1: Novelty compared to Once-for-All. To our best knowledge, we are the first to introduce this finding. We believe our findings will open up new opportunities for on-device learning.