Specifically, forthesearchspacesandtasks, we use NAS-Bench-101 (CIFAR-10), NAS-Bench-201 (CIFAR-10, CIFAR-100, and ImageNet16-120), NAS-Bench-301 (CIFAR-10), and TransN AS-Bench-101 Microand Macro (Jigsaw, Object Classification, Scene Classification, Autoencoder) from NAS-Bench-Suite. Weconsiderall 44Karchitecturesreferencedin Table 2. See Table 3 and Appendix Dforthefullresults.

Neural Information Processing Systems http://nips.cc/

Wefirstprovethat"delayed"Exp3achievesthe O p (KT +D)lnK regret bound conjectured by Cesa-Bianchi et al. [2019] in the case of variable, but bounded delays. Here,K is the number of actions andD isthe total delay overT rounds.

This is implemented using a variant of the transformer architecture that ingests optical flow, where each query vector specifies an object and its layer for the entire video.