Bootstrap Your Own Latent (BYOL) learns useful representations without explicit negative examples.
Our researchers show that it also works without using batch statistics at all, challenging the recent "implicit negatives from BatchNorm" hypothesis: https://t.co/SMDyTiOHTC
Ape-X DQfD completes the first level on Montezuma's Revenge, which has proven challenging for deep RL. It extends DQN using demonstrations and modifications to handle high discount factors and varying reward distributions. https://t.co/dfCvCpyeXB
Video: https://t.co/G5glodYsHK