1 Photos and videos
Mar 6
Had a lot of fun recreating the 13yr old DeepMind paper on playing Atari with deep RL entirely from scratch!
Here, an agent learns from only raw video frames to match human experts in 29/49 diff. games using the same hyperparameters! The most interesting thing revisiting is how the authors like to motivate experience replay and DQN from a neuroscientific/bio perspective, even though--
2
6
432
Mar 6
-- both choices seem done entirely from a DL perspective.
Namely, experience replay is linked to the idea that recent human experiences are reaactivated during waking rest/sleep perhaps for memory consolidation.
Also, a Q-network that doesn't separate feature extraction(CNN) from value estimation draws on claim that reward signals not only influence the actions you take but your internal representations of visual perception-why you might unconsciously focus on certain things over others.
3
201