Deep Reinforcement Learning: Pong from Pixels. As promised, here’s a bit of a flashback on reinforcement learning, a neat older result on using reinforcement learning to train a network to play Atari video games. It’s important to recognize in this work, too, just like with the AlphaGo Zero work, that the resulting network does not understand what it’s doing. It can’t explain the rules, doesn’t have any abstractions. It’s just very, very, very good at pattern recognition.