Our work shows that adding geometric inductive biases in neural nets enables spatio-temporally consistent (hundreds of steps) object keypoints. This enables agents that play Atari games on a single machine with less than 100k steps + deeply explore hard envs without rewards.
Our work shows that adding geometric inductive biases in neural nets enables spatio-temporally consistent (hundreds of steps) object keypoints. This enables agents that play Atari games on a single machine with less than 100k steps + deeply explore hard envs without rewards.
@tejasdkulkarni I love you. Now please apply this to Starcraft 2 AI, we’re eagerly waiting to watch the transporter create representations of probes, scvs and utilize those to great effect.
@tejasdkulkarni Could you elaborate on what is “geometric” about the inductive bias? 1) if that’s asked of the audience to know beforehand 2) if the paper covers it. (Couldn’t see on abstract) That’s fine too.