Papers of the day   All papers

Unsupervised Learning of Object Keypoints for Perception and Control


Jul 01 2019 DeepMind

Deep RL agents are data hungry and often learn task-specific representations. Our model learns object-centric abstractions from raw videos. This enables highly data-efficient RL and structured exploration.
9 replies, 1144 likes

Jul 01 2019 Tejas Kulkarni

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.
4 replies, 364 likes

Oct 24 2019 Tejas Kulkarni

We have released the code for transporter — a neural network architecture for unsupervised learning of object keypoints (now a NeurIPS paper):
1 replies, 286 likes

Jul 01 2019 David Pfau

Getting sensible object detection is key to interpretable, data-efficient and general-purpose agents (i.e. agents that don't fall flat on their face if you tweak one little thing about the environment). This is some really interesting work in that direction.
0 replies, 51 likes

Dec 11 2019 Tejas Kulkarni

I will be at the @NeurIPSConf to present this work tomorrow and more (Wednesday, 10:45 AM -- 12:45 PM @ East Exhibition Hall B + C. Poster #124). Come say hello if you are around!
0 replies, 50 likes

Jul 01 2019 Arthur Juliani

Typical DeepRL algorithms lack any priors to enable them to learn to reason over persistent objects from raw pixels. This is a nice work exploring one way that might be possible.
0 replies, 35 likes

Jul 01 2019 Vlad Mnih

Excited to share some recent work on unsupervised learning of object-centric representations for reinforcement learning.
0 replies, 11 likes

Oct 26 2019 Daisuke Okanohara

For efficient control and exploration in RL, Transporter discovers object keypoints. It learns from raws videos by 1) extract key points from source/target frames 2) manipulate source feature map around keypoints 3) reconstruct target frame.
0 replies, 7 likes

Jul 01 2019 IntuitionMachine

DeepMind trains Atari play with only 100k interactions: .
0 replies, 3 likes