Papers of the day   All papers

CURL: Contrastive Unsupervised Representations for Reinforcement Learning

Comments

Aravind Srinivas: New paper - CURL: Contrastive Unsupervised Representations for RL! We use the simplest form of contrastive learning (instance-based) as an auxiliary task in model-free RL. SoTA by *significant* margin on DMControl and Atari for data-efficiency. https://arxiv.org/abs/2004.04136 https://t.co/3ZSfzmxemE

12 replies, 943 likes


hardmaru: Contrastive unsupervised features might be the ingredient that makes RL much more sample efficient. Nice results demonstrating the method for Atari and continuous control (from pixels) tasks. @Aravind7694 @MishaLaskin @pabbeel https://arxiv.org/abs/2004.04136 https://mishalaskin.github.io/curl/

1 replies, 207 likes


Michael Laskin: Can pixel-based RL be as data-efficient as state-based RL? We show for the first time that the answer is yes, new work with @Aravind7694 and @pabbeel website 👉 http://mishalaskin.github.io/curl code 👉 https://github.com/MishaLaskin/curl

2 replies, 104 likes


hardmaru: Another datapoint: https://twitter.com/Aravind7694/status/1248049713149906945

1 replies, 34 likes


Alex Nichol: Amazing work! This is a very important moment for RL.

1 replies, 31 likes


Max Jaderberg: Simple and effective, fantastic! 👏👏

0 replies, 30 likes


roadrunner01: CURL: Contrastive Unsupervised Representations for Reinforcement Learning pdf: https://arxiv.org/pdf/2004.04136.pdf abs: https://arxiv.org/abs/2004.04136 webpage: https://mishalaskin.github.io/curl/ github: https://github.com/MishaLaskin/curl https://t.co/nML3KnRzKr

0 replies, 18 likes


Aravind Srinivas: New paper - Reinforcement Learning with Augmented Data (RAD)! Data augmentation *alone* can achieve SoTA on DMControl and test-time generalization on ProcGen! Paper: https://arxiv.org/abs/2004.04136 Project Page: https://mlaksin.github.io/rad Code: https://github.com/MishaLaskin/rad https://t.co/VtA2oHRBvY

1 replies, 11 likes


Timothy O'Hear: Fascinating results, would love to a @CShorten30 or @ykilcher video on this paper :-)

1 replies, 8 likes


rewon: Simple, powerful, and efficient!

0 replies, 4 likes


Daisuke Okanohara: CURL performs unsupervised contrastive representation learning simultaneously with pixel-based off-policy RL training, which enables to reduce the required training samples significantly. Simple and very effective for data-efficient RL. https://arxiv.org/abs/2004.04136

0 replies, 4 likes


ALife Papers: CURL: Contrastive Unsupervised Representations for Reinforcement Learning "CURL extracts high-level features from raw pixels using contrastive learning and performs off policy control on top of the extracted features." https://github.com/MishaLaskin/curl https://arxiv.org/pdf/2004.04136.pdf https://t.co/YqjA19deZV

0 replies, 1 likes


Xander Steenbrugge: When applied to RL, these techniques yield vastly more sample-efficient agents, but are still quite far off human data efficiency... The open question is how to learn generalizable priors that transfer across a wide range of environments & tasks! https://twitter.com/Aravind7694/status/1248049713149906945

1 replies, 0 likes


Content

Found on Apr 09 2020 at https://arxiv.org/pdf/2004.04136.pdf

PDF content of a computer science paper: CURL: Contrastive Unsupervised Representations for Reinforcement Learning