Papers of the day   All papers

Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels

Comments

Denis Yarats: Exciting to announce our new work together with @ikostrikov and @rob_fergus: Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels. Paper: https://arxiv.org/abs/2004.13649 Code: https://github.com/denisyarats/drq Website: https://sites.google.com/view/data-regularized-q [1/N]

6 replies, 399 likes


hardmaru: Deep RL from Pixels is moving so quickly, and the ideas that achieve large improvements are the simple ones “SOTA on the DeepMind control suite, surpassing model-based (Dreamer, PlaNet, and SLAC) methods and recently proposed contrastive learning (CURL)” https://arxiv.org/abs/2004.13649 https://t.co/w6crfG9kfv

5 replies, 310 likes


roadrunner01: Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels pdf: https://arxiv.org/pdf/2004.13649.pdf abs: https://arxiv.org/abs/2004.13649 project page: https://sites.google.com/view/data-regularized-q github: https://github.com/denisyarats/drq https://t.co/VNmMkOJuZw

0 replies, 295 likes


Soumith Chintala: This paper is doing rounds, SOTA on DeepMind Control Suite by adding simple data regularization. So simple! The code is pretty concise, easy to build on top: https://github.com/denisyarats/drq

2 replies, 245 likes


Yann LeCun: DrQ: awesome new RL technique from my NYU colleagues @Ikostrikov, @denisyarats and @rob_fergus . Beats SOTA on the DeepMind Control Suite (including model-based methods).

0 replies, 62 likes


Denis Yarats: People have been asking us whether our (together with @ikostrikov and @rob_fergus ) data augmentation techniques (DrQ https://arxiv.org/pdf/2004.13649.pdf) can also improve sample efficiency in Atari? The answer is YES! See the tweet chain for further information. [1/N] https://t.co/oStSQ6MH5y

1 replies, 41 likes


Kai Arulkumaran: * When you have a good prior. Data augmentation has been critical to many successes recently, but in domains such as vision/text where we know how to make meaningful interventions on the data. Have we replaced feature engineering with augmentation engineering?

2 replies, 38 likes


Denis Yarats: We extend our data augmentation techniques from DrQ (https://arxiv.org/abs/2004.13649) to on-policy setting which results to much improved generalization on ProcGen! Now DrQ is proven to be effective for a variety of RL algorithms, both off and on policy!

1 replies, 36 likes


Kyunghyun Cho: spicy! contrastive learning is also not needed but just clever augmentation at multiple points in an algorithm is for pocel-level control. https://arxiv.org/abs/2004.13649 @ikostrikov @denisyarats @rob_fergus

3 replies, 26 likes


Ting Chen: Data augmentation should not be overlooked. We have shown it is critical for contrastive learning. This work shows it is also very important for RL!

1 replies, 18 likes


Markus Wulfmeier 🏡: 'Nothing is as powerful as an idea whose time has come!' Image augmentations are enabling some considerable performance boost in Deep RL https://arxiv.org/abs/2004.13649 https://arxiv.org/abs/2004.14990 #ReinforcementLearning

1 replies, 15 likes


arXiv CS-CV: Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels http://arxiv.org/abs/2004.13649

0 replies, 12 likes


MONTREAL.AI: Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels Kostrikov et al.: https://arxiv.org/abs/2004.13649 Code: https://github.com/denisyarats/drq Website: https://sites.google.com/view/data-regularized-q #DeepLearning #MachineLearning #ReinforcementLearning https://t.co/pLplBpQZw8

1 replies, 10 likes


HotComputerScience: Most popular computer science paper of the day: "Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels" https://hotcomputerscience.com/paper/image-augmentation-is-all-you-need-regularizing-deep-reinforcement-learning-from-pixels https://twitter.com/denisyarats/status/1255325685628968961

0 replies, 1 likes


Rob Fergus: Great work from my PhD students @denisyarats and @ikostrikov !

0 replies, 1 likes


Denis Yarats: People have been asking us whether our (together with @ikostrikov and @robfergus) data augmentation techniques of DrQ (https://arxiv.org/pdf/2004.13649.pdf) can also improve sample efficiency in Atari? The answer is YES! [1/N] https://t.co/ook5oEB3q4

1 replies, 1 likes


Estesis: Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels Kostrikov et al.: https://arxiv.org/abs/2004.13649 Code: https://github.com/denisyarats/drq Website: https://sites.google.com/view/data-regularized-q #DeepLearning #MachineLearning #ReinforcementLearning https://t.co/ZHyX00LULK

0 replies, 1 likes


arXiv CS-CV: Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels http://arxiv.org/abs/2004.13649

0 replies, 1 likes


Brundage Bot: Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels. Ilya Kostrikov, Denis Yarats, and Rob Fergus http://arxiv.org/abs/2004.13649

1 replies, 1 likes


cs.LG Papers: Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels. Ilya Kostrikov, Denis Yarats, and Rob Fergus http://arxiv.org/abs/2004.13649

1 replies, 1 likes


Content

Found on Apr 29 2020 at https://arxiv.org/pdf/2004.13649.pdf

PDF content of a computer science paper: Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels