Co-Founder at Ineffable Intelligence Previously Research Scientist at DeepMind
London, England
Joined March 2015
- Tweets20
- Following77
- Followers937
- Likes36
Excited to announce that our work on “Discovering state-of-the-art RL algorithms” is finally published in @Nature! In this work, we meta-learned RL algorithms at scale.
Paper: nature.com/articles/s41586-0…
Blog: google-deepmind.github.io/di…
See thread 👇
What did it discover? DiscoRL discovered entirely new prediction semantics.
While humans rely on concepts like "value functions," DiscoRL learned to predict salient future events - like high rewards or changes in policy entropy - that complement traditional RL concepts.
Work done with the amazing team: @iurii_kemaev [tech lead], @greg_far, @dancalian, @luisa_zintgraf, @matteohessel, Satinder Singh, @hado, David Silver
Code & weights: github.com/google-deepmind/d…
Paper: nature.com/articles/s41586-0…
Junhyuk Oh retweeted
Excited to share our new paper, "DataRater: Meta-Learned Dataset Curation"!
We explore a fundamental question: How can we *automatically* learn which data is most valuable for training foundation models?
Paper: arxiv.org/pdf/2505.17895 to appear @NeurIPSConf
Thread 👇
I will be briefly talking about how I used JAX to implement my recent #NeurIPS2020 work on Discovering RL Algorithms (arxiv.org/pdf/2007.08794.pdf). Stop by our livestream if you are interested. :)
Join our team at the #NeurIPS2020 JAX Ecosystem meet up to learn more about JAX and why it's effective for research in reinforcement learning, GANs, meta-gradients and more.
Today at 11am PST / 2pm ET/ 7pm GMT
bit.ly/3oO8Lx9 (calendar invite)
Junhyuk Oh retweeted
Our most recent work is out in Nature! We're reporting on (reinforcement) learning to navigate Loon stratospheric balloons and minimizing the sim2real gap. Results from a 39-day Pacific Ocean experiment show RL keeps its strong lead in real conditions. nature.com/articles/s41586-0…
Junhyuk Oh retweeted
In a major scientific breakthrough, the latest version of #AlphaFold has been recognised as a solution to one of biology's grand challenges - the “protein folding problem”. It was validated today at #CASP14, the biennial Critical Assessment of protein Structure Prediction (1/3)
Junhyuk Oh retweeted
CfP for the @iclr2019 workshop on structure and priors in reinforcement learning (SPiRL), deadline 3/7! spirl.info/2019/call/
Junhyuk Oh retweeted
really happy to announce the next version of our #RL framework: Dopamine 2.0!
beyond atari: now we support general discrete-domain gym environments.
we've been using this internally for our research and it allows us to test out new ideas very quickly.
try it out!
Introducing Dopamine 2.0, a major update to the open source framework for flexible reinforcement learning.
No longer limited to the Arcade Learning Environment, researchers can now iterate faster on ideas before testing them on larger Atari games. opensource.googleblog.com/20…
Junhyuk Oh retweeted
Join us and @Blizzard_Ent this Thursday at 6:00pm GMT for an exciting #StarCraft demonstration, hosted by @Artosis and @RotterdaM08!
Livestream on YouTube: youtube.com/c/deepmind
Read more about #StarCraft2 as an environment for AI research: deepmind.com/blog/deepmind-a…
Junhyuk Oh retweeted
Delighted to welcome reinforcement learning pioneer Satinder Singh to @DeepMindAI. He’ll bring some incredible experience to the team and I'm really looking forward to working with him!
Junhyuk Oh retweeted
Excited to share a quite extensive introduction to deep reinforcement learning! With @astro_pyotr @riashatislam @marcgbellemare and Joelle Pineau, we hope it will be useful to the community. Print version available at #NeurIPS! arxiv.org/abs/1811.12560
Junhyuk Oh retweeted
"Self-Imitation Learning," Oh and Guo et al.: arxiv.org/abs/1806.05635
Imitating past good experiences in the replay buffer leads to big improvements over A2C, PPO, inc. good Montezuma performance in fewer frames than prior approaches
Junhyuk Oh retweeted
NIPS Deep RL Symposium Schedule now available: sites.google.com/view/deeprl…
includes over 70 contributed papers/posters, and invited talks by David Silver, Joelle Pineau, Ruslan @rsalakhu , Ben Van Roy, Michael Bowling. Thursday 12/7 @NipsConference
Junhyuk Oh retweeted
NIPS Deep RL Symposium deadline Fri Nov 3. We welcome your #ICLR Deep RL submissions! (+other recent work)
sites.google.com/view/deeprl…
Junhyuk Oh retweeted
Check out the code (with trained models) for our ICLR 2017 paper on video prediction.
github.com/rubenvillegas/icl…
@albert_swart Thank you for mentioning my code. :) They will merge Jeff Donahue's code soon. See the following PR: github.com/BVLC/caffe/pull/2…
