Can a large model learn generalizable concepts of physics?
We present TDMs, transformer-based large dynamics models with emergent generalization capabilities to unseen environments.
Preprint: https://t.co/MJMBOcR0WH
Thanks to my collaborators at @DeepMind: Jingwei Zhang, Jake Bruce, Sarah Bechtle, Emilio Parisotto, @MartinRiedmill1, Jost Tobias Springenberg, @arunbyravan, @lqh20, Nicolas Heess!
Can a large model learn generalizable concepts of physics?
We present TDMs, transformer-based large dynamics models with emergent generalization capabilities to unseen environments.
Preprint: https://t.co/MJMBOcR0WH
While not our primary focus, TDMs are remarkably capable as “classic” specialist dynamics models as well. We use TDMs for planning in a range of difficult tasks in the DeepMind control suite.
At NeurIPS, we present an approach that combines the strengths of reinforcement learning and planning by learning policies that can execute plans.
Paper: https://t.co/Bq83GE4abg
Video: https://t.co/HzFJkyZJFc
Code: https://t.co/qnm2uJvEbU
Visit our poster on Dec 8, 8.30 am UTC!
By relaxing some of the optimality guarantees of reward shaping to a guarantee of preserved long-time-behaviour, the sample efficiency of reward shaping can be increased dramatically!
To learn more, join our session at ICLR 2021 at 4 p.m. UTC: https://t.co/s8P0WVYkUv
Our lab unites researchers from many backgrounds. @IngmarSchubert for example was a quantum physicist before joining us. Check out his recent work theoretically modelling quantum correlation effects in multi-electron ionization of large C60 molecules): https://t.co/WqXfzvdAkz
Die #CoronaWarnApp herunterladen – und gemeinsam das #Coronavirus bekämpfen:
Die App ist jetzt im App Store und im Google Play Store verfügbar. https://t.co/2MzLOViTdz
With ICRA starting soon, also check out our recent work: Robust Task and Motion Planning for Long-Horizon Architectural Construction Planning (https://t.co/blXlLThKh0)
Check out our ICRA paper: Deep Visual Heuristics: Learning Feasibility of Mixed-Integer Programs for Manipulation Planning.
Nominated for Best Paper Award in Automation.
Presentation: https://t.co/ZMNWicT07B
PDF: https://t.co/QkqcLAsZj1
Slack: https://t.co/E1SCv0APlg
Check out our paper at the #MLPC workshop at #ICRA2020 on how to extend any off-the-shelf #ReinforcementLearning algorithm for learning a value function to continuous action spaces:
Presentation: https://t.co/tV7zS4bclT
@tobilindh Do I understand it correctly that they only expect the approval process to be done in autumn this year? Does that mean construction is stopped until then?
Announcing TensorFlow Quantum (#TFQ), an open-source library for the rapid prototyping of quantum #MachineLearning models, bringing the quantum computing and ML research communities together to accelerate the discovery of new quantum algorithms. https://t.co/y624QyVjV8