Papers of the day   All papers

Up to two billion times acceleration of scientific simulations with deep neural architecture search

Comments

Duncan Watson-Parris: Our recent #ML paper describing a new, general physics emulator (DENSE; https://arxiv.org/abs/2001.08055) just had a really nice write-up in @sciencemagazine, check it out! https://www.sciencemag.org/news/2020/02/models-galaxies-atoms-simple-ai-shortcuts-speed-simulations-billions-times https://t.co/BHfsTtCI6B

7 replies, 207 likes


Quoc Le: Nice work on speeding up physics simulations with convolutional nets & neural architecture search.

1 replies, 168 likes


Eirini Malliaraki: 'Up to two billion times acceleration of scientific simulations with deep neural architecture search' ! https://arxiv.org/abs/2001.08055

1 replies, 42 likes


Jan Jensen: From models of galaxies to atoms, simple AI shortcuts speed up simulations by billions of times https://www.sciencemag.org/news/2020/02/models-galaxies-atoms-simple-ai-shortcuts-speed-simulations-billions-times

0 replies, 32 likes


News from Science: New research is showing how #AI can help accelerate simulations, like ones that model the atmosphere, by billions of times. It focuses on improving the accuracy of algorithms known as emulators. https://fcld.ly/z838rss

1 replies, 20 likes


Daisuke Okanohara: To emulate slow simulators by NN, we can combine network architecture search to find the right model and build an accurate emulator even with small training data. Succeeded in 10 scientific simulations and achieved up to 2 billion times acceleration. http://arxiv.org/abs/2001.08055

0 replies, 19 likes


José Luis Ricón (Artir): Tired: Thinking about problems Wired: High-throughput search your way through problems https://twitter.com/irinimalliaraki/status/1230558582346641415 c @AdamMarblestone

1 replies, 8 likes


Yves Hilpisch: Up to two billion times acceleration of scientific simulations with deep neural architecture search https://arxiv.org/abs/2001.08055

0 replies, 5 likes


Brad Neuberg: Idea: train a deep net to be able to emulate a slow simulation but much faster, such as in the work at https://arxiv.org/pdf/2001.08055.pdf. Then, use this much faster deep net emulator to train a reinforcement learning algorithm such as PPO to find a policy for solving some problem.

1 replies, 3 likes


JuliaGo: #Deeplearning based #emulators can speed up complex simulations in scientific cases including astrophysics and climate science while retaining accuracy. #DENSE #DeepEmulatorNetworkSearch Paper: https://arxiv.org/abs/2001.08055

0 replies, 3 likes


Jan Jensen: Here's the preprint https://arxiv.org/abs/2001.08055

1 replies, 3 likes


Chris Adams at Bosch Connected World: I don't know enough about neutral networks to know if this would actually help much with simulations for climate in practice, but the numbers are astronomical, and I learn loads from following Eirini, so....

1 replies, 2 likes


akira: https://arxiv.org/abs/2001.08055 Up to two billion times acceleration of scientific simulations by using Neural Architecture Search. They model well with small data such as ~100 since the CNN architecture are a good prior. https://t.co/2aIgmpbuTN

0 replies, 1 likes


Zook: i heard you like deep learning for science. so i put deep learning in your science so you can learn while you learn https://arxiv.org/abs/2001.08055

0 replies, 1 likes


Content

Found on Feb 13 2020 at https://arxiv.org/pdf/2001.08055.pdf

PDF content of a computer science paper: Up to two billion times acceleration of scientific simulations with deep neural architecture search