Sam Greydanus: 1/5 Introducing “Lagrangian Neural Networks”
We show how to parameterize arbitrary Lagrangians and learn them with neural networks. Work done w/ with @MilesCranmer @shoyer @PeterWBattaglia @DavidSpergel @cosmo_shirley https://t.co/9BLumnzwsL
11 replies, 525 likes
Miles Cranmer: 1/10 Very excited to present Lagrangian Neural Networks, a new type of architecture that conserves energy in a learned simulator without requiring canonical coordinates. w/ @samgreydanus, @shoyer, @PeterWBattaglia, @DavidSpergel, @cosmo_shirley: https://arxiv.org/abs/2003.04630 https://t.co/8r4sloBNnH
11 replies, 390 likes
Simone Scardapane: *Lagrangian Neural Networks* 1/4
Minimalist, beautiful paper and experiments.
They model a generic Lagrangian with a neural net, then use it to integrate the dynamics of the system itself.
Code in #JAX is almost as beautiful as the paper!
2 replies, 84 likes
Ryan Reece: To philosophers concerning the scientific realism debate: this is a way to make manifest that certain models are "true" descriptions of the data independent of us. Neural nets can find them.
3 replies, 27 likes
Miles Cranmer: Excited for ICLR's virtual workshop tomorrow on Neural Networks X Differential Equations. Tune in to see our "poster talk" on Lagrangian Neural Networks!
Live-streamed from http://iclr2020deepdiffeq.rice.edu/ - our talk is at 6:00 PM Pacific (UTC-7). Q&A session afterward!
0 replies, 23 likes
Statistics Papers: Lagrangian Neural Networks. http://arxiv.org/abs/2003.04630
0 replies, 21 likes
Shirley Ho: Proud of our new architecture that conserves energy in a learned simulator. Wonder what kind of #physics and #RealWorld problems we can solve with this new architecture? @MilesCranmer @samgreydanus @shoyer
0 replies, 18 likes
federica bianco 🔭👊: Should we just replace publishing with twitter threads? This thread is clear + gives a great overview of the paper. Ok, I still want to read the paper for details, but THIS is what I mean when I tell my students your figures+captions should tell the whole story! #scicomm
1 replies, 13 likes
Scott H. Hawley: @osprangers @Lorenzo85710163 @hardmaru For physical systems such as @hardmaru’s example, reviews like these would do well to include the excellent work by @samgreydanus et al re. Hamiltonian (https://greydanus.github.io/2019/05/15/hamiltonian-nns/) & Lagrangian neural networks; most recently @MilesCranmer et al https://arxiv.org/abs/2003.04630
1 replies, 5 likes
Roberta: So this week I had this very same problem: my model didn't conserve energy nor mass. My advisor showed me this paper and I'm just "UOW"
2 replies, 1 likes
Miles Cranmer: 2/10 First, papers referenced in main tweet:
LNNs - https://arxiv.org/abs/2003.04630
HNNs - https://arxiv.org/abs/1906.01563
Graph Nets - https://arxiv.org/abs/1806.01261 (+ refs therein...)
Group-CNNs - http://proceedings.mlr.press/v48/cohenc16.pdf https://t.co/ic8jH0Y9fc
1 replies, 1 likes
matthew miller: Recently did myself, or sort of. Had to derive the Lagrangian for a relativistic Schrödinger-like equation. Required knowledge of higher order Largrangians but ultimately hacked my way through to a solution. Interesting it can be automated, though this algorithm does much more...
0 replies, 1 likes
Found on Mar 11 2020 at https://arxiv.org/pdf/2003.04630.pdf