Papers of the day   All papers

Discovering Symbolic Models from Deep Learning with Inductive Biases

Comments

Miles Cranmer: Very excited to share our new paper "Discovering Symbolic Models from Deep Learning with Inductive Biases"! We describe an approach to convert a deep model into an equivalent symbolic equation. Blog/code: https://astroautomata.com/paper/symbolic-neural-nets/ Paper: https://arxiv.org/abs/2006.11287 Thread👇 1/n https://t.co/VGdMWDFSAs

29 replies, 2009 likes


hardmaru: Their network rediscovers known equations like force laws and Hamiltonians, and finds new interpretable closed-form expression for astrophysics. Cool work! I wonder if this type of approach can be applied outside the physical sciences, to social sciences, politics and economics.

12 replies, 520 likes


Miles Cranmer: Happy to announce that our work on converting deep models to symbolic equations has been accepted to NeurIPS! 🍾 @PeterWBattaglia @cosmo_shirley @DavidSpergel @KyleCranmer

3 replies, 329 likes


Shirley Ho: Thanks to @ykilcher for explaining our paper with a youtube video! Paper led by @MilesCranmer w/ @PeterWBattaglia @KyleCranmer @DavidSpergel https://arxiv.org/abs/2006.11287

2 replies, 308 likes


Miles Cranmer: Here's a talk I gave at http://mlclub.net on (1) how to convert a neural network into a symbolic equation, and (2) why I think symbolic regression should be a primary ML algorithm in astrophysics. Lots of insightful questions from the audience! https://www.youtube.com/watch?v=wmQIcTOzH0k

2 replies, 135 likes


Yann LeCun: Awesome paper.

0 replies, 126 likes


ScriptOnRoblox: People are sleeping on this right now, but this is AMAZING. They're using neural networks to figure out mathematical formulas to explain the universe in a way that humans can understand. These are like cheat codes for figuring out science https://twitter.com/MilesCranmer/status/1275241300829384704?s=09

3 replies, 95 likes


Miles Cranmer: Some slides for my talk later today on the main ideas from our neural network=>symbolic model paper! (https://arxiv.org/abs/2006.11287) Astro/ML seminar series link: https://docs.google.com/document/d/1GGtE-YIuAWlmpKSr38_kyiF-Fklszhkh4FkiYWzBAho/pub https://t.co/VpKcDKgFck

4 replies, 94 likes


Ricard Solé: Can deep learning networks uncover mathematical expressions describing unknown models of physical processes? I think this study might be the start for a major leap in that direction. A great paper by @MilesCranmer & co. here:https://arxiv.org/abs/2006.11287 @rogertgn https://t.co/LJuY06WF66

4 replies, 84 likes


Bindu Reddy 🔥❤️: Fascinating work that uses ML to accelerate scientific research GNNs are used to discover equations in Astrophysics Imagine if you could extract symbolic equations from NNs and then use math to operate on them to further discover how the universe works: https://arxiv.org/abs/2006.11287 https://t.co/WnL5FEp9a7

1 replies, 67 likes


Miles Cranmer: I'm giving an astrophysics seminar tomorrow on neural nets, symbolic regression, and dark matter. Should be an interesting discussion! https://docs.google.com/document/d/1GGtE-YIuAWlmpKSr38_kyiF-Fklszhkh4FkiYWzBAho/pub

1 replies, 61 likes


Kyle Cranmer: Some fun follow up work on graph networks / symbolic regression hybrid approaches to learn physics. Great work led by @MilesCranmer (no relation). Check out the thread 👇

2 replies, 42 likes


Sourabh Katoch: Discovering Symbolic Models from #DeepLearning with Inductive #Biases Paper-https://arxiv.org/abs/2006.11287 #100DaysOfCode #100DaysOfMLCode #Machinelearning #javascript30 #RStats #hourofcode #womenintech #code #CodeNewbie #DataScience #ArtificialInteligence #AI #Python https://t.co/ZA84250N5j

0 replies, 16 likes


Socially Distant Brian Bucklew ₑͤ>∿<ₑͤ ∞🌮: caves of qud

1 replies, 14 likes


Jelle Zuidema: Interesting parallels between the uses of deep learning in astrophysics and in linguistics. This work uses Graph NNs, building on a paper (Battaglia et al. 2018) that starts its discussion with Von Humboldt and Chomsky!

0 replies, 14 likes


Dagmar Monett: #AI hybrids have been always the oft unreservedly forgotten, unsung heroes. Here 👇, they excel once more. #GeneticAlgorithms & #NeuralNetworks #symbolic & #connectionist #AI #GOFAI #NN #DL

1 replies, 12 likes


Song Huang: Sounds exciting to me for now, but probably not so much when a neural network replaces me as a researcher in 10 yrs...

1 replies, 10 likes


Jude Gomila: It's exciting how AI may find new physical laws. Symbolic modeling using graph neural networks may allow us to short cut finding useful laws and formula for practical use e.g. this method discovered a new dark matter formula https://arxiv.org/abs/2006.11287 https://www.youtube.com/watch?v=LMb5tvW-UoQ&feature=youtu.be

1 replies, 9 likes


arxiv: Discovering Symbolic Models from Deep Learning with Inductive Biases. http://arxiv.org/abs/2006.11287 https://t.co/929F0M3V9Y

0 replies, 8 likes


MONTREAL.AI: Discovering Symbolic Models from Deep Learning with Inductive Biases Cranmer et al.: https://arxiv.org/abs/2006.11287 Blog and code: https://astroautomata.com/paper/symbolic-neural-nets/ #Cosmology #Astrophysics #ComputationalPhysics https://t.co/diF5RAnYwQ

0 replies, 8 likes


Dr. Abhay Alok Singh: Congrats Miles for this excellent research.

0 replies, 4 likes


Fabian Offert: This is all over ML Twitter already but retweeting it anyway because it is so cool.

0 replies, 3 likes


higgsinocat: Discovering Symbolic Models from Deep Learning with Inductive Biases. (arXiv:2006.11287v1 [cs.LG]) relevance:100% http://arxiv.org/abs/2006.11287 #darkmatter @StatRui @KyleCranmer @DavidSpergel https://t.co/auZfK2NKBr

0 replies, 2 likes


Shlomo Engelson Argamon: Fascinating.

1 replies, 1 likes


DelocalizedDanny: Interesting approach to #NeuralNetworks providing access to equations. Stems my phycisist 🧡 hopeful for the future of #MachineLearning .🙌

0 replies, 1 likes


Parisa Rashidi: Converting deep learning models to equivalent symbolic equations in physics.

0 replies, 1 likes


Brant Robertson: Excellent talk just now by @MilesCranmer, seems like symbolic regression + deep learning has lots of potential in the physical sciences.

0 replies, 1 likes


Marco Manca: Thread... & this -> https://twitter.com/tegmark/status/1275865930858483713?s=20

0 replies, 1 likes


Content

Found on Jun 23 2020 at https://arxiv.org/pdf/2006.11287.pdf

PDF content of a computer science paper: Discovering Symbolic Models from Deep Learning with Inductive Biases