"When one does a theoretical calculation, there are two ways of doing it. Either you can have a clear physical model in mind or you should have a rigorous mathematical basis." -Freeman Dyon recounting his meeting with Enrico Fermi.
This is absolutely amazing! It feels like this is the biggest step (that I have seen) towards machines literally making discoveries and teaching us! And the fact that I understood most of it despite me having no experience in AI or programming makes me happy and is a giant praise to the presenter! I love it and I'll tell everyone about it! :D
Very interesting! I wish psychology and neuroscience applied similar models! The brain and associated behaviour would be a lot easier to organise and interpret if we had symbolic methods of representing the interactions.
**Start of the video** Miles: if you forget everything from this, i need you to remember:... Me: oh man i hope he's not saying this cause it's boring **End of the video** Me: this man is a god
Could you add a reference to the work on extracting fluid dynamics PDEs from a trained GNN that was/is co-led by Elaine Cui (Flatiron Inst)? Link to pre-prints would be fine too. Thanks!
Hey Steve, are you saying that neural networks are models of the brain? What if that brain was Newton's brain? Congratulations, Mr. Cranmer. A major breakthrough.
Wow, super cool, very happy I stumbled upon this video. A question, at the end of this process do you just get a "point estimate" in the space of functions that have the basis you gave? If so, how difficult would it be to obtain a probabilistic generative model that gives you a "distribution" of functions? I am completely in the dark about this topic, but definitely wanting to know more. Cool stuff!
Why does a DL model be only converted to Math models, but not into an algorithm or combination of both. That means you can take a DL black box model into Algorithm+ math equation forming a code.
The fact that it is conceptually simple does not mean the applications can't be great. Most "big" scienctific discoveries are built upon thousands of "conceptually simple discoveries" like this. Edit: I always think of it like this: given enough time, could I have created this myself? Possibly, but unlikely. It is always easy to tell yourself "I could have done that" once you already know it.
@@jeroenritmeester73 Thanks for the comment. I re-read my comment and realized it had a negative tone. I'm actually super impressed with this guys work! It's a huge step in the right direction. I love the idea of developing analytical formulas directly from data by using machine learning :)
You compress the multivariable into a plane like dropping marbles onto a floor. And wait till they roll far enough so the space of vector fields that play with them doesn't interact.
If you are talking about a vector with 212 components then you will have 212 marbles in this plane before dropping through an optimized vector field. I think what he is talking about is the vector sums of the position functions of all the marbles interacting through these vector fields and averaging them to have a net movement to generalize the answer for that one particular component of the 212 vector.
Why is he making such sounds in between the explanations? That happens when you're thinking while presenting! Try to go through the presentation before recording!
![Physics Informed Neural Networks (PINNs) [Physics Informed Machine Learning]](http://i.ytimg.com/vi/-zrY7P2dVC4/mqdefault.jpg)
![Physics Informed Neural Networks (PINNs) [Physics Informed Machine Learning]](/img/tr.png)
A python package with a julia backend ?
What a time to be alive!
Also, I can't wait for the Elixir NX / Axon version. ; )
Wow
Next: school homework doing itself.
What a time to be alive!
two-minute papers? :)
@@tanchienhao
"Dear Fellow Scholars, this is Two Minute Papers with Dr. Károly Zsolnai-Fehér."
From now on, when I am fooling around, trying out different formulas, I will call it "manual symbolic regression". Thanks a lot! :D
This is one of the best works I have read about earlier this year, thanks for the lecture that makes it is more clear.
"When one does a theoretical calculation, there are two ways of doing it. Either you can have a clear physical model in mind or you should have a rigorous mathematical basis." -Freeman Dyon recounting his meeting with Enrico Fermi.
You got the sources of this quote?
Start at 1:15 end at 1:27 on UA-cam Freeman Dyson - Fermi's rejection of our work (94/157) Cheers!
Absolutely incredible lecture - continually coming back to the well on this one
One step closer to stop having neural networks be complete black boxes. Great work.
This is absolutely amazing!
It feels like this is the biggest step (that I have seen) towards machines literally making discoveries and teaching us!
And the fact that I understood most of it despite me having no experience in AI or programming makes me happy and is a giant praise to the presenter!
I love it and I'll tell everyone about it! :D
IMO, this is a profound direction for research. Congratulations to the entire team.
The improved generalization gives me chills.
very interesting work，but one thing is the efficiency of genetic algorithm, maybe the MCMC or HMC method can obtain a higher search efficiency
Very interesting! I wish psychology and neuroscience applied similar models!
The brain and associated behaviour would be a lot easier to organise and interpret if we had symbolic methods of representing the interactions.
Wow, the world needs more videos like this
Amazing stuff! Miles Cranmer is a genius!
Absolutely amazing video!
This kid is a pure empiricist at heart.
**Start of the video**
Miles: if you forget everything from this, i need you to remember:...
Me: oh man i hope he's not saying this cause it's boring
**End of the video**
Me: this man is a god
Very interesting, I want to apply it in Geophysics
let me know what you find!
This will start a new field
Great, will try to apply it to Economics
Could you add a reference to the work on extracting fluid dynamics PDEs from a trained GNN that was/is co-led by Elaine Cui (Flatiron Inst)? Link to pre-prints would be fine too. Thanks!
2:58 for the immunologists out there, this is somatic hypermutation for expressions!
Brilliant idea!!
Hey Steve, are you saying that neural networks are models of the brain? What if that brain was Newton's brain?
Congratulations, Mr. Cranmer. A major breakthrough.
Wow, super cool, very happy I stumbled upon this video. A question, at the end of this process do you just get a "point estimate" in the space of functions that have the basis you gave? If so, how difficult would it be to obtain a probabilistic generative model that gives you a "distribution" of functions? I am completely in the dark about this topic, but definitely wanting to know more. Cool stuff!
Why does a DL model be only converted to Math models, but not into an algorithm or combination of both. That means you can take a DL black box model into Algorithm+ math equation forming a code.
Thank you, I nice introduction and examples of the possibilities.
Well it looks like what I did in Excel making the solver optimize mathematical operators through numbers to minimize the residuals.
Thank you very much! This is awesome :o
Thank you. This is really cool. Is your package available also in Julia itself?
Entropical relatives with a minimal potential of power isn't not look foward under the Maxwell equations?
Wow i would love to see more videos like this, so many videos were just theory.
I am curious how this is different or better than what was done in the AI Feynman papers
I think this is like the PCA of one subclass of a vector component like a matrix column and not the row.
Amazing work!!
Very cool! Respect!!!
This was great!
18:14 why is it rotated?
Deep learning physics era is soon to be discovered...!
Why not try it on the 3 body problem?
Cool but the genetic programming idea by Koza (1992) and several works following that come to mind.
This is literally the holy grail... get a machine to reason scientifically and be able to communicate the results!
What are the implications of something being "low dimensional"? Please eli5
Isn't this just the genetic optimization algorithm + machine learning?
The fact that it is conceptually simple does not mean the applications can't be great. Most "big" scienctific discoveries are built upon thousands of "conceptually simple discoveries" like this.
Edit: I always think of it like this: given enough time, could I have created this myself? Possibly, but unlikely. It is always easy to tell yourself "I could have done that" once you already know it.
@@jeroenritmeester73 Thanks for the comment. I re-read my comment and realized it had a negative tone. I'm actually super impressed with this guys work! It's a huge step in the right direction. I love the idea of developing analytical formulas directly from data by using machine learning :)
I like it. PINN and now fitting on Deep net!
Can this do particle physics?
"low dimensional" is it like pure functions in funcional programming?
You compress the multivariable into a plane like dropping marbles onto a floor. And wait till they roll far enough so the space of vector fields that play with them doesn't interact.
Cool I roll marbles for a hobby.
I mean if you have 212. Then how many 101’s are there.
I say there’s 3
If you are talking about a vector with 212 components then you will have 212 marbles in this plane before dropping through an optimized vector field. I think what he is talking about is the vector sums of the position functions of all the marbles interacting through these vector fields and averaging them to have a net movement to generalize the answer for that one particular component of the 212 vector.
@@WhenThoughtsConnect a 212 vector. Hi can it add more than 10072
Very cool!
Wolfram does both. Also probabilistic programming ppl
One needs highly fluid dark matter to understand this 🙂
So, will this thing finally figure out how to get a functional hoverboard?
nice one
Wigerian Prior?
nice
This is amazing. But dude you need to breath. Im getting out of breathe listening to this.
Lol, I've been there. Just nerves and excitement. Remedied by practice.
Nice! 😏
TLDR; Math is dead.
Why is he making such sounds in between the explanations? That happens when you're thinking while presenting! Try to go through the presentation before recording!
Amazing stuff. You were doing great until you started talking about dark matter. Let’s call it what it is ‘ plasma and dust’.