Thank you so much for sharing about Symbolic Regression! I'm not in the development of SR, but have been testing a lot of the variants for some time for engineering and finance. It's surprisingly useful for HFT. It's incredibly relevant today despite discouragement simply because it's an old concept. One paper showed that it could compress data, two papers showed some could outperform SVM despite being much faster for inference (800x + faster in one test of my own). It's shown use cases in electrical engineering, civil engineering, and physics, and finance. The solutions are low level, without the need of libraries. Some are robust to noise too. Again, thanks for your discussion and sharing!
@@taumag Sorry, I haven't used Pysindy yet. All I can say, is that the deeper the tree, the more difficult it is to get a globally optimal solution using genetic programming. PySR has some nifty features that alleviate this weakness, and it's fast. It's ability to add a myriad of permutations of operators can help avoid imaginary numbers and div/low explosions of numbers, and help it find an optimal solution faster. Personally, I prefer to keep the depth low yet keep the width wide using multi-gene genetic programming. It allows for lower latency, parallelism, and faster convergence to the solution. Have fun!
@@yensteel Thanks for the feedback! I sincerely appreciate it. You might be interested in a search for "SymINDy: Symbolic Identification of Nonlinear Dynamics" It's a combination of SR with SINDy.
It would be beyond great to see a video on Kolmogorov-Arnold Networks (KANs) leveraging their interpretability for Physics Informed ML somehow. Perhaps, KANs could be used to replace MLP/FFN blocks in existing Physics Informed ML models?
It would be awesome to see videos on SPINDE and Neural SDEs too. Can symbolic regression be used to learn/find SDE terms to fit to data as an alternative to Neural SDEs?
great video, I want to know what is this pysr model or library is good for fitting the predetermined equations or you can fit the data as well, i mean can i give this model a bunch of data and it will be able to tell me the equation.
Sir can you pls structure all of your videos, I will be starting my undergrad soon so this will help a lot, we would be extremely grateful to you, THANKU 🙏🙏
I asked the same thing. Have you found anyone who made an attempt? PySR can't do differential equations where PySINDy can. They are complementary areas of research, not exclusive.
@@taumag unfortunately none of it really matters any more because the world is run by psychopaths and they are openly and freely committing a genocide among other things(hijacking other countries, undermining peoples sanity, undermining the financial system, etc).
![Residual Networks (ResNet) [Physics Informed Machine Learning]](http://i.ytimg.com/vi/w1UsKanMatM/mqdefault.jpg)
![Residual Networks (ResNet) [Physics Informed Machine Learning]](/img/tr.png)
![Lagrangian Neural Network (LNN) [Physics Informed Machine Learning]](http://i.ytimg.com/vi/HLUIx6FqAvg/mqdefault.jpg)
![AI/ML+Physics Part 5: Employing an Optimization Algorithm [Physics Informed Machine Learning]](/img/1.gif)
As someone in this field, your work is having a 1000x impact with these easy to digest explainers. Absolutely fantastic work!
Thank you so much for sharing about Symbolic Regression! I'm not in the development of SR, but have been testing a lot of the variants for some time for engineering and finance. It's surprisingly useful for HFT.
It's incredibly relevant today despite discouragement simply because it's an old concept. One paper showed that it could compress data, two papers showed some could outperform SVM despite being much faster for inference (800x + faster in one test of my own). It's shown use cases in electrical engineering, civil engineering, and physics, and finance. The solutions are low level, without the need of libraries. Some are robust to noise too.
Again, thanks for your discussion and sharing!
I've been working on PySINDy for HFT. How does PySR help improve the process?
@@taumag Sorry, I haven't used Pysindy yet. All I can say, is that the deeper the tree, the more difficult it is to get a globally optimal solution using genetic programming. PySR has some nifty features that alleviate this weakness, and it's fast. It's ability to add a myriad of permutations of operators can help avoid imaginary numbers and div/low explosions of numbers, and help it find an optimal solution faster.
Personally, I prefer to keep the depth low yet keep the width wide using multi-gene genetic programming. It allows for lower latency, parallelism, and faster convergence to the solution. Have fun!
@@yensteel Thanks for the feedback! I sincerely appreciate it. You might be interested in a search for "SymINDy: Symbolic Identification of Nonlinear Dynamics" It's a combination of SR with SINDy.
13:48 Interesting that the Planck and Rydberg benchmarks, which I assume are data from quantum systems, have a 0/5 in every method tested
Simply fantastic outreach work Professor Brunton. Thank you so much for the incredible contributions you bring to your channel!
Congrats to Miles for getting coverage here
Love your work! Makes me want to lean more into this filed of research
Vivek here - awesome video! What about KANs (Kolmogorov Arnold Networks)? - would you say they belong to the family of "interpretable ML models"?
It would be beyond great to see a video on Kolmogorov-Arnold Networks (KANs) leveraging their interpretability for Physics Informed ML somehow. Perhaps, KANs could be used to replace MLP/FFN blocks in existing Physics Informed ML models?
Curious as to why PySR failed with Planck. Was that due to weakness in modeling stochastic diffeqs?
It would be awesome to see videos on SPINDE and Neural SDEs too. Can symbolic regression be used to learn/find SDE terms to fit to data as an alternative to Neural SDEs?
Great video. Thank you Prof
great video, I want to know what is this pysr model or library is good for fitting the predetermined equations or you can fit the data as well, i mean can i give this model a bunch of data and it will be able to tell me the equation.
Why is PySR considered N/A for DE?
❤thank you sir very informative. I kinda understood it.
What are the best entry level books for ML, AI, any other related topic?
This is amazing. Thank you!
Sir can you pls structure all of your videos, I will be starting my undergrad soon so this will help a lot, we would be extremely grateful to you, THANKU 🙏🙏
what does “structure all of your videos” mean?
I believe genetic programming enables the evolution of computer programs, not just mathematical expressions like symbolic regression
Thank you...
Wow! just wow
So now someone just has to combine Sindy with PySR. It should be pretty simple.
I asked the same thing. Have you found anyone who made an attempt? PySR can't do differential equations where PySINDy can. They are complementary areas of research, not exclusive.
@@taumag unfortunately none of it really matters any more because the world is run by psychopaths and they are openly and freely committing a genocide among other things(hijacking other countries, undermining peoples sanity, undermining the financial system, etc).
Search for "SymINDy: Symbolic Identification of Nonlinear Dynamics" It's a combination of SR with SINDy.
it is just trial and error, nothing wrong with that but it is not very sophisticated
Funny to see scientists still stuck in python even after industry has moved away from python 🤣🤣🤣🤣
moved to where? 98% of ML engineering and data science is in Python.
@herewegoagain2 This superiority complex type of comment stems from insecurity 99% of the time.
Moved where? SQL ? 🤣🤣
What industry? Trolling industry?