Data-driven model discovery: Targeted use of deep neural networks for physics and engineering
2024 ж. 14 Мам.
25 950 Рет қаралды
website: faculty.washington.edu/kutz
This video highlights physics-informed machine learning architectures that allow for the simultaneous discovery of physics models and their associated coordinate systems from data alone. The targeted use of neural networks and enforcement of parsimonious models allows for a robust architecture that can be broadly applied in the sciences.
So much information packed into a 45 minute presentation. The insights are amazing! Would you mind adding links to the work you cite in the description as well? Thank you again for posting these!
Excellent content again. Thank you Prof. Kutz!
Thanks for putting the time and energy to publish these. They are quite unique even though the topics are covered so many other places.
Seems like this technique has enormous potential. Great talk!
This is simply amazing Dr. Kutz.
Brilliant talk from Brilliant prof Kutz!!
Fantastic fantastic fantastic! Especially the history of science intro puts everything into perspective.
I got printing presented papers one by one and each time I said to myself: "Here is what you may use for your work":)) Wonderful presentation of great order and logic! One of the best I have ever seen from you on youtube. Thanks a lot
Hi Nathan, this topic is absolutely mind blowing, congrats for the whole team for this job!!
Heavily loaded presentation. Thank you so much
Wonderful talks, many thanks, it is very inspiring.
Great work.. Thank you for sharing
Awesome lecture!
Great material
Excellent!
honestly content is Amazing but unorganized. please make some Playlist for related contents ! thank you!!!
Fantastic! Your contents are awesome. But you need to make people click on your videos to make them see it. Unfortunately you must use eye catching thumbnail and a caption that is both exciting and accurate. I want your channel to boost. You deserve the fame.
Very interesting
Thank you for the video Prof Nathan. I have a question though: are SINDY libraries baked in the model as a model layer? Or are they simply used in the loss function? Thank you.
Thanks for the great work! Is the phase for "finding coordinates" just "feature engineering" in machine learning terminology or any difference?
I feel it's more similar to looking at the neuron activations and trying to come up with the co-ordinates.
yes they are precisely the same things
Koopman theory reminds me of Hilbert-Huang Transform.
How to make collective excitations out of walking droplets?
lol during the previous video I was like "oh, this notation for the model is exactly like what I saw in that paper. I guess that's pretty much standard then." Now I noticed the paper was from this guy...
Fast Transform fixed-filter-bank neural networks are a thing.
A lot of what I'm starting to think is that non-linearity is actually misrepresentation. Maybe everything is linear? Obviously Koopman sort of says this but what if non-linearity is a product of our representation of systems? Maybe because we try to force things in to our mathematical systems and this requires, in our system, non-linear representation. In some sense this is also obvious but then it seems the natural thing would be to understand precisely how to represent things in their canonical system. This may not be possible, maybe the devil is really binding us to what we think is linear.
Old wine in a new bottle!! What is the use of Neural networks in this framework? I mean, if you define the polynomials in advance, why do we need NNs? Why cannot we find the parameters by using the standard regression techniques, like in established methods? I guess this is what you are doing anyway. How is this different from standard well-established system identification methods?
It's Ludacris The only neural network you need is in your head there will be no progress in physics until first principles are revisited not going to be able to black box this one
Not that it's super important but, Netwon wasn't made more famous because he "invented" a heuristic. Both did that in different ways. What made Newton more famous than Kepler is more likely the wide pop cultural applicability of the heuristic.