AI/ML+Physics Part 2: Curating Training Data [Physics Informed Machine Learning]

This video discusses the second stage of the machine learning process: (2) collecting and curating training data to inform the model. There are opportunities to incorporate physics into this stage of the process, such as data augmentation to incorporate known symmetries.

This video was produced at the University of Washington, and we acknowledge funding support from the Boeing Company

%%% CHAPTERS %%%
00:00 Intro
03:02 Augmenting Data with Physics
04:16 Coordinates Matter!
06:38 Simulated vs Experimental Data
10:42 Big Data vs Diverse Data
12:48 Generalizing Models with Physics
16:31 Data is Expensive
17:42 Data is Biased
18:58 Rare Events
21:24 Small Signals
24:13 Galileo Dropped the Ball
27:10 Hidden Variables
29:22 Preview: Discovering Governing Equations
30:42 The Digital Twin
35:09 Outro

[mostafasayahkarajy]

Thank you very much for releasing this series on Youtube. Please continue! I am doing a phd and your videos have become my online courses.

[marktahu]

I add my thanks to the many others with regard your presentation, I feel that what you offer is suitable for every field and not restricted to physics, it allows amateurs like me to get a clear sense of the scientific process in relation to ML. Looking forward to the maths components. Very much appreciate your work.👍👍👍

[GuruKal]

This is very quickly becoming my favorite channel. Thanks for putting this stuff together! 🙏 Also, if you could touch on how new architectures are usually designed in your next video, then that would be tremendously appreciated. I see the rationale behind most of the fundamental architectures (e.g. CNNs, Autoencoders, etc.) but I’d like to know how researchers usually go about the design phase in putting together a new architecture to tackle a specific problem. Some of the things (e.g. use of residual connections) seem so random to me

[tumi]

Hi Prof. Brunton

Thanks so much for the this fantastic series,

As a question/kind request : do you plan on having a hands-on examples / tutorials for this series ?

Where we can play around with some code and datasets. I think this would go a long way in internalizing the information.

Otherwise, this is one of the best corners of educational content on the internet 👌🏿

[lewisryan]

I’m a mechanical engineering student who’s very interested in machine learning and how we can use it in the engineering world, the information on PINNs is very limited even to the point where there’s not many examples of people coding them on python or other languages.

This series has been great for me as I’m trying right now to model my own physics models .

What would be cool is an episode near the end where you go over an example of a working machine learning model going over the data you used and the results of this model

[raphango]

Really looking forward for next Thanks!

[ds]

Excellent videos by a very talented teacher👏

[wiredrabbit]

I would love to see a deep dive into the rogue wave problem. It's one thing to know that it must be accounted for, quite another to actually do it.
Thanks for the videos

[rito_ghosh]

Thanks for this video. Some really interesting insights!

But, I would have liked some code and building some models live, or did some curating of data with code, along with the required math. Like you did for the video series for your databook (Data Driven Science and Engineering by you and your colleagues).

[rito_ghosh]

Was waiting for it. Just started watching! ❤

[AI-Youtube-pjxz]

for me, AI-informed physics videos seem appealing but I believe it would be better to have a video with solid slides and intuitive examples, structured and with examples. Instead of doing it in a Webinar style. This is really an important topic. The physics notes are really informative

[Starcfd]

Great job. I'm still waiting for the digital twin video in this Playlist 😢

[splaytrees]

I would like to see some out-of-sample performance plots for baking in physics vs without. A year ago I trained the UNIT GAN to translate daily weather from one model to another. I baked physics into the discriminator, but did not see a performance boost. I think there's more to it. My suspicion was that the physics was tracking at a macro/low-res level but paled in comparison to the patterns the model discovered in the data texture; hence, the texture patterns got priority because their affect on the loss function dominated. One trick that gained some traction was dedicating a specific range of channels for training specific phenomenon, this allowed the model to learn and maintain weak patterns, while giving access to these representation channels to the model as a whole.

[Jiri_Klic]

@Eigensteve thank you very much for your content. I love your lectures and your book. Regarding machine learning with hidden variables, could you please cover the Hidden Markov Model in your series? 🙏 I understand the mathematical theory of HMM and Markov Models in general, but I struggle to grasp the principle behind finding the unknown parameters using the Viterbi or Baum-Welch algorithm. Thanks again for your time.

[vishwafuru]

Prof. Brunton, Please make a video on Kernelization.

[sen]

Dear Prof. Brunton, when you mentioned how we learned about gravity or F=ma, you mentioned that we learned by thousands of years of observation. But, instead what happened is that a 21 old boy (Newton) actually noticed it for us. So, maybe a 21 years of data collection lead to this. So, my question is, do we really need an infinite amount of data to understand any physical phenomena? Or a sufficiently large amount is enough to reach a conclusion?

[PatrickFaith]

Hmmm … by now I was hoping to see some examples in something like Jupyter notebooks.

[AnjuIITMandi]

sir please upload the third part as well design an architecture

[MatthewGale-sw]

😊 good luck teaching others something you don't really have a teaching inside yourself to teach
....😊 You have to learn a different way
😊 A different way indeed

[الاستاذ_فهد_الفعاس]

Dear professor :
Can you help me to get a studio like you have?
I am from Iraq