Our MIT class “6.S184: Introduction to Flow Matching and Diffusion Models” is now available on YouTube!
We teach state-of-the-art generative AI algorithms for images, videos, proteins, etc. together with the mathematical tools to understand them.
Flow and diffusion models are mathematically demanding subjects - which is why many lectures restrict themselves to teaching high level intuition. Here, we give a mathematically rigorous and self-contained introduction yet aimed at beginners in AI. We hope you will like it!
By the way, I was trying to go through the MIT Optics [1] course, but the audio/video quality is ... terrible. Could somebody fix that? (Maybe with diffusion models? ;)
I can put this course on our radar to put up (as part of the work soul.mit.edu does).
One difficulty is that it's not usually taught in a room with an automatic video recording setup, so it's not that easy, in terms of logistics, to get the course recorded. But I'll see what we can do next fall (it's taught in the fall).
> Flow and diffusion models are mathematically demanding subjects - which is why many lectures restrict themselves to teaching high level intuition.
I appreciate this - I hope norms develop to clearly identify whether learning materials / courses are about intuition or deeper applications that don't shy away from full prerequisites. They both have their place, but can be hard to find the latter amidst the sea of introductory materials that merely give intuition.
Awasemoness! Since my undergrad days (it was a long time ago), I’ve found myself turning to MIT classes to help me grasp some challenging subjects. It's been such a great resource!
Conditional normalizing flows are one of the most beautiful solutions to inverse design problems that I’ve come across, if you have the data to train them. Something about the notion of carefully deforming a base distribution by pushing and pulling its probability mass around until it’s in the right location by using bijective functions (which themselves have very clever constructions) is just so elegant…
I’ve had some trickiness trying to get them to work when some of the targets are continuous and some categorical, but regardless just a really cool method… really nailed it on the name imo!
Cool course, can't wait to go through it! I noticed that this is focused strictly on continuous spaces, but there's a lot of cool stuff going on in discrete diffusion. Any plans for a follow up? I couldn't help but notice that the course teacher Peter just came out with a paper for discrete diffusion too.
I'm incredibly grateful for MIT OCW and consorts. I've been using it as a secondary resource for my subjects and learning about the same topic in two different ways is incredibly helpful, especially hard to grasp ones.
Would one of you, who is familiar with this topic, help me understand the primary use case(s) along with a few words, just your overall take on these techniques?
It’s the fundamentals that underly Stable Diffusion, Dalle, and various other SOTA image generation models, video, and audio generation models. They’ve also started taking off in the field of robotics control [1]. These models are trained to incrementally nudge samples of pure noise onto the distributions of their training data. Because they’re trained on noised versions of the training set, the models are able to better explore, navigate, and make use of the regions near the true data distribution in the denoising process. One of the biggest issues with GANs is a thing called “mode collapse” [2].
Our MIT class “6.S184: Introduction to Flow Matching and Diffusion Models” is now available on YouTube!
We teach state-of-the-art generative AI algorithms for images, videos, proteins, etc. together with the mathematical tools to understand them.
Flow and diffusion models are mathematically demanding subjects - which is why many lectures restrict themselves to teaching high level intuition. Here, we give a mathematically rigorous and self-contained introduction yet aimed at beginners in AI. We hope you will like it!
From: https://x.com/peholderrieth
Thanks!
By the way, I was trying to go through the MIT Optics [1] course, but the audio/video quality is ... terrible. Could somebody fix that? (Maybe with diffusion models? ;)
[1] https://ocw.mit.edu/courses/2-71-optics-spring-2009/resource...
I can put this course on our radar to put up (as part of the work soul.mit.edu does).
One difficulty is that it's not usually taught in a room with an automatic video recording setup, so it's not that easy, in terms of logistics, to get the course recorded. But I'll see what we can do next fall (it's taught in the fall).
Youtube link to playlist: https://www.youtube.com/watch?v=GCoP2w-Cqtg&list=PL57nT7tSGA...
> Flow and diffusion models are mathematically demanding subjects - which is why many lectures restrict themselves to teaching high level intuition.
I appreciate this - I hope norms develop to clearly identify whether learning materials / courses are about intuition or deeper applications that don't shy away from full prerequisites. They both have their place, but can be hard to find the latter amidst the sea of introductory materials that merely give intuition.
Thank you, for making the effort of making this so accessible! Danke :)
Awasemoness! Since my undergrad days (it was a long time ago), I’ve found myself turning to MIT classes to help me grasp some challenging subjects. It's been such a great resource!
Conditional normalizing flows are one of the most beautiful solutions to inverse design problems that I’ve come across, if you have the data to train them. Something about the notion of carefully deforming a base distribution by pushing and pulling its probability mass around until it’s in the right location by using bijective functions (which themselves have very clever constructions) is just so elegant…
I’ve had some trickiness trying to get them to work when some of the targets are continuous and some categorical, but regardless just a really cool method… really nailed it on the name imo!
Cool course, can't wait to go through it! I noticed that this is focused strictly on continuous spaces, but there's a lot of cool stuff going on in discrete diffusion. Any plans for a follow up? I couldn't help but notice that the course teacher Peter just came out with a paper for discrete diffusion too.
https://x.com/peholderrieth/status/1891846309952282661
https://github.com/kuleshov-group/awesome-discrete-diffusion...
Does anyone have a collection of all public courses on latest AI techniques?
Just start an "awesome AI courses" repo on GitHub and invite PRs. Or update these:
https://github.com/luspr/awesome-ml-courses
https://github.com/owainlewis/awesome-artificial-intelligenc...
This
I'm incredibly grateful for MIT OCW and consorts. I've been using it as a secondary resource for my subjects and learning about the same topic in two different ways is incredibly helpful, especially hard to grasp ones.
I'm so happy to find this here. LLMs seem to have diverted a lot of attention away from this incredibly useful technique.
Would one of you, who is familiar with this topic, help me understand the primary use case(s) along with a few words, just your overall take on these techniques?
Thanks and appreciated in advance.
It’s the fundamentals that underly Stable Diffusion, Dalle, and various other SOTA image generation models, video, and audio generation models. They’ve also started taking off in the field of robotics control [1]. These models are trained to incrementally nudge samples of pure noise onto the distributions of their training data. Because they’re trained on noised versions of the training set, the models are able to better explore, navigate, and make use of the regions near the true data distribution in the denoising process. One of the biggest issues with GANs is a thing called “mode collapse” [2].
[1] https://www.physicalintelligence.company/blog/pi0
[2] https://en.wikipedia.org/wiki/Mode_collapse
Thank you
Thanks again. Past decade has been golden era for deep learning education. I love the fights of who will make high quality learning content free
Great for MIT to be putting out such timely and relevant content for free!
Thank you so much, what other OCW courses exist on modern AI?
soul.mit.edu is putting some courses up. You can find another course on diffusion here (https://mitsoul.org/courses/mit/course-6/6-S185/), and a course on data-centric AI here (https://mitsoul.org/courses/mit/course-6/6-DCAI/).
This is exactly what I was looking for! Thanks for sharing
Well done, folks. Congrats!
Nice