Molei Tao
@MoleiTaoMath
Followers
1K
Following
1K
Media
13
Statuses
295
Georgia Tech Prof; Tsinghua, Caltech, NYU Courant * deep learning theory * (diffusion) generative model, probabilistic ML * AI4Science * applied & comput. math
Joined October 2021
What is variational optimization?. Why can continuous dynamics help? . Optimization is already a profound field, what can it bring in?. Check out blog Comment/Retweet/Like will be deeply appreciated!. 1/6.
1
42
152
Great work by my former student Yuqing Wang and her amazing collaborator @ShangdingG95714 !.I'm not part of this work, but I hope you wouldn't mind this proud advisor moment.
0
0
0
When analyzing training dynamics, less work focused on the role of (real) data. rigorously shows that uniformality of data is a key to accelerated training, via a new framework beyond NTK. Theoretical predictions are verified by LLaMA-1-13B experiments.
1
16
79
RT @BachFrancis: Big thanks to the COLT 2025 organizers for an awesome event in Lyon! Here are the slides from my keynote this morning in c….
0
21
0
Kudos to @YeHeMath and @KevRojas1499 !. @YeHeMath also wrote better tweet about it with more details. Happy to discuss with anyone interested!.
0
0
4
Analysis of Classifer Free Guidance for masked diffusion: * 1-token conditional generation is exact, unlike continuous CFG (not sampling the tilted distribution).* guidance suppresses the overlap between classes.* multi-token generation differs from single.
1
20
120
Fantastic reading, written by Prof. Damek Davis.
i wrote some notes on GPUs. > how are they organized, what are the bottlenecks, how to measure and increase performance. they're based on what i learned reading @cHHillee and @Si_Boehm's blog posts. i wrote this just so I don't forget what i read.
0
2
10
As some asked (thank you!) -. Title: Implicit Biases of Large Learning Rates in Machine Learning. Time: Fri 6/20 2-3:30pm. Location: Conference Room A (3F), Seoul AI Hub, 108 Taebong-ro, Seocho-gu, Seoul.
0
0
10
Visiting Seoul, Korea and honored to give a National AI Research Lab Invited Seminar on large learning rates. Don't hesitate to let me know if you'd like to chat!.
1
1
28
Big shout out to my amazing students and collaborators @YuchenZhu_ZYC , @KevRojas1499 , @SichenZhu & Felix Ye. @YuchenZhu_ZYC and @KevRojas1499 also wrote much better tweets about it!. Care to chat? Feel free to DM/email any of us!.
0
0
3
Generative modeling data with multiple modalities (e.g.continuous,discrete,manifold,constrained)?. ppl often tokenize everything into 1 modality->use AR transformer. Want an encoder-free native-multimodal diffusion model? #icml2025 is a general approach.
9
31
166
RT @mathNAb: Durastante, Gnazzo, Meini: A Riemannian Optimization Approach for Finding the Nearest Rev. https://t….
0
2
0
Will be at #ICLR2025 and love to chat about any of our 5 papers, or .probabilistic ML / diffusion model,.deep learning theory / optimization, and .AI4Science in general. Feel free to DM!
3
7
90
RT @YuanqiD: Scientific Knowledge Emerges in LLMs and YOU CAN Access It (via sampling)! . 🔥🔥🔥New blog to summarize what we have learned fro….
0
17
0
BTW: does anyone know how to fix this:. We knew our paper was cited by more than a handful of papers arXiv'ed months ago, but google scholar still shows 0 citation. Could it be a bug?.
0
0
3
New blog after only 2 years - big thanks to my excellent student @yuchen4975 for the lead. Blog is based on one of our ICLR'25 papers. Happy to talk to anyone at Singapore. Plz just DM!.
1
0
7
If you know data live on a manifold, you can hardwire this prior knowledge in diffusion model to make generation more accurate & data efficient. What if there is also a group structure, like in protein design & quantum problem? Use it to do even better -.
3
20
149
I'm trying to compile a reading list for math graduate students, on existing/popular theories of generalization in deep learning. I can think of tens of beautiful research articles, but any book / survey you would recommend?. Any suggestion would be appreciated!.
2
4
45
