Wei Hu (@weihu_) 's Twitter Profile
Wei Hu

@weihu_

Assistant professor @UMich; previously @UCBerkeley @Princeton @Tsinghua_Uni. Building the theoretical and scientific foundations of deep learning.

ID: 2716076245

linkhttp://weihu.me calendar_today08-08-2014 03:40:48

139 Tweet

2,2K Followers

1,1K Following

Behnam Neyshabur (@bneyshabur) 's Twitter Profile Photo

💡💡What is the best acc an MLP can get on CIFAR10❓ 65%❓ No, 85%‼️ Trying to understand convolutions, we look at MDL and come up with a variant of LASSO that when applied to MLPs, it learns local connections and achieves amazing accuracy! Paper: arxiv.org/abs/2007.13657 1/n

💡💡What is the best acc an MLP can get on CIFAR10❓

65%❓ No, 85%‼️

Trying to understand convolutions, we look at MDL and come up with a variant of LASSO that when applied to MLPs, it learns local connections and achieves amazing accuracy!

Paper: arxiv.org/abs/2007.13657

1/n
Wei Hu (@weihu_) 's Twitter Profile Photo

I'm thrilled to be one of the #SiebelScholars Class of 2021! Many thanks to Siebel Scholars for creating this great community, and congrats to everyone on the recognition! #SiebelClassOf2021

Sebastien Bubeck (@sebastienbubeck) 's Twitter Profile Photo

Starting tomorrow (with livestream): Simons Institute workshop on *Learning and Testing in High Dimensions*. We have a great line-up of talks, featuring many of the recent exciting results in high-dimensional learning! simons.berkeley.edu/workshops/sche…

Jenn Wortman Vaughan (@jennwvaughan) 's Twitter Profile Photo

My fully-vaccinated husband came down with a fever on Friday. On Saturday he tested positive for COVID. The past few days he's had headaches, muscle aches, fatigue, appetite loss, a cough... the whole gamut of COVID symptoms. But it could be worse. [1/n]

Aleksander Madry (@aleks_madry) 's Twitter Profile Photo

ImageNet is the new CIFAR! My students made FFCV (ffcv.io), a drop-in data loading library for training models *fast* (e.g., ImageNet in half an hour on 1 GPU, CIFAR in half a minute). FFCV speeds up ~any existing training code (no training tricks needed) (1/3)

ImageNet is the new CIFAR! My students made FFCV (ffcv.io), a drop-in data loading library for training models *fast* (e.g., ImageNet in half an hour on 1 GPU, CIFAR in half a minute).
FFCV speeds up ~any existing training code (no training tricks needed) (1/3)
Zeyuan Allen-Zhu, Sc.D. (@zeyuanallenzhu) 's Twitter Profile Photo

MSR Asia Theory Center (Beijing) is recruiting! A lovely place where I had a wonderful year of internship and published my first theory paper. Website: microsoft.com/en-us/research… and application link: careers.microsoft.com/us/en/job/1241…. They may also consider full-time for strong candidates.

Rada Mihalcea (@radamihalcea) 's Twitter Profile Photo

“Nearly half of the [Michigan AI] lab’s faculty are women, much higher than the global average of 16% reported in the 2021 AI Index Report.” We know a community of diverse viewpoints is essential for the growing world of #AI. So so proud MichiganAI! cse.engin.umich.edu/stories/divers…

XY Han (@xyhan_) 's Twitter Profile Photo

Neural collapse observes last-layer class variation collapses 𝘵𝘰𝘸𝘢𝘳𝘥𝘴 0 with training. 𝗕𝘂𝘁: As it does, one can 𝘴𝘵𝘪𝘭𝘭 find informative, fine-grained structures in the residual small variations at 𝘧𝘪𝘹𝘦𝘥 epochs (even ones that look “collapsed”!). Check out this

MichiganAI (@michigan_ai) 's Twitter Profile Photo

Looking forward to tomorrow's #AI Seminar w/ Samuel Marks Samuel Marks! Samuel Marks will introduce methods for discovering and applying sparse feature circuits: cse.engin.umich.edu/event/sparse-f…

Looking forward to tomorrow's #AI Seminar w/ Samuel Marks <a href="/saprmarks/">Samuel Marks</a>!

<a href="/saprmarks/">Samuel Marks</a> will introduce methods for discovering and applying sparse feature circuits:
cse.engin.umich.edu/event/sparse-f…
Wei Hu (@weihu_) 's Twitter Profile Photo

Check out our new LLM quantization algorithm that is extremely fast, requires minimal calibration data, and enables flexible bit allocation! Led by Yongyi Yang