Irena Gao
@irena_gao
Followers
746
Following
231
Media
5
Statuses
67
Explore trending content on Musk Viewer
ميلان
• 1226913 Tweets
BRICS
• 283516 Tweets
Öcalan
• 269410 Tweets
Özgür Özel
• 78700 Tweets
Ümit Özdağ
• 70603 Tweets
NEVER BORED ON NINGNING DAY
• 57510 Tweets
#النصر_الاستقلال
• 53873 Tweets
Eminem
• 52773 Tweets
Blade
• 52702 Tweets
Valdivia
• 52516 Tweets
桜井政博
• 49100 Tweets
Siya
• 42690 Tweets
Celtics
• 39942 Tweets
Dortmund
• 36612 Tweets
Bicol
• 32627 Tweets
APO PİÇTİR PİÇ KALACAK
• 29415 Tweets
Türkiye Türk'tür
• 25957 Tweets
桜井さん
• 23837 Tweets
Knicks
• 19543 Tweets
Yoru
• 17372 Tweets
TXT YOU ARE LOVED
• 14430 Tweets
Yunus Akgün
• 13677 Tweets
#راكان_السهلي_في_شاعر_الرايه
• 11867 Tweets
Pulisic
• 10573 Tweets
Denji
• 10398 Tweets
Applying to Stanford's CS PhD program? Current graduate students are running a SoP + CV feedback program for URM applicants (broadly defined). Apply to SASP by Oct. 25! Info:
3
98
475
@shiorisagawa
and I will be at the ICML 11am poster session tomorrow, Wednesday, presenting "Targeted Augmentations Improve Out-of-Domain Robustness." () 🏝️Come chat with us about how to design data augmentations for robustness!
0
9
52
Excited to be at
#ICCV2023
this week presenting our paper "Adaptive Testing of Computer Vision Models."
Work w/ the incredible
@gabriel_ilharco
,
@scottlundberg
, and
@marcotcr
.
Oral presentation: Wednesday 13:30-14:30
Poster session: Wednesday 14:30-16:30
2
7
47
We're excited to announce a new release of OpenFlamingo models. Check out the blog post and keep an eye out for our technical report!
We are excited to announce OpenFlamingo V2 🦩! We are releasing five new multimodal models, across the 3B, 4B, and 9B scales, that outperform our previous model. w/
@irena_gao
.
Repo:
Demo:
Blog:
13
144
659
0
2
27
Check out our benchmark evaluating the compositionality of CLIP-like models!
Have vision-language models achieved human-level compositional reasoning? Our research suggests: not quite yet.
We’re excited to present CREPE – a large-scale Compositional REPresentation Evaluation benchmark for vision-language models – as a 🌟highlight🌟at
#CVPR2023
.
🧵1/7
3
14
68
0
2
12
To read more, please check out our paper at .
And if you're at ICCV this week, I'd love to talk more at the poster session!
0
1
5
Vision models can have unexpected failure modes. For example, object detection models identify stop signs in sunny environments, but can fail when the stop signs are covered in snow.
How can we efficiently test models for such bugs before deployment?
1
0
3
Rather than having users generate tests on their own, we use foundation models (e.g. CLIP and GPT-3) to rapidly propose tests. Throughout the testing process, we also adapt these models to learn from previously found failures.
1
0
3
We find in user studies that AdaVision helps users find significantly more failures than a non-adaptive baseline. This holds across classification, detection, and captioning tasks.
We also show that finetuning ViT-H/14 on AdaVision tests fixes behavior on buggy topics.
1
0
3
Human-in-the-loop testing (e.g. red teaming) enables open-ended bug discovery, but is slow and limited by user creativity. How can we address these issues?
We propose Adaptive Testing for computer vision models (AdaVision)!
1
0
2
Unfortunately, testing for bugs is hard. Most evaluation sets are small and underestimate OOD vulnerabilities. Automatic testing methods, like suites of synthetic perturbations, are fast but close-ended: they can only test along a set of pre-defined axes.
1
0
2
@dailyaibrief
ICCV registrants should be able to view a livestream in the Imagina app, under the "Paris Nord" room.
0
0
0
@JKhindkar
We only read PhD applications, unfortunately. I'm not aware of a similar program for the MS, but I could be wrong. Good luck!
1
0
1