Kyle Hsu (@kylehkhsu) header
Kyle Hsu (@kylehkhsu) profile picture

research @openai. prev @tesla_optimus, phd @stanfordailab, basc @uoft engsci

🇹🇼🇨🇦🇺🇸
Joined September 2012
Is it possible to obtain zero-shot generalization of vision-based RL agents without data augmentation, task-centric representations, etc? Yes! by disentangling the latent space and, you might’ve heard about this given the latest news, associative memory i.e. Hopfield models!🧵
5
27
2
163
26,510
After two years, it is my pleasure to introduce “DROID: A Large-Scale In-the-Wild Robot Manipulation Dataset” DROID is the most diverse robotic interaction dataset ever released, including 385 hours of data collected across 564 diverse scenes in real-world households and offices
6
80
45
298
119,353
[1/6] Excited to share “RLVF: Learning from Verbal Feedback without Overgeneralization” Our method C3PO fine-tunes an LLM from 1 sentence of feedback and decreases overgeneralization (=applying the feedback when it should not be applied). Details: austrian-code-wizard.github.…
2
17
6
88
71,176
As new VLMs get released everyday, its easy to get lost in the myriad of details and design choices each one adopts. With Prismatic VLMs we try an cut through the noise - presenting a modular VLM training codebase that enables rigorous investigation of the VLM design space.
What design choices matter when developing a visually-conditioned language model (VLM)? Check out our paper – Prismatic VLMs – and open-source training code, evaluation suite, and 42 pretrained VLMs at the 7B-13B scale! 📜 arxiv.org/abs/2402.07865 ⚙️ + 🤗 github.com/TRI-ML/prismatic-…
3
1
13
3,405
Introducing our new work @corl_conf 2023, a novel brain-robot interface system: NOIR (Neural Signal Operated Intelligent Robots). Website: noir-corl.github.io/ Paper: arxiv.org/abs/2311.01454 🧠🤖
18
170
70
716
287,804
Kyle Hsu retweeted
How can we build a modular and compositional system that understands 3D scenes? Excited to introduce our @CVPR paper — NS3D: Neuro-Symbolic Grounding of 3D Objects and Relations, w @maojiayuan and @jiajunwu_cs . Check out our poster next week at Tue-AM-249.
1
26
97
17,415
🎠🎠 We're presenting MagicPony at CVPR! It turns a photo of an animal into a fully articulated 3D model ready for animation. It is trained using just online images, without any 3D data, keypoints or template shapes. 3dmagicpony.github.io/ Drop by our poster on Wed morning!
9
70
6
343
72,184
Abstaining from prediction is often better than making up an answer, especially in safety-critical scenarios. Can we help models know what they don’t know? We tackle this problem in our new paper, "Conservative Prediction via Data-Driven Confidence Minimization"
4
22
2
132
30,851
Our new paper on RL From Human Feedback is out: arxiv.org/abs/2305.18290. In Direct Preference Optimization (DPO) we reparameterize the reward model in a suitable way without any loss in generality and optimize the EXACT RLHF objective directly with a simple classification loss.
2
29
1
119
22,846
Hong Kong protests are in their fifth month despite an escalating crackdown. How? One surprising answer: the fate of China's Uyghurs. Many talked to me about it. They watched and learned. They've decided that they "may as well go down fighting." My latest. theatlantic.com/internationa…
133
5,103
309
6,978
Had a great time speaking on the student researcher panel today at the AI Squared Forum!
1
9
How an elite university research center concealed its relationship with a sex offender—documents show MIT Media Lab accepted donations directed by Jeffrey Epstein far in excess of what the university has admitted to, and worked to cover it up: newyorker.com/news/news-desk…
667
9,212
1,552
19,128
With the #Raptors about to hit the global stage, I’m going to share a story about what makes Toronto, and Canada, so special. It’s about a guy you’re going to be seeing a whole lot of. A Thread. [Hint: It's about racism, sports, and why you'll end up cheering for 🍁]
1,350
25,073
8,133
61,665
Meta-learning replaces the problem designing efficient algorithms with the problem of making meta-training tasks. But this is still hard. Can we generate tasks automatically instead? @kylehkhsu will be presenting Unsupervised Meta-Learning Wed at 11 am at #ICLR! w/ @chelseabfinn
29
1
128
The Max Planck Society (@maxplanckpress) is canceling its Elsevier subscriptions. The current agreement expires at the end of this month and will not be renewed. mpdl.mpg.de/en/505
15
386
72
764
my favorite thing in the world is responding to my hateful messages and seeing how fast i can change their mood.
811
37,244
2,423
208,784
Kyle Hsu retweeted
Dear @JustinTrudeau @VectorInst @MILAMontreal @CIFAR_News, How can Canada lead in AI if we cannot even process visas for top AI researchers gives MONTHS of advance warning? Please RT.
7
107
7
347
"Learning to Learn with Gradients", dissertation by @chelseabfinn. This is a great reference for MAML, with the recent extensions and applications! www2.eecs.berkeley.edu/Pubs/…
16
50
CACTUs: an unsupervised learning algorithm that learns to learn tasks constructed from unlabeled data. Leads to significantly more effective downstream learning & enables few-shot learning *without* labeled meta-learning datasets arxiv.org/abs/1810.02334 w/ @kylehkhsu, @svlevine
113
3
438