Pinned
✨New blog post✨: my attempt as a vision researcher at finally understanding RLHF -- a deep dive into PPO & DeepSeek's GRPO!
No hot take, I promise.
Yuge Shi (Jimmy)
694 posts
Yuge Shi (Jimmy)
@YugeTen
London, England
Joined December 2017
- Dear conference organisers: I promise I will choose a better birth place next time, but in the meantime can we please 1) host conferences at more visa-friendly locations or 2) leave more time between paper decision time and conference so we have enough time to apply for visas?
- I am starting work at @DeepMind as a research scientist tomorrow ☀️☀️ I will be joining the Openendedness team led by Tim Rocktäschel @_rockt. Excited (and equally nervous, admittedly) to start this new chapter of life and meet & work with amazing people!
- Successfully defended my thesis with Max Welling (@wellingmax) and Andrew Zisserman today! No professors were physically harmed in the process. A huge thank you to both for the insightful discussions!
- Replying to @YugeTenNot everyone can get into the US by taking 20 minutes to fill out an online form, some has to dig up their entire employment/travelling history and refresh the visa appointment website desperately for days and days
- After much procrastination, finally uploaded my first blog post "Gaussian Processes, not quite for dummies"! Have a lookie here: yugeten.github.io/posts/2019/09/… Share if like :)
- Paper + Code release! We propose Fish🐟, an effective algorithm for domain generalisation. It learns invariant features by maximising gradient inner product across domains. 📜:arxiv.org/abs/2104.09937 👩💻:github.com/YugeTen/fish Work done during internship at FAIR with @syhw. 🧵
- Pretty pissed off about not getting to ICML for the third time because of visa, but super proud to have 2 orals, 1 spotlight and 1 workshop at the conference! I'm not going to fix open borders with this tweet but I do hope it nudges you to go talk to my amazing collaborators --
- Random pixel mask is NOT random word mask in image space: we propose ADIOS, which learns semantically meaningful masks for masked image modelling in a self-supervised manner. Joint work with Adam Kosiorek (@arkosiorek), now accepted to #ICML2022. 📜: arxiv.org/abs/2201.13100
- Compared to the much anticipated unveiling of #Veo3 I am more excited to share my usage of “no cap” in generating a rapping video
00:00 - The school of informatics at the University of Edinburgh and DeepMind are offering an ML PhD scholarship for students who identify as gender/racial/ethnic minorities in 2022/23. See thread for details. (1/n)
- New blog post: How I learned to stop worrying and write ELBO in a billion ways. I talk about a list of interesting approaches including IWAE and DReG that are important to the evolution of ELBO over the past few years. Have fun reading, RT if like!
- Dare I say...computer vision with human feedback is on its way in? Super happy to have contributed to this important piece of work during my internship at Google Brain!Vision meets RL! We reveal that policy gradient can be used for tuning vision models to optimize complex metrics, such as mAP, PQ or “color diversity”, observing large performance boosts on tasks like object detection, panoptic segmentation, etc. arxiv.org/abs/2302.08242
- How robust are unsupervised representation learning methods (e.g. SSL) to distirbution shift compared to supervised learning? 𝐒𝐡𝐨𝐫𝐭 𝐚𝐧𝐬𝐰𝐞𝐫: Quite! 𝐋𝐨𝐧𝐠 𝐚𝐧𝐬𝐰𝐞𝐫: Our #ICLR2023 paper arxiv.org/pdf/2206.08871… Joint work with Imant Daunhawer & @AmartyaSanyal.
