🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Found a very interesting pattern in ML conference reviews: for a theoretical paper, the reviewers will ask for experiments instead; for an experimental paper, the reviewers will for sure ask for more experiments, often to compare your apple to other oranges.

I always tell all my students: do not take outcome of *any* conferences seriously, no matter what others tell you. Focus on doing significant research and writing good papers. Treat conference submissions as a drill for sharpening your academic skills - that is all what they are

Flying to New Orleans to attend NeurIPS 2022 with my research group. I am so proud to present the following three exciting papers from our group! 🧵

Why are diffusion models so good? Our NeurIPS work by Dohyun Kwon , Ying Fan and Kangwook Lee presents a plausible explanation for it. 🧵

How to build *TruthGPT*? I listened to a talk by the legendary John Schulman. It's densely packed with lots of deep insight. Key takeaways: - Supervised finetuning (or behavior cloning) makes the model prone to hallucination, while RL mitigates it. - NLP is far from done! 1/🧵

1/10: The summer break is the perfect time to share recent research from my lab. Our first story revolves around a fresh interpretation of diffusion-based generative modeling by my brilliant student Ying Fan. She proposed "diffusion models are solving a control problem".

🔥Check out our ICML Conference ICML23' work on training diffusion models with policy gradient for shortcuts, which is the first work to use RL for training diffusion models to our knowledge. Check out our Arxiv paper arxiv.org/abs/2301.13362 & an exciting follow-up work coming soon!

❓ What is an effective approach for fine-tuning pre-trained t2i diffusion models using a reward function? 💡 I'm excited to share "DPOK: Reinforcement Learning for Fine-tuning Text-to-Image Diffusion Models" co-led by Ying Fan Website: sites.google.com/view/dpok-t2i-… 🧵 1/N

🔥Check out our work on training diffusion models with reinforcement learning: We show that with proper KL regularization, RL is better at obtaining both high text-image alignment and image quality than supervised fine-tuning!

DPOK: Reinforcement Learning for Fine-tuning Text-to-Image Diffusion Models propose using online reinforcement learning (RL) to fine-tune text-to-image models. We focus on diffusion models, defining the fine-tuning task as an RL problem, and updating the pre-trained

We'll present the paper tomorrow! Find us at 11am, Exhibit Hall 1 #427

🧵Four amazing presentations lined up for the final day of #ICML2023! Our group will cover topics from teaching Transformers arithmetic and iterative in-context learning to understanding weight decay and speeding up GPT! Stay tuned! (1/5)

Come to our poster at #NeurIPS to discuss about RLHF for t2i diffusion models and more! We will also share some new results compared to the preprint version in x.com/kimin_le2/stat….

two poster sessions at #NeurIPS2023 today! * #1415: "Guide Your Agent with Adaptive Multimodal Rewards" with Changyeon Kim x.com/cykim1006/stat… * #542 "DPOK: Reinforcement Learning for Fine-tuning Text-to-Image Diffusion Models" with Ying Fan x.com/kimin_le2/stat…

Check out our #NeurIPS poster #542 (Tue afternoon) on RLHF for diffusion model. TLDR; Our new method DPOK can significantly improve the text/image alignment of text-to-image models, eg #StableDiffusion neurips.cc/virtual/2023/p… Led by Ying Fan and Kimin. See you soon!

🚀 Excited to share our latest research on Looped Transformers for Length Generalization! TL;DR: We trained a Looped Transformer that dynamically adjusts the number of iterations based on input difficulty—and it achieves near-perfect length generalization on various tasks! 🧵👇

Huge thanks to UMD Center for Machine Learning for the recognition, and grateful to Kangwook Lee for being an amazing advisor!

While I couldn't make it to #NeurIPS2024 this time, Ching-An Cheng and Aditya Modi will present our work on offline contextual goal-oriented RL @ West Ballroom A-D #6206 on Thu (Poster Session 3 West)! Also check our paper here: arxiv.org/abs/2408.07753.

Notes: - Two models, R1-Zero (V3-Base + RL, no SFT), R1 (SFT [CoT from R1-Zero] -> RL [reasoning] -> SFT [general] -> RL [alignment, reasoning]) - Six distillation models, i.e., SFT from R1 on Qwen, Llama. Outperforms RL-only on those models, RL on distilled models would improve

Excited to share that our paper Looped Transformers for Length Generalization (arxiv.org/abs/2409.15647) has been accepted at #ICLR2025! 🎉 Feel free to stop at Poster Session 3 (Fri 25 Apr 10 a.m. +08 — 12:30 p.m. +08) in Hall 3 + Hall 2B #475