Thanks AK for sharing our work. We propose Peekaboo - a training-free method for incorporating spatio-temporal control in video diffusion models. Work done in collaboration with Yash Jain Harkirat Behl Vibhav Vineet . Longer thread + code coming soon!

Another really cool work from our team, this one led by Harkirat Behl !

Come check out our spotlight work at Room 252 - Room 254 (Level 2) and let's talk more about fairness and distribution shifts :) Amazing conversations at the morning poster session AFME 2024 See you at the talk at 13:00 and the afternoon poster at 16:50 to 1730! #NeurIPS2023

Extending an LLM for new knowledge sources is tedious—fine-tuning is expensive/causes forgetting, LoRA is restrictive. Excited to share our work where we show that an LLM can be efficiently *composed* with specialized (L)LMs to enable new tasks! arxiv.org/abs/2401.02412 🧵(1/8)

While large-scale models are pushing the boundaries of video generation, spatio-temporal control of their outputs is still a problem. Peekaboo (now accepted at #CVPR2024) provides a simple plug-and-play solution to tackle this! 🔗- jinga-lala.github.io/projects/Peeka…

📢 New Paper! Ever wondered why transformers are able to capture hierarchical structure of human language without incorporating an explicit 🌲 structure in their architecture? In this work we delve deep into understanding hierarchical generalization in transformers. (1/n)

Looking forward to presenting this work at #ICLR2024 next week in Vienna! 🇦🇹 Please stop by our poster on 8th (10:45am) if you are interested in efficient, modular, decentralized development of large models!

Waiting for my timeline to be flooded with NeurIPS preprints

🚨 New Paper 🚨 Evaluations can have a lot of variance, throwing off model comparisons especially during pre-training. In our latest work, “Quantifying Variance in Evaluation Benchmarks”, we explore this phenomenon in depth. A thread [1/n]

Will training on AI-generated synthetic data lead to the next frontier of vision models?🤔 Our new paper suggests NO—for now. Synthetic data doesn't magically enable generalization beyond the generator's original training set. 📜: arxiv.org/abs/2406.05184 Details below🧵(1/n)

How can we align foundation models with populations of diverse users with different preferences? We are excited to share our work on Personalizing RLHF using Variational Preference Learning! 🧵 📜: arxiv.org/abs/2408.10075 🌎: weirdlabuw.github.io/vpl/