[1/2] #ICLR2024 Vision-language models still struggle with open-domain visual concepts, while finetuning without proper regularization tends to overfit. We introduce OGEN to improve OOD GENeralization of finetuned models by synthesizing OOD features and adaptive self-distillation
Our paper on uncertainty estimation for offline RL by our Intern Yue Wu (https://t.co/aZ0rGJOwov) has recently been accepted to #ICML2021. It's a simple, yet intuitive method that results in performance gains and training stability. Check it out!
New work on Geometric Capsules: Learning to group 3D points into parts & parts into the whole object in unsupervised way. Each capsule represents a visual entity consisting of a pose & feature representing "where" & ''what'' it is.
https://t.co/czOjsenGpO
w/t @nitishsr & @Hanlin
Check out our latest work on modeling the "what's" and "where's" of objects and parts with geometric capsule representations from 3D point clouds. https://t.co/B6Xd7rarq9 (with @nitishsr & @rsalakhu)... I'm at #NeurIPS2019 all week if you want to chat about it!