Compositional Learning for Object-centric Attributes and Relations

In this talk, I will cover my group's recent works on compositional learning for object-centric attributes and relations. We will briefly discuss learning disentangled representations for objects and attributes, and our efforts to create large-scale datasets and scalable approaches for open vocabulary recognition. I will also highlight some of our latest contributions in compositional image generation.Furthermore, I will offer a glimpse into other research areas we work on, including concept discovery, diffusion-based generative models, neural representations for videos, and predictive perception and robotics.