Title: Generative Factorization For Object-centric Representation Learning
Abstract: Human understanding of the natural world is built upon the understanding of lower-level concepts/entities, which allows humans to handle the seemingly infinite expressions of the natural world. Inspired by this, object-centric representation learning has emerged as a promising approach towards machine intelligence, facilitating high-level reasoning and control from visual sensory data. It aims at discovering compositional structures around objects from the raw sensory input data, i.e. a binding problem, where the segregation (i.e. factorization) is the major challenge, especially in cases of no supervision. Towards this, we consider object-centric representation inference as the inverse problem of an observation generation problem and solve the neural scene factorization problem in a generative framework. In this talk, I will present two of our recent works in this direction where one focus on factoring the spatial structure of 3D scenes and the other focus on learning latent representations of dynamic scenes as functions of space and time.