Object-Oriented Learning (OOL): Perception, Representation, and Reasoning

International Conference on Machine Learning (ICML)

Friday July 17, 2020, Virtual Workshop

Access the virtual workshop page

Library learning for structured object concepts

The ability to represent semantic structure in the environment --- objects, parts, and relations --- is a core aspect of human visual perception and cognition. As a testament to this ability, humans are capable of expressing rich conceptual knowledge in simple iconic images. Here we leverage recent advances in program synthesis to develop a self-supervised and data-efficient algorithm for learning the the part-based structure of objects, as represented by graphics programs. Our algorithm iteratively learns higher-order subroutines that can be used to more compactly represent these objects, exploiting commonalities between learned subroutines to infer a part-based grammar. Our experiments explore how the resulting library of learned subroutines is jointly determined by the data distribution and the cost of learning additional subroutines.