Learning, Reasoning, and Planning with Neuro-Symbolic Concepts

This lecture from the Allen School Colloquia Series features MIT PhD student Jiayuan Mao presenting her research on neuro-symbolic concepts for artificial intelligence. Explore a framework for building intelligent agents that can continually learn, reason, and plan using compositional abstractions of the physical world. Discover how neuro-symbolic concepts—represented through symbolic programs and modular neural networks—enable superior data efficiency, faster reasoning and planning, and strong generalization capabilities. The presentation demonstrates applications in visual reasoning across 2D/3D environments, motion analysis, video data, and decision-making tasks in both virtual and robotic contexts. Mao, recognized as a Rising Star in EECS and Generative AI (2024) with multiple Best Paper awards, shares insights from her research at the intersection of visual reasoning, robotic manipulation, scene understanding, and language acquisition.