CMU Artificial Intelligence Seminar Series ^{sponsored by}

Tuesday, Apr 23, 2024

Time: 12:00 - 01:00 PM ET. NSH 3305

Yexiang Xue -- Vertical Reasoning Enhanced Learning, Generation and Scientific Discovery

Abstract: Automated reasoning and machine learning are two fundamental pillars of artificial intelligence. Despite much recent progress, the full integration of reasoning and learning is beyond reach. This talk presents three cases where integrated vertical reasoning significantly enhances learning. Our first case is in neural generation, where state-of-the-art models struggle to generate pleasing images while satisfying complex specifications. We introduce Spatial Reasoning INtegrated Generator (SPRING). SPRING embeds a spatial reasoning module inside the deep generative network to reason object locations. This embedded approach guarantees constraint satisfaction, offers interpretability, and facilitates zero-shot transfer learning. Our second case is in AI-driven scientific discovery, where we embed vertical reasoning to expedite symbolic regression. Vertical reasoning builds from reduced models that involve a subset of variables (or processes) to full models, inspired by the human scientific approach. Vertical discovery outperforms horizontal ones at discovering equations involving many variables and complex processes, especially in learning PDEs in computational materials science. In the third case, we demonstrate that vertical reasoning enables constant approximation guarantees in solving Satisfiable Modulo Counting (SMC). SMC involves model counting as predicates in Boolean satisfiability. It encompasses many problems that require both symbolic decision-making and statistical reasoning, e.g., stochastic optimization, hypothesis testing, solving quantal-response leader-follower games, and learning (inverse reinforcement learning) with provable guarantees. Using vertical reasoning that streamlines XOR constraints, our proposed XOR-SMC reduces highly intractable SMC problems into solving satisfiability instances, while obtaining a constant approximation guarantee.