Probabilistic and Relational Inferences in Dynamic Environments with Compositionally Organized Learning To Reason About Novel Experiences (PRIDE with COLTRANE)

Current AI systems are brittle to novel changes in the world. This limits their applicability, leads to large engineering efforts to create robustness in spite of the brittleness, and is a cause of unpredictable failures. Both knowledge-based, first-wave AI and data-driven, second-wave AI techniques have been unable to overcome this brittleness. Charles River Analytics and MIT’s Computational Cognitive Science and Probabilistic Computing groups are developing a new, third-wave approach that merges symbolic and sub-symbolic reasoning for effective operation in open worlds. This approach builds on emerging technologies that combine probabilistic programming (PP) and symbolic reasoning to create PRIDE with COLTRANE (Probabilistic and Relational Inferences in Dynamic Environments with Compositionally Organized Learning To Reason About Novel Experiences). In contrast to traditional symbolic reasoning systems, PRIDE with COLTRANE recognizes novelty, synthesizes new representations automatically, and reasons using the new representations to adapt in real-time to novel situations. Unlike traditional learning systems that require large amounts of training data for every situation, PRIDE with COLTRANE uses its knowledge structures to learn about new situations from a very small set of instances and applies the structures again to adapt on the fly.