SGER: Computational Basis for Cognition: Concept Learning for Vision and Planning

This research is intended to result in a novel approach to concept learning. The basic premise of this project is that intelligent systems ought to be able to learn concepts autonomously through a combination of learning methods (e.g., boosting, statistical analyis and reinforcement learning) without requiring extensive hand-engineering (e.g., knowledge representation, concept taxonomies). The hypothesis to be explored in this project is that concept formation and related tasks such as classification can be accomplished by computationally exploitating underlying domain structure using methods that encourage parsimony and compactness of system modules that correspond to concepts. The long term goals for this line of research include better understanding of the nature of systems that underlie and embody symbols and the discovery and demonstration of methods that enable such systems to be created through a combination of programming and autonomous learning. Nearer term project goals include exploring how to harness what we know about learning in both machines and humans, understanding the role of experience vs. "innate" structures, and better understanding the role of credit assignment, focus of attention, and evolution in such systems. This project will explore these research questions in two particular contexts: concept learning in vision, particularly in the domain of interstate highway traffic viewed by remote cameras, and concept learning in planning, particularly in the domain of Sokoban problems.