Compositional hierarchies, layered structures of part-of relationships, underlie many forms of data, and representations involving these structures lie at the heart of much of AI. Despite this importance, methods for learning CHs from data are scarce. We present an unsupervised technique for learning CHs by an on-line, bottom-up chunking process that repeatedly identifies frequently occurring substructures and forms new nodes for them in the hierarchy. The hierarchy is encoded in a predictive model, an extension of Boltzmann Machines, that can at any point make predictions about new data. This work can be seen as a continuation of work that makes predictions using hand-crafted compositional hierarchies, such as early blackboard systems and McClelland and Rumelhart's Interactive Activation Model of Context Effects in Letter Perception.
 

Return to the seminar schedule