Reference: Pfleger, K. & Hayes-Roth, B. Learning of Compositional Hierarchies for the modeling of context effects. Knowledge systems Laboratory, January, 1998.
Abstract: Compositional, or part-whole, hierarchies underlie many forms of data, and representations involving these structures lie at the heart of much of the work in Artificial Intelligence and Cognitive Science. However, despite their prevalence, general methods for learning such structures from data are scarce. This paper presents a learning and prediction system that learns compositional hierarchies and uses them to mediate context effects in making predictions. The model is a hybrid system based on an early psychological neural network system, the Interactive Activation model of context effects in letter perception, and an elegant new symbolic hierarchy-generation algorithm called Sequitur. The composite system overcomes an important limitation in each of its parents.
Full paper available as ps.