An Integrated Model of Memory

Abstract

Memory is a foundational building block of human cognition. Computational models of memoryhave traditionally segmented memory by function, dividing it into three separate but interacting systems: semantic memory, episodic memory, and working memory. This dissertation instead takes a foundational approach to studying memory, outlining five computational principles underlying prior computational models: statistical learning, distributed representation, similarity-based structure, graded abstraction, and context dependence. Through a series of twelve studies, I construct a set of computational models that instantiate these five principles and demonstrate how these models provide a coherent explanation of human behavior across a wide range of paradigms including property identification, inductive inference, similarity judgment, reward learning, sequence learning, and category learning. I show that the principles-based approach to understanding memory provides insights into human behaviors that lie at the boundaries of traditional, functionally-divided memory systems, explaining, for example, how representations traditionally associated with working memory reflect semantic structure and how representations traditionally associated with episodic memory can be dynamically controlled to optimize behavior. Overall, the results demonstrate the value of understanding cognitive systems through computational principles rather than functional units, and they further support the connectionist principle that complex behavior can emerge from interactions among a simple set of rules.