ATHENA for Memory-Based Inference

Instance models are widely used to model memory and inference and have been a staple of computational cognitive modeling since their inception. However, there have recently been important developments in our understanding of energy-based memory models and their ability to make reasonable inferences. In this thesis, I argue that one class of energy-based model, known as modern Hopfield Networks, is closely related to extant instance models of cognition, and that understanding the relationship betweenthese approaches suggests new ways that instance models might be understood and used. To do this, I introduce ATHENA, a novel instance model that makes inferences via retrieval in a modern Hopfield network and investigate its properties and capabilities. I then turn to cognitive development, using ATHENA to help understand the different features that children and adults use when reasoning and their differing ability to maintain distinguishable memory traces. Finally, I discuss the implications of thesefindings for early cognition and cognitive modeling.