Adaptive Abstraction and the Computational Basis of Compositional Cognition

Abstract

The world is too rich to represent in full detail, yet we navigate it successfully. We quickly reason about physical events without solving complex equations, imagine vivid pictures without thinking about every pixel, and combine ideas in novel ways without storing every combination in advance. How does the mind use the right level of abstraction in different scenarios? This dissertation examines how the mind constructs adaptive abstractions that are selective to be efficient, structured to support multiple computations, and compositional to enable flexible recombination. I examine these phenomena across domains of progressively greater generality. Chapter 1 focuses on intuitive physics and shows that physical reasoning relies on a coarse body representation that is different from the fine-grained shape representation used for visual recognition. Chapter 2 investigates mental imagery and shows that internally constructed scenes are not fully rendered replicas of perception. Instead, scene construction unfolds hierarchically: people first commit to broad semantic and spatial scaffolds, and only later, if at all, to surface properties. Chapter 3 turns to domain-general reasoning and presents computational models of compositional cognition, explaining how a neural system can build and manipulate complex semantic structures. These models include three core ingredients (working-memory control, variable binding, and pointer-like indirection) and exhibit systematic generalization on compositional benchmarks where alternative architectures fail. Together, these studies reveal how the mind adaptively abstracts information across contexts to support efficient and effective perception, imagination, and reasoning. These adaptive abstractions help explain how human cognition is both resource-rational and endlessly productive.