UC San Diego

Humans must continually generalize from past experience to novel scenarios. For instance, when we see a new coffee mug, although it does not look exactly the same as the ones we have seen before, we can still categorize it as a coffee mug, recognize the handle, and infer that it will probably break when dropped to the ground. What mental representations enable people to make predictions about objects in new environments? In this dissertation, I explore different possible computational constraints on what people infer about object properties and their interactions that enable them to generalize. In Chapter 1, I begin by investigating how people learn to recognize novel objects. I develop a paradigm where I systematically manipulate the underlying rules that generated the stimuli. I find that people can learn both specific exemplars and abstract object patterns depending on the statistics of the environment and generalize to novel instances. I show that people’s behavior can be explained by a program learning model that simulates the generating process of the stimuli. In Chapter 2, I analyze how people make generalizable judgements about the physical interactions between objects. People are asked to predict the behavior of an object in an environment it has never been in before. Results suggest that people simultaneously infer multiple physical variables (i.e., object mass, force in the environment) based on past observations of how objects behave and compose knowledge about these variables to make predictions. In light of these findings, Chapter 3 examines how people generalize their understanding of physical interactions across multiple scenarios. Across two studies with seven different physical scenarios (e.g. collision, containment, support, etc.), I find that people are accurate and consistent with each other in their predictions about whether and where two objects would contact. I then show that a computational model that runs simulations of noisy physics generalizes in human-like ways across all scenarios. Altogether, this dissertation suggests that representing entities and how these entities interact in an abstract internal model plays a key role in, and potentially provides a coherent account of, human generalization.