Hierarchical Bayesian Causal Inference and Natural Statistics Explain Heaviness Perception
GOAL: The goal of this dissertation is to explore the factors contributing to humans' expectations about our interactions with objects, and perceptual experiences of these interactions, from a sensory integration and computational perspective. More specifically, this project focuses on the regular patterns that exist in the environment for properties of objects typically manipulated by human observers, and how our sensitivity to and representation of these properties contributes to the subjective, perceptual experience of an object's heaviness. These ideas are developed through the following aims:
AIM 1: Sensitivity to complex environmental regularities. This aim first investigates regularities in the joint distribution of the properties of size and weight (i.e., the distribution of the hidden property of density) for liftable objects in the environment. Previous studies have identified univariate environmental regularities in the visual domain, and have demonstrated that sensitivity to such statistics can strongly influence perceptual experiences. This study series identifies a previously-unreported complex environmental statistic linking size with weight, and suggests that human observers represent and can use quantitative knowledge of this statistic to predict the weight of novel, nondescript objects. These findings imply that humans perform inference on the hidden variable of density when judging a novel object's potential weight given its size.
AIM 2: How inference about hidden properties contributes to visuohaptic perception. The next series of studies examines the consequences of AIM 1's conclusions for the visuohaptic percept of an object's heaviness. It is well known that an object's perceived heaviness is inversely proportional to its volume: With weight held constant, a smaller object feels heavier than a larger object. Until now, this Size-Weight Illusion (SWI) has defied modern computational theories of perception, which rest on the notion of a Bayesian ideal observer. Through combining expectations with sensory evidence, Bayesian inference can describe much of perceptual experience across a wide variety of domains - but not the SWI. Through multiple behavioral studies and development of a computational model, this study series demonstrates that assuming human observers perform inference about both observable (size, weight) and hidden (density) variables allows this perceptual illusion to be explained by Bayesian inference.