Establishing the Limits of Perceptual Inference for Visual Motion

关于世界的统计知识在决定我们的感官知觉方面起着重要作用。当传入的

Learned statistics about the world play an important role in dictating our sensory perception. When incoming sensory inputs carry limited information, such as in low-contrast conditions like dusk, percepts appear to be heavily dictated by implicit assumptions about the probability of different sensory experiences. More specifically, contemporary research suggests that human visual motion perception is well-described by an ideal observer model that gathers environmental information about motion, but also assumes that objects in the environment are most likely either stationary or moving relatively slowly. Theoretical work implementing this model, referred to as a Bayesian observer with a "slow speed prior", has successfully explained many disparate perceptual studies that found curious biases in perceived motion, and has had far-reaching influence on how we think about human spatial vision. As useful as this slow speed prior hypothesis is, it makes several critical, untested assumptions: namely that the visual system represents a motion prior in an accurate, world- based coordinate system. While this is ideal for a Bayesian observer, it is at odds with evidence from the psychophysical literature on visual motion perception. It is also unclear how flexible this prior is in adults in the face of changing environmental conditions or stimuli. Thus the overarching hypotheses of this proposal are that human representations of motion statistics (1) are updated in the light of strong evidence for either general changes in environmental statistics or changes in stimulus-specific statistics, and (2) are best characterized by coordinate system that is intermediate between retinal and world systems. The proposed research will address hypothesis (1) in Specific Aim 1 by testing whether changes in perceived visual motion following exposure to altered motion statistics are well-explained by updates to slow speed prior that generalizes across stimuli and tasks. This proposal will address hypothesis (2) in Specific Aim 2 by estimating priors under conditions that dissociate retinal and world motion statistics. In each aim, the questions will be studied using a combination of visual psychophysics and computational modeling that formalizes the representation of the priors. Together, these aims will address the substantial gap in the current literature that is marked by a vast number of perceptual learning studies and relatively few studies addressing the ways in which the visual system updates its representation of sensory statistics. These studies will contribute to our knowledge on the fundamental computations involved in the transformation of sensory evidence from the periphery into robust percepts. This fellowship proposal includes a detailed training plan at a world-class research institution (University of California, Berkeley) with several specialists in psychophysical research and access to modern visual display technology and computational resources. The proposed research will take advantage of these resources to design a next-generation model of perceptual inference in the visual system that is well-rooted in physical and biological constraints.