Neural Codes Underlying Visual Segmentation

点击翻译按钮获取中文摘要

Project Summary/Abstract In natural vision, it is rare to encounter an isolated object presented on a blank background. Instead, natural scenes are often complex and contain multiple entities. Image segmentation refers to the process of partitioning visual scenes into distinct objects and surfaces, which includes segmenting a figure from the background (figure- ground segregation) and segmenting multiple objects/surfaces from each other. Segmentation is a fundamental function of vision and is a gateway to perception, recognition and visually guided action. However, the neural underpinning of segmentation remains to be understood. A key question is to understand how the brain represents multiple visual stimuli such that information regarding individual stimuli can be extracted from the activity of populations of neurons. We address this question in the proposed project to elucidate the neural mechanisms underlying segmentation and the principles of coding sensory information in neuronal populations. Visual motion and depth provide potent cues for segmentation. Therefore we focus on understanding how the brain uses motion and depth cues to achieve segmentation. We have made substantial progress in defining how middle-temporal (MT) cortex, an area important for motion and depth processing, represents multiple overlapping visual stimuli. We found that MT neurons show various types of response biases toward one component of multiple stimuli, revealing a set of novel rules by which multiple stimuli interact within neurons’ receptive fields. These physiological findings together with our preliminary data on natural scene statistics led us to hypothesize that the visual system exploits the statistical regularities in natural scenes that differentiate figure from the background and represents multiple visual stimuli efficiently to achieve segmentation. To test this overarching hypothesis, we will integrate the approaches of natural scene statistics, neurophysiology, and theoretical consideration of optimal coding. Specifically, we will characterize natural scene statistics of depth and motion pertinent to image segmentation, elucidate the functional roles of stereoscopic depth in figure-ground segregation, define the rules by which neurons in area MT represent multiple spatially-separated stimuli, which are commonly encountered in natural vision, and determine the signal transformation across multiple brain areas in the dorsal visual pathway to achieve segmentation. Finally, we will use an Information-Maximization approach to determine whether the neural representation of multiple visual stimuli is optimal for segmentation. The proposed study rigorously explores the interaction of multiple stimuli and is expected to provide important insight into how the visual system solves the challenging problem of segmentation in natural vision.