Towards a computationally precise characterization of the human ventral visual pathway

人类腹侧视觉通路的计算精确表征

• 批准号: 10191834
• 负责人: N Apurva Ratan Murty
• 金额: $ 11.02万
• 依托单位: MASSACHUSETTS INSTITUTE OF TECHNOLOGY
• 依托单位国家: 美国
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-09-01 至 2023-08-31
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/10191834
• 关键词: Address; Algorithms; Animals; Area; Behavior; Brain; Brain Injuries; Brain region; Categories; Cognition; Computer Models; Computing Methodologies; Data; Data Analyses; Databases; Development; Disease; Encapsulated; Evolution; Eye; Face; Functional Magnetic Resonance Imaging; Goals; Human; Image; Impairment; Individual; Lead; Location; Mentors; Methods; Modeling; Monkeys; Network-based; Neurosciences; Pattern; Perception; Performance; Phase; Population; Positioning Attribute; Reading; Recovery; Research; Resolution; Rest; Series; Social Functioning; Stimulus; Supervision; System; Testing; Text; Vision; Visual; Visual Cortex; Visual Pathways; Visual Perception; Work; analytical method; base; cognitive ability; cohort; data quality; deep neural network; design; experimental study; extrastriate; fusiform face area; insight; interest; neuroimaging; neuroprosthesis; neurosurgery; object recognition; public health relevance; relating to nervous system; response; screening; theories; ultra high resolution; vision development; visual information; visual processing; visual stimulus

Project Summary/Abstract: Humans are extraordinarily visual animals, allocating a third of their cortex just to seeing what is in front of them. Visual recognition is supported by a series of hierarchically organized brain regions known collectively as the ventral visual cortex (VVC). Despite extensive research, we still lack a computationally precise understanding of how visual information is represented and transformed over stages of the human VVC. A key barrier has been the limitations of methods like functional MRI (fMRI) which make it difficult to test a large number of experimental stimuli. The research in this proposal will overcome this barrier by collecting fMRI responses to hundreds of stimuli, and analyzing these data using deep neural network based computational models and human interpretable algorithms such as image-synthesis and saliency mapping. In Aim 1 (K99 phase), I will focus on the category-selective regions of the VVC, that respond preferentially to images of faces (fusiform face area), scenes (parahippocampal place area), and bodies (extrastriate body area). I will develop and use new computational methods together with closed-loop experiments to address open questions such as: Is the hypothesized selectivity for these regions even correct? What is represented in the intermediate stages of processing? Are there functionally distinct regions within the category-selective regions? In Aim 2 (R00 phase), I will venture into the ~65% of VVC that lies outside the category-selective regions. I will develop and apply new data-driven clustering to divide these regions into their native components, and characterize them individually. Together, this endeavor will reveal the computational and neural basis of visual recognition in humans with an unprecedented precision. My background in experimental and analytical methods in monkey and human vision puts me in a unique position to accomplish this proposal which requires a seamless integration between neuroimaging experiments and state-of-the-art computational modeling. The proposed work will be initiated in the lab of Prof. Nancy Kanwisher (mentor). During the K99 phase, I will continue to be mentored by Prof. Kanwisher, and will also advance my expertise with computational modeling under the supervision of Dr. Jim DiCarlo (co-mentor), and ultra-high-resolution 7T neuroimaging with Dr. Jon Polimeni (collaborator). This proposed plan will significantly augment my theoretical understanding and experimental abilities, and put me on a path to independence.

项目摘要/摘要：人类是非常视觉的动物，他们大脑皮层的三分之一只用来 看看他们面前的是什么。视觉识别是由一系列等级组织的大脑 这些区域统称为腹侧视觉皮层（VVC）。尽管进行了广泛的研究，我们仍然缺乏一个 在计算上精确理解视觉信息是如何在不同的阶段被表示和转换的。 人体VVC一个关键的障碍是功能性磁共振成像（fMRI）等方法的局限性， 难以测试大量的实验刺激。本提案中的研究将克服这一障碍， 收集对数百种刺激的fMRI反应，并使用基于深度神经网络的 计算模型和人类可解释的算法，如图像合成和显着映射。在 目标1（K99期），我将集中在VVC的类别选择性区域，优先响应于 面部图像（梭形面部区域）、场景图像（海马旁部位区域）和身体图像（纹外体区域）。 我将开发和使用新的计算方法与闭环实验一起解决开放性问题。 问题如：假设的这些区域的选择性是否正确？什么是代表在 加工的中间阶段？在类别选择区域中是否存在功能不同的区域？ 在目标2（R 00阶段）中，我将冒险进入位于类别选择区域之外的约65%的VVC。我会 开发和应用新的数据驱动聚类，将这些区域划分为本地组成部分， 把它们个别化。总之，这一奋进将揭示视觉的计算和神经基础， 以前所未有的精确度识别人类。我在实验和分析方法方面的背景 在猴子和人类的视觉中，我处于一个独特的位置来完成这个需要无缝连接的提议。 神经成像实验和最先进的计算建模之间的整合。拟议工作 将在Nancy Kanwisher教授（导师）的实验室启动。在K99阶段，我将继续接受指导 Kanwisher教授，也将提高我的专业知识与计算建模的监督下， 博士Jim DiCarlo（共同导师）和Jon Polimeni博士（合作者）的超高分辨率7 T神经成像。这 提出的计划将大大提高我的理论理解和实验能力，并把我放在 一条通往独立的道路