LEARNING IN MODULAR SYSTEMS: A COMPUTATIONAL APPROACH

模块化系统中的学习：一种计算方法

基本信息

批准号：
2890719
负责人：
ROBERT ALAN JACOBS
金额：
$ 9.47万
依托单位：
UNIVERSITY OF ROCHESTER
依托单位国家：
美国
项目类别：
财政年份：
1995
资助国家：
美国
起止时间：
1995-07-01 至 2001-06-30
项目状态：
已结题

来源：
https://reporter.nih.gov/project-details/2890719
关键词：
association learning behavioral /social science research tag cognition computational neuroscience human subject learning social competition

项目摘要

Nearly all of the research investigating the modular nature of the mind/brain in general, and the acquisition of functional specializations in particular, is behavioral or neuroscientific in character. The progress of this research effort has been impeded by a lack of computational studies that attempt to relate functional properties with underlying cognitive and neural mechanisms. Jacobs's research program has developed a family of computational architectures, referred to as mixtures-of- experts architectures, that acquire functional specializations by combining associative learning mechanisms and competitive learning mechanisms. Within the mixtures-of-experts computational framework, a surprising result is that the mechanisms underlying the acquisition of functional specializations and the mechanisms underlying the acquisition of modular integrations are identical. The goal of the proposed research is to develop these mechanisms for the purpose of acquiring modular integrations. Studies one through three investigate bootstrap learning; that is, the ability to use solutions to simpler tasks as "building blocks" for more difficult tasks. One aim of these studies is to investigate the role of context-dependent visual object representations in the acquisition of invariant visual representations. Studies four and five use the mixtures-of-experts architecture to investigate the aggregation of multiple expert opinions. In particular, they examine the issue of whether modular integrations during visual processing should be based on selection or on combination.

几乎所有研究了研究的模块化性质一般思维/大脑，以及功能专业的获取特别是，特征是行为或神经科学的。进度缺乏计算，这项研究工作阻碍了试图将功能特性与潜在的研究认知和神经机制。雅各布斯的研究计划已经制定一个计算体系结构家族，称为混合物专家体系结构，通过结合关联学习机制和竞争性学习机制。在专家的混合物中，计算框架，一个令人惊讶的结果是，采集的机制功能专业和收购的基础机制模块化整合是相同的。拟议研究的目标是为了获得模块化而开发这些机制集成。研究一到三个研究引导学习；也就是说，将解决方案用于简单任务的能力作为“构建 “对于更艰巨的任务。这些研究的目的是研究与上下文相关的视觉对象表示的作用获取不变的视觉表示。研究四和五使用杂货结构的混合物来研究多种专家意见。特别是，他们研究了是否视觉处理过程中的模块化集成应基于选择或组合。