权益分类	功能权益	普通用户	{{item.name}}会员
{{category.name}}	{{benefitItem.name}}

Studies on the formation of intellectual concepts with neural network models.

用神经网络模型研究智力概念的形成。

基本信息

批准号：
03831011
负责人：
KISHI Koichiro
金额：
$ 1.34万
依托单位：
Jichi Medical School
依托单位国家：
日本
项目类别：
Grant-in-Aid for General Scientific Research (C)
财政年份：
1991
资助国家：
日本
起止时间：
1991 至 1993
项目状态：
已结题

项目摘要

In the cerebral cortex, it is assumed that information is represented in the activity pattern of an assembly of neurons. In this study, in order to investigate the formation of intellectual concepts, such a distributed representation was incorporated in the output layr of a neural network with the error-back-propagation algorithm. The network recognized the learned data with complete accuracy. Several units of the hidden layr responded to a series of related data. The network showed high tolerance for the inactivation of the hidden layr units. In addition, it recognized the newly constructed data with high accuracy. These results showed some advantages of the formation of intellectual concepts with a distributed representation. Furthermore, according to the current knowledge of cognitive neuroscience, a recurrent neural network model for controlling multiple intelligence was proposed. The model had multiple neural networks for multiple intelligence. Information from multiple inputs was integrated in the higher association layr through secondary sensory layrs. The output of the higher association layr was transmitted to the filter layr, and the filter provided each secondary sensory layr with each momentary bias, which directed the successive expression of intelligence. The recurrent neural network learned the learning data set with the recurrent back-propagation algorithm. After learning, it was found that this network processed input data according to the previous command. The information representation in the higher association layr may be regarded as the intellectual concepts and will be analyzed in further studies.

在大脑皮层，信息被认为是以一组神经元的活动模式来表示的。在这项研究中，为了研究智能概念的形成，采用误差反向传播算法将这种分布式表示结合到神经网络的输出层中。网络完全准确地识别了学习到的数据。隐藏层的几个单位对一系列相关数据做出了回应。该网络对隐藏的LayR单元的失活表现出很高的耐受性。此外，它对新构建的数据进行了高精度的识别。这些结果显示了用分布式表征形成智力概念的一些优势。在此基础上，根据认知神经科学的现有知识，提出了一种控制多元智能的递归神经网络模型。该模型具有多个神经网络，以实现多种智能。来自多个输入的信息通过次级感觉层整合到较高的关联层。较高关联层的输出被传输到过滤器层，过滤器向每个二级感觉层提供每个瞬间的偏向，这指导了智力的连续表达。递归神经网络用递归反向传播算法学习学习数据集。经过学习发现，这个网络是按照之前的命令处理输入数据的。较高关联层的信息表征可以看作是智力概念，将在进一步的研究中加以分析。