Computational Study on the Structure and Learning in the Integrated Neural Networks for Different Sensors

不同传感器的集成神经网络结构和学习的计算研究

基本信息

批准号：
09680365
负责人：
ISHII Naohiro
金额：
$ 2.3万
依托单位：
Nagoya Institute of Technology
依托单位国家：
日本
项目类别：
Grant-in-Aid for Scientific Research (C)
财政年份：
1997
资助国家：
日本
起止时间：
1997 至 1998
项目状态：
已结题

来源：
https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-09680365/
关键词：
Neural network Asymmetric network Structure of network Function of network Nonlinear analysis Integrated network 異種センサーネットワークの構造追加学習

项目摘要

Biological neural networks play important roles in the sensory infor- mation processing, which is different from artificial neural networks. The retinal neural networks in the visual system, are classified into symmetrical and asymmetrical structures. The asymmetrical networks, which are shown in catfish and cat et al., consist of the linear pathway and the nonlinear pathway. This asymmetrical networks have characteristic properties in the sensory perceptions.The asymmetric structures reflect the function of the neural networks. Thus, this study, first analyzes what is the functions of the asymmetric neural networks in the sensory perception. Second, the integrated neural netwoks are discussed to respond and learn different sensory informations from different perceptual networks. The integrated network consist of several sub-netwoks to recognize objects correctly by using several sensors."Forward Network" receives inputs from corresponding sensor. "Integrating Unit" integrates outputs of all. Forward Networks."Backward Network" receives inputs from Integrating Unit and recostructs the sensory information as its outputs. In the recognition phase, the integ- rated system gets a correct output by modifying the inputs of Forward Networks, the output of Integrating Unit and the confidence of the sensory information repeatedly to maximize a likelihood function , which can be derived by a Bayesian method. It was clarified the system has the improved ability in the integrated networks.

生物神经网络在感觉信息处理中发挥着与人工神经网络不同的重要作用。视觉系统中的视网膜神经网络分为对称结构和不对称结构。鲶鱼和猫等人中显示的不对称网络由线性路径和非线性路径组成。这种不对称网络在感官知觉方面具有特征性。不对称结构反映了神经网络的功能。因此，本研究首先分析非对称神经网络在感官知觉中的作用。其次，讨论了集成神经网络来响应和学习来自不同感知网络的不同感觉信息。集成网络由多个子网络组成，通过使用多个传感器来正确识别物体。“前向网络”接收来自相应传感器的输入。 “整合单元”整合所有的输出。前向网络。“后向网络”接收来自整合单元的输入，并重建感觉信息作为其输出。在识别阶段，集成系统通过反复修改前向网络的输入、集成单元的输出和感觉信息的置信度来最大化似然函数，从而得到正确的输出，该似然函数可以通过贝叶斯方法导出。明确了系统在综合网络方面的能力得到了提高。