INFORMATION-FUSION LEARNING METHODS OF NEURAL NETWORKS AND THEIR APPLICATIONS

人脑是现存最高级的信息处理机器。它们融合了五个传感器获得的各种信息，并识别外部世界。本研究项目的目的是建立生物系统信息融合的数学模型，开发利用人工神经网络实现信息融合的方法。提出了一种能够融合各种信息的人工神经网络学习方法。我们还研究了神经网络信息融合学习的应用。研究结果表明：在一般框架下处理信息融合问题是非常困难的。首先，为了简单起见，我们考虑了不同类型信息之间的内在联系作为教学数据的情况。考虑到不同类型信息之间的内在联系，提出了一种不仅实现…的输入输出的学习方法更多的指标体系，也是从体系中获得的各种数据的内在联系。我们开发了多层网络和递归神经网络的信息融合学习算法。对于可以直接处理复数(信号)的复值神经网络也进行了类似的研究。其次，针对不同类型信息之间没有内在联系的情况，对学习方法进行了扩展。我们建立了一个包含学习回路中目标系统模型的神经网络学习框架，并进一步研究了这些信息融合学习方法在几个工程问题中的应用，如计算机视觉中的遮挡形状和运动场估计、流动速度场估计、联想记忆的实现、神经振荡器的合成等。在一些应用中，信息融合学习方法成功地改善了它们的性能。较少

Brains of human beings are the most superior information processing machine in existence. They fuse various information obtained with five sensors and recognize the external world. The objective of this research project is to develop mathematical model of information fusion performed by biological systems and develop methods for realize information fusion by artificial neural networks. We propose learning methods of artificial neural networks, which are capable of fusing various informations. We also investigate applications of the information-fusion learning of neural networks. The results are summarized as follows.It is very difficult to deal with the problem of information fusion in general framework. Firstly, for simplicity, we considered the case that internal relations among different kinds of informations are available as teaching data. Taking account of internal relations among different kinds of informations we proposed learning methods which realize not only input-output of a … More target system but also the internal relations of various kinds of date obtained from the system. We developed the information-fusion learning algorithms for both multilayer networks and recurrent neural networks. Similar investigations were made for complex valued neural networks, which can directly deal with complex numbers (signals).Secondly, in order to deal with the case that internal relations among different kinds of informations are not available, we extend the learning methods. We developed a framework of neural network learning which includes the model of a target system in the learning loop.Furthermore we investigated applications of the information-fusion learning methods this developed to several engineering problems such as problem of shape from shading and estimation of motion fields in computer vision, estimation of flow velocity fields, realizations of associative memories, synthesis of neural oscillators and so on. It has been shown that in several applications the information-fusion learning methods successfully improve their performance. Less