Coherent neural networks : Construction of prototypal subsystems

尽管“脑型系统”的定义还不明确，但脑型信息处理系统的研究仍十分活跃。预期的脑型系统将至少是对所谓的冯-诺伊曼型系统的补充。换句话说，未来脑型系统最令人期待的特征应该是基于高度并行架构的分布式和灵活的（自组织和/或学习）信息处理。也就是说，我们必须实现一个在图灵机原理之外有自己基础的大规模并行和灵活的硬件。本课题旨在利用光波、电磁波等波场实现神经网络，构建未来脑型系统的原型子系统；也就是相干神经网络。相干神经网络是申请人课题组在1992年提出的复值神经网络的扩展。在本项目中，我们的目标是建立光波（或电磁波/电子波）相干神经网络的动态行为的理论基础，其中可以操纵信息载体的振幅和相位，并实现原型子系统硬件。在两年的研究中，我们(1)阐明了行为不稳定性的机制，(2)提出并证明了一种新的学习方法来抑制不稳定性。我们还(3)构建了一个具有复幅信息处理阶段的雷达系统，通过该系统可以阐明电磁波物理特性与复值神经网络之间的动力学关系。这些卓有成效的成果被认为几乎实现了这个项目的目标。我们期望在下一阶段进行进一步的理论和实验研究，以构建更现实的相干神经网络。

Brain-type information processing systems are actively studied widely, although there is no clear definition of the "brain-type systems'yet. The expected brain-type systems will be, at least, complementary to the so-called von-Neumann-type systems. In other words, the most expected feature of the future brain-type systems should be a distributed and flexible (self-organizing and/or learning) information processing based on a highly parallel architecture. That is to say, we have to realize a massively parallel and flexible hardware which has its own basis outside the Turing machine principle.The research project aims at a construction of the prototypal subsystems of the future brain-type systems by way of realizing neural networks using wave fields such as lightwave and electromagnetic wave ; i.e., the coherent neural networks. The coherent neural networks are an expansion of the complex-valued neural networks proposed by the applicant's research group in 1992. In the present project, we have been aiming to establish the theoretical basis of the dynamic behavior of the lightwave (or electromagnetic-wave/electron-wave) coherent neural networks in which the amplitude and the phase of the information carrier are manipulated, and to realize the prototypal subsystem hardware.In the two-year research, we have (1) elucidated the mechanism of the behavioral instability, (2) proposed and demonstrated a novel learning method to suppress the instability. We have also (3) constructed a radar system that has a complex-amplitude information processing stage by which the dynamics relation between the physics of electromagnetic wave and the complex-valued neural networks can be elucidated. The aims of this project are considered almost accomplished by these fruitful results. We expect that further theoretical and experimental research on the next stage will be carried out for constructing a more realistic coherent neural networks.