Fundamental studies on quantum neural devices

以下是1995-1996年度科学研究资助项目(C)“07650393：量子量子器件基础研究”的研究成果要点。这个项目是为我们的最终目标“实现量子神经设备”提供线索，并将遵循下一阶段的项目。为器件设计和定量分析奠定理论基础对以光波为信息载体的相干光神经网络中信息载体能量与学习过程的关系进行了理论分析和仿真研究。相干神经网络利用载波频率空间作为神经信息处理的一种新的并行度。然而，人们发现，在学习有意义的频域泛化之前，其初始状态存在某种限制。在反向传播学习理论中，我们在神经连接延迟空间和载波频率空间中发现了周期性误差最小谷，这在仿真中也观察到了。在大多数情况下，在这样的山谷和山脉的褶皱上进行学习过程是不可能的。因此，网络的初始状态必须在一定的参数范围内（特别是在延迟域）才能顺利泛化。然而，也发现相干网络的限制非常宽松，因此，实现未来的量子神经设备不是一个重大问题。用宏观电子电路验证基本动力学用微波和毫米波域的宏观高频电子电路验证用于未来量子神经装置的相干神经网络的基本动力学也在进行中。构造了相敏相干雷达样机，利用该样机获得了雷达目标的二维复值数据。在本初步实验中，原始数据在包含相干网络和量子神经器件关键技术的复值神经网络中进行处理，仅在冯·诺伊曼型计算机上通过软件进行明显的并行处理。在本实验中，我们发现物体上的相位信息对于物体图像的认知或重建是非常有效的。这一结果表明相干神经网络将产生一种利用信息载体（光子、电子等）相干方面的新型神经装置，即量子神经装置。少

The follows are the key points of results of the research project "07650393 : Fundamental studies on quantum neral devices" supported by the Grant-in-Aid for Scientific Research (C), fiscal year 1995-1996. This project is to clues us to the final goal "realization of quantum neural devices" and will be followed by next-stage projects.1. Theoretical establishment of fundamentals for designing the devices and quantitative analyzesTheoretical analysis and simulations have been carried out to investigate the relation between the energy of the information carrier and the learning process in coherent optical neural networks using lightwave as the information carrier. The coherent neural networ can use the carrier-frequency space as a new degree of parallelism of neural information processing. However, it has been found that there is a sort of limitation in its initial state before learning for a meaningful generalization in the frequency domain. In the theory of backpropagation learning, we … More found periodic error-minimal valleys in the neural-connection delay space and the carrier frequency space, which are observed also in simulations. A learning process over the fold of such valley and mountain range cannot be performed in most cases. Therefore the network initial state have to be within a certain range of parameters (especially in the delay domain) for a smooth generalization. However, it has also been found that the limitation is very loose for the coherent networks and, hence, it is not a significant problem to realize a future quantum neural devices.2. Verification of basic dynamics by experiments using macroscopic electronic circuitsThe verification of the basic dynamics of the coherent neural networks for a future quantum neural devices are also in progress by using macroscopic high-frequency electronic circuits in the microwave and millimeterwave domain. A prototype of phase-sensitive coherent radar is constructed by which two-dimensional complex-valued data of rader objects are obtained. The raw data are process in a complex-valued neural network, which contains the key technology of the coherent networks and the quantum neural devices, in parallel only apparently by software on a von Neumann-type computer in the present preliminary experiment. In this experiment, it is found that the phase information on the objects is significantly effective for cognition or reconstruction of the object image. This result suggests that the coherent neural networks will give rise to a novel neural devices utilizing the coherent aspects of the information carrier (photons, electrons, ets.), i. e., the quantum neural devices. Less