Wafer-Scale Dynamic Neural-Network-System with Optical Waveguide

提出了一种具有光波导的动态神经网络系统。活体神经有几千个连接，即突触。这些连接在生命信息处理系统中起着重要的作用，为人类提供识别和联想。光波导可以像突触一样同时汇总多个输入信号。本研究旨在研究这种具有光波导的光神经网络系统的可能性，并制作该系统的测试芯片和测试模块。对具有光波导的动态神经网络系统采用同步摄动学习算法。我们可以在同时摄动的学习算法中以相同的增量同时更新突触的权值。因此，同时摄动学习算法非常适合于光波导动态神经网络系统。我们提出了新的突触电路和学习电路。我们还提出了一种三维堆叠结构的动态神经网络系统。该系统除采用光波导进行突触连接外，还采用了可重构电路。为了制造该系统的测试芯片和测试模块，开发了基于波导的光互连技术和基于可重构电路的三维集成技术等基础技术。

We proposed a new dynamic neural network system with optical wave guides. Living nerves have several thousand connections, namely synapses. These connections play an important role in living information processing system and provide recognition and association for human. The optical wave guide can simultaneously sum up many input signals like synapses. This study aimed in study the possibility of such optical neural network system with optical wave guides and to fabricate the test chips and test modules for the system. We adopted the learning algorithm using simultaneous perturbation for the dynamic neural network system with optical wave guides. We can simultaneously update the weights of synapses by the same increment in the learning algorithm using simultaneous perturbation. Therefore, the learning algorithm using simultaneous perturbation is very suitable for the dynamic neural network system with optical wave guides. We proposed new synapse circuit and learning circuit for this system. We also proposed a dynamic neural network system with three-dimensionally stacked structure. The re configurable circuits are used in this system in addition to the optical waveguides for synaptic connection. In order in fabricate the test chips and teat modules for this system, basic technologies such as the optical interconnection technology using waveguides and the three-dimensional integration technology using re-configurable circuits were developed.