Using Kerr-Type Nonlinear Optical Materials for Artificial Neural Network Implimentations

这个项目的目的是研究克尔型非线性光学材料在全光实现具有光误差反向传播训练的前馈人工神经网络中的有用性。这些类型的材料具有皮秒级的响应时间，并且允许使用由自由空间分隔的薄材料层来实现加权连接和非线性神经元处理。这使得这种网络优于其他需要复杂硬件设备来实现加权连接和单独硬件进行神经元处理的光学实现。此外，kerr型材料的分层网络可以同时处理前向计算信号和后向误差传播。非线性层作为处理层，线性层（费用空间）作为连接层。每个非线性层都可以看作是神经元的平面连续体，通过连续的光路与前后各层相连。可用神经元和连接的数量仅受系统光学的可用分辨率的限制。输入和权值是二维空间中辐照度的分布，权值改变非线性介质的折射率分布。通过施加在非线性层上的传播波与光信号耦合的自作用效应，给出了神经计算所需的非线性强迫和处理函数。对该网络进行了二维数值模拟，并训练其执行逻辑函数，并构造了该网络的简单光学演示。提出了对网络进行进一步的仿真和训练，以执行更复杂的任务，并在二维中实现这些网络，然后将仿真和实现扩展到三维。这将涉及理论和实验研究，并与计算机模拟并行，以确定最佳的网络几何形状（层数和类型，带场和信号场的大小和所需分辨率）以及控制重量光场应用的更好技术。＊＊＊

9312345 Skinner The objective of this project is to investigate the usefulness of Kerr- Type nonlinear optical materials for all-optical implementations of feed forward artificial neural networks with optical error back- propagation training. These types of materials have pico-second response times and allow both weighted connections and nonlinear neuron processing to be implemented using thin materials layers separated by free space. This gives such networks and advantage over other optical implementations which require complicated hardware devices to implement the weighted connections and separate hardware for neuron processing. In addition, a layered network of Kerr-Type material can process both forward calculation signals and backward error propagation simultaneously. The nonlinear layers serve as processing layers and the linear layers (fee space) serve as connection layers. Each nonlinear layer can be thought of as a plane continuum of neurons connected with the layers before and after it by a continuum of optical paths. The number of available neurons and connections are limited only by the available resolution of the system optics. The inputs and weights are a distribution of irradiance in 2D space and the weights vary the refractive index profile of the nonlinear medium. The nonlinear forcing and processing functions necessary for neural computing are given by the self-action effects of the propagating wave coupled with the optical with signals applied to the nonlinear layers. Such a network has been numerically simulated in two dimensions and trained to perform logic functions and a simple optical demonstration of this network is constructed. Further simulation and training of the network to preform more complicated tasks as well as implementation of these networks in two dimensions is proposed, followed by an extension of the simulation and implementation to three dimensions. This well involve both theoretical and experimental investigations in parallel with computer simulation to determine the optimal network geometry (number and type of layers, size and required resolution of with and signal fields) as well as better techniques for controlling the application of the weight light fields. ***