基于紧致编码的高阶神经网络光计算

High-order Neural Optical Computing Based on a Compact-encoding Method

提出一种能压缩三阶神经网络的互连张量规模近一半的紧致编码方法。设计两种光电混和系统-非相干多重成象系统和非相干相关系统,并用它实现了三阶互连的二维神经网络,构成图象平方关联存贮器。在光学实现中,借助于一种多重矩阵排列方式,作者用一块二维掩模板表示六维的互连张量,用二维的发光二极管(LED)阵列表示四维的输入张量,实验结果与数值模拟的结果吻合。

A compact-encoding method, which can reduce nearly the half size of the interconnection tensor in the third-order neural network, is proposed. Two types of incoherent optical systems, based on multiple-imaging and correlation, are used to realize the 2-D neural network with a third-order interconnection for image quadratic associative memory. In optucal implementations, a 6 D memory interconnection tensor is realized with a 2-D mask of triple matrix arrangement and a 4-D input tensor is realized with a 2-D LD array of double matrix arrangement. The experimental results are also given.