基于透明介质成像的光学散射神经网络

Optical scattering neural network based on transparent medium imaging

近年来,随着神经网络的发展,多种光电神经网络框架相继提出,在图像和语音处理等方面表现出强大的应用潜力。设计一种新型的基于透明介质成像的光学散射神经网络框架。描述透明介质双向散射传播的物理特性,采用多层堆叠散射介质构建神经网络单层智能单元,并利用非线性光电材料薄膜实现非线性激活,级联单层智能单元构建深度网络物理模型;根据双向散射物理模型构造对应的双向数字复值神经网络,并根据复值求导链式法则推导数字权重更新的后向传播算法,其复值网络权重参数通过调控散射介质的物理参数来实现;在MNIST数据集上验证其智能分类的有效性。

With the development of neural networks in recent years,a variety of photoelectric neural network frameworks have been successively proposed and presented strong application potentials in image and language processing and so on.A new type of optical scattering neural network based on transparent medium imaging is designed.Firstly a physical characteristic of the bidirectional propagation in scattering medium is presented,a single layer intelligent unit of optical neural network is formulated by cascading a stack of medium layer,and the nonlinear photoelectric material film is appended to act as nonlinear activation response.Then,the corresponding digital bidirectional complex-valued neural network is established according to the bidirectional scattering physical model,and the backpropagation algorithm for its digital weight update is derived.The weight parameters of a digital complex-valued neural network can be realized by adjusting the physical parameters of scattering medium.The numerical experiments about the effectiveness of the intelligent classification of proposed optical scattering neural network is verified on the MNIST data set.