基于梯度下降算法的光学神经网络性能的研究

Research on the performance of optical neural network based on gradient descent algorithm

光电混合人工智能计算芯片在人工智能应用中通过人工智能算法实现高速和高效的计算,其中光学神经网络(Optical Neural Networks,ONNs)算法在实现大量矩阵运算方面尤为重要.通过使用由马赫曾德尔干涉仪(Mach-Zehnder interferometers,MZI)搭建的快速傅里叶变换(Fast Fourier transform,FFT)型光学神经网络来实现手写数字的高精确度识别.在模型构建方面,利用奇异值分解将神经网络的线性层进行分解,从而实现数据降维,主要特征提取.在对该ONN的训练中,分别采用了带动量的随机梯度下降算法(Stochastic Gradient Descent with momentum,SGD with momentum)和均方根传递(Root Mean Square propagation,RMSprop)算法,分析了在不同训练算法下该ONN对手写数字的识别精度.此外,还深入剖析了两种训练算法背后的数学理论,探究造成两种训练算法实验结果差异的本质原因.最后,通过实验对比,发现RMSprop算法在FFT型光学神经网络上具有较高的识别精确度,达到97.4%;并且采用RMSprop算法的ONN计算速度优于SGD with momentum算法.

Photoelectric hybrid artificial intelligence computing chip realizes high-speed and efficient computing through artificial intelligence algorithms in artificial intelligence applications.Particularly,Optical neural networks algorithm is important in realizing a large number of matrix operations among them.We use a fast Fourier transform type optical neural network built with Mach-Zehnder interferometers to achieve high-precision recognition of handwritten digits.In terms of model construction,the linear layer of the neural network is decomposed by singular value decomposition,so as to realize data dimension reduction and main feature extraction.In the training of the optical neural networks,the stochastic gradient descent with momentum algorithm and the root mean square propagation algorithm were used respectively,and the recognition accuracy of the optical neural networks for handwritten digits was analyzed under the different training algorithms.In addition,we also deeply analyze the mathematical theory behind the two training algorithms,and explore the essential reasons for the difference between the experimental results of the two training algorithms.Finally,through experimental comparison,we found that the root mean square propagation algorithm has a high recognition accuracy on the fast Fourier transform type optical neural network,reaching 97.4%.what′s more,the optical neural networks calculation speed using the root mean square propagation algorithm is better than the stochastic gradient descent with momentum algorithm.