近年来人工智能受到各国重视，被多国??升至国家战略地位，我国连续三年 在政府工作报告中??及人工智能。随着计算力的??高，人工智能快速发展，在计 算机视觉和自然语言处理领域表现较好。但由于神经网络模型参数众多，因此需 要通过大量的计算资源和存储资源，这成为部署在移动端的一个瓶颈。而随着人 工智能的发展，存储资源和计算资源较小的智能终端产品必将日益普及。因此对 深度学习模型进行压缩，减少模型的部署所需的存储资源显得尤为必要。本文针 对该问题，研究总结了深度学习领域的基础知识和模型压缩领域的工作，并??出 了一种深度神经网络压缩的方法。 为达到模型压缩的目的，本文从神经网络参数矩阵稀疏化方向入手，??出了 一种将矩阵参数分解，深度学习模型中的参数大多包含在全连接层和卷积层。对 于全连接层中的神经元，将与每个神经元相连的权重向量分解为模长与单位特征 向量(二范数为 1)的乘积;对于卷积层中的卷积核，将卷积核分解为矩阵模长 和二维特征矩阵(F 范数为 1)。本文将模长定义为神经元与卷积核的重要程度 的度量指标，通过对模长进行 L1 惩罚达到裁剪压缩神经网络的目的，为保证模 型的可识别，在优化过程中对模长和单位特征向量进行调整。 本文在 Windows 环境下基于 Tensorflow 框架，使用 mnist 数据集和 cifar-10 数据集分别对 LeNet-500-300 模型，LeNet-5 模型和 VGG-16 模型进行实验。结 果表明使用本文??出的方法对模型裁剪，将 LeNet-500-300 模型参数压缩了 85.4% 后，测试集上的泛化误差仅??升了 0.09%;仅对 LeNet-5 模型的全连接层进行压 缩，在网络参数压缩了 82.2%后，测试集上的泛化误差并未??升;将 VGG-16 模 型参数压缩了 90.51%后，测试集上的泛化误差略有下降。以上试验表明本文?? 出的模型在维持网络效果的情况下，能够有效的压缩神经网络模型。

In recent years, artificial intelligence has been valued by many countries and promoted to national strategic status. China has mentioned artificial intelligence in the government work report for three consecutive years. With the improvement of computing power, artificial intelligence develops rapidly and performs well in the fields of computer vision and natural language process. However, due to the numerous parameters of the neural network model, a large number of computing resources and storage resources are needed, which becomes a bottleneck for deployment on the mobile terminal. With the development of artificial intelligence, intelligent terminal products with small storage resources and computing resources will be increasingly popular. Therefore, it is necessary to compress the deep learning model and reduce the storage resources required for model deployment. Aiming at this problem, this paper studies and summarizes the basic knowledge in the field of deep learning and the work in the field of model compression, and proposes a method of deep neural network compression. In order to achieve the purpose of model compression, this paper proposes a method to decompose matrix parameters from the direction of sparse neural network parameter matrix, and the parameters in the deep learning model are mostly contained in the full connection layer and convolution layer. For neurons in the full connection layer, the weight vector connected to each neuron is decomposed into the product of the module length and the unit eigenvector (L2 norm is 1). For the convolution kernel in the convolution layer, the convolution kernel is decomposed into matrix modulus and two-dimensional eigenmatrix (F norm is 1). In this paper, module length is defined as a measure of the importance of the neuron or the convolution kernel. The purpose of II 基于矩阵稀疏化的深度神经网络压缩研究 clipping and compressing the neural network is achieved by imposing L1 penalty on the module length. In order to ensure the recognition of the model, the module length and unit eigenvector will be adjusted in the optimization process. In this paper, based on the Tensorflow framework in the Windows environment, the mnist data set and cifar-10 data set were used to conduct experiments on the LeNet -500-300 model, LeNet-5 model and vgg-16 model, respectively. The results show that after the model parameters are compressed by 85.4% using the method proposed in this paper, the test error is only improved by 0.09%. Only the full connection layer of LeNet -5 model was compressed. After the network parameters were compressed by 82.2%, the test error was not improved. After the vgg-16 model parameters were compressed by 90.51%, the test error decreased. The above experiments show that the proposed model can effectively compress the neural network model while maintaining the network effect.