神经网络模型压缩方法综述

深度神经网络模型被广泛应用在图像分类、物体检测等机器视觉任务中，并取得了巨大成功。然而，由于存储空间和功耗的限制，神经网络模型在嵌入式设备上的存储与计算仍然是一个巨大的挑战。模型压缩作为一种有效的解决方法，受到了越来越多研究者的关注。首先针对卷积神经网络模型进行了研究，分析了模型中存在的冗余信息；随后对国内外学者在神经网络模型压缩方面的研究成果整理，从参数剪枝，权重共享和权重矩阵分解等方面总结了神经网络压缩的主要方法；最后针对神经网络模型发展现状及目前面临的若干主要问题进行了讨论，指出了下一步的研究方向。

Survey of neural network model compression methods

Deep neural network has been widely utilized in the image classification, recognition, object detection and other machine vision fields with an unprecedented success. However, limited by the memory space and power, it is a challenging task to deploy the deep neural network model in embedded system. As an effective solution, model compression has attracted tremendously attention for researchers. At first, the basic deep neural network model is introduced and the redundant computation in the model is analyzed. Then, the existing methods for model compression are presented from the aspects of parameter pruning, parameter sharing and weight matrix decomposition. Finally, the potential challenges in deep neural network and direction of future research are discussed.