深度卷积神经网络的数据表示方法分析与实践

发布时间：2018-01-11 02:28

本文关键词：深度卷积神经网络的数据表示方法分析与实践　出处：《计算机研究与发展》2017年06期 　论文类型：期刊论文

【摘要】：深度卷积神经网络在多个领域展现了不凡的性能,并被广泛应用.随着网络深度的增加和网络结构不断复杂化,计算资源和存储资源的需求也在不断攀升.专用硬件可以很好地解决对计算和存储的双重需求,在低功耗同时满足较高的计算性能,从而应用在一些无法使用通用CPU和GPU的场景中.在专用硬件设计过程中仍存在着很多亟待解决的问题,例如选择何种数据表示方法、如何平衡数据表示精度与硬件实现代价等.为解决上述问题,针对定点数和浮点数建立误差分析模型,从理论角度分析如何选择表示精度及选择结果对网络准确率的影响,并通过实验探究不同数据表示方法对硬件实现代价的影响.通过理论分析和实验验证可知,在一般情况下,满足同等精度要求时浮点表示方法在硬件实现开销上占有一定优势.除此之外,还根据浮点表示特征对神经网络中卷积操作进行了硬件实现,与定点数相比在功耗和面积上分别降低92.9%和77.2%.
[Abstract]:The depth of the convolutional neural network show the extraordinary performance in many fields, and has been widely used. With the increase of the depth of the network and the network structure is more complex, computing resources and storage resources demand is also rising. Dedicated hardware can solve the dual demand of computing and storage, in low power consumption while satisfying calculation high performance, which is used in some cannot use the general CPU and GPU in the scene. In the special hardware design process, there are still many problems to be solved, such as the choice of data representation, accuracy and hardware cost. How to balance the data representation to solve the above problems, the fixed-point and floating-point error analysis of model selection from the perspective of theoretical analysis how to express the accuracy and influence the accuracy of network selection results, and different data representation method of hardware experiment research Effect of implementation cost. Through theoretical analysis and experimental verification shows that, in general, to meet the same requirements of precision floating-point representation method has some advantages in hardware cost. In addition, also features of the hardware implementation of convolutional neural network in operation according to the floating point, compared with the fixed point number in power and area were reduced by 92.9% and 77.2%.

【作者单位】：清华大学计算机科学与技术系;清华信息科学与技术国家实验室(筹);
【基金】：国家自然科学基金项目(61373025) 国家重点研发计划项目(2016YFB1000303)~~
【分类号】：TP183
【正文快照】： This work was supported by the National Natural Science Foundation of China(61373025)and the National Key Research andDevelopment Program of China(2016YFB1000303).(wpq14@mails.tsinghua.edu.cn)卷积神经网络(convolution neural network,CNN)因为其高准确率,广

【相似文献】