基于数据冗余性的近似计算技术  

Approximate computing technique based on data redundancy

作　　者：陈折桂 王晶[1,2] 张伟功[1,3] CHEN Zhe-gui;WANG Jing;ZHANG Wei-gong(Information Engineering College,Capital Normal University,Beijing 100048,China;Beijing Advanced Innovation Center for Imaging Theory and Technology,Capital Normal University,Beijing 100048,China;Beijing Key Laboratory of Electronic System Reliability Technology,Capital Normal University,Beijing 100048,China)

机构地区：[1]首都师范大学信息工程学院,北京100048 [2]首都师范大学北京成像理论与技术高精尖创新中心,北京100048 [3]首都师范大学电子系统可靠性技术北京市重点实验室,北京100048

出　　处：《计算机工程与设计》2021年第3期649-655,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61772350);共有信息系统装备预先研究基金项目(公开)(JZX2017-0988/Y300);北京市科技新星计划基金项目(Z181100006218093);北京未来芯片技术高精尖创新中心科研基金项目(KYJJ2018008);北京市高水平教师队伍建设计划基金项目(CIT&TCD201704082);科技创新服务能力建设-基本科研业务费基金项目(科研类)(19530050173)。

摘　　要：为解决神经网络中爆炸式增长的数据量所加剧的计算机系统存储墙问题,针对经典近似方法在平衡精确度和开销问题上的不足,提出一种利用相邻数据之间相似特征以及神经网络中数据冗余性的优化方法。通过分析数据密集型应用的数据特征,提出针对权值的稀疏矩阵压缩,针对输入数据的重复值压缩和相邻小范围浮动数据近似的方案,设计神经网络拓扑感知的冗余数据近似方法,在合理范围内实现数据近似。实验结果表明,相比于SNNAP方案输出质量平均提高了8.16%,压缩率达到51.38%,为提高数据密集型应用的系统加速比提供了可行性。To address the problem of computer system storage walls exacerbated by the explosive amount of data in neural networks,and aiming at the shortcomings of classical approximation methods in balancing accuracy and cost,an optimization method using similarity between adjacent data and data redundancy in neural networks was proposed.By analyzing the data cha-racteristics of data-intensive applications,a scheme was proposed for sparse matrix compression for weight,aiming at repeated value compression for input data and approximation for adjacent low dynamic range data,an approximation method of redundant data for neural network topology perception was designed to realize data approximation within a reasonable range.Experimental results show that compared with SNNAP scheme,the output quality is improved by 8.16%on average and the compression rate reaches 51.38%.It is feasible to improve the system speedup ratio for data-intensive applications.

关 键 词：神经网络 数据冗余 数据相似 近似 压缩 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]