文献类型：会议论文

英文题名：Evaluating Data Resilience in CNNs from an Approximate Memory Perspective

作者：Chen, Yuanchang[1];Zhu, Yizhe[4];Qiao, Fei[1];Han, Jie[2];Liu, Yuansheng[3];Yang, Huazhong[1]

第一作者：Chen, Yuanchang

通讯作者：Qiao, F[1]

机构：[1]Tsinghua Univ, Dept Elect Engn, Beijing, Peoples R China;[2]Univ Alberta, Dept Elect & Comp Engn, Edmonton, AB, Canada;[3]Beijing Union Univ, Coll Robot, Beijing, Peoples R China;[4]Beijing Univ Posts & Telecommun, Beijing, Peoples R China

第一机构：Tsinghua Univ, Dept Elect Engn, Beijing, Peoples R China

通讯机构：[1]corresponding author), Tsinghua Univ, Dept Elect Engn, Beijing, Peoples R China.

会议论文集：Great Lakes Symposium on VLSI (GLSVLSI)

会议日期：MAY 10-12, 2017

会议地点：Banff, CANADA

语种：英文

外文关键词：Data Resilience Evaluation; Convolutional Neural Network; Approximate Memory

摘要：Due to the large volumes of data that need to be processed, efficient memory access and data transmission are crucial for high-performance implementations of convolutional neural networks (CNNs). Approximate memory is a promising technique to achieve efficient memory access and data transmission in CNN hardware implementations. To assess the feasibility of applying approximate memory techniques, we propose a framework for the data resilience evaluation (DRE) of CNNs and verify its effectiveness on a suite of prevalent CNNs. Simulation results show that a high degree of data resilience exists in these networks. By scaling the bit-width of the first five dominant data subsets, the data volume can be reduced by 80.38% on average with a 2.69% loss in relative prediction accuracy. For approximate memory with random errors, all the synaptic weights can be stored in the approximate part when the error rate is less than 10(-4), while 3 MSBs must be protected if the error rate is fixed at 10(-3). These results indicate a great potential for exploiting approximate memory techniques in CNN hardware design.