Estimating Redundancy-Reliability of CNNs Based on Strip-Median Attributes
Estimating Redundancy-Reliability of CNNs Based on Strip-Median Attributes
复制标题
DOI：
10.1109/tvlsi.2023.3297125
复制
发表时间：
2023-10
影响因子：
2.8
通讯作者：
Jie Xiao;Yujian Yang;Haixia Long;Rongzhen Qin;Jungang Lou
中科院分区：
文献类型：
--
作者：
Jie Xiao;Yujian Yang;Haixia Long;Rongzhen Qin;Jungang Lou
深度神经网络的冗余可靠性计算对于其轻量化操作具有重要意义。提出了一种基于条带中值属性的卷积神经网络冗余可靠性计算方法。首先，基于共因原理，构造了一种面向输入样本的自动边界检测可靠性计算方法，以度量正常分类边界变化引起的网络可靠性差异。然后，使用自适应聚类算法并结合条带中值属性来识别网络中每个卷积层（CONV）的冗余条带。然后，基于中心极限定理和Pauta准则，构造了一种具有自适应收敛能力的故障注入方法，有效地计算了CNN面向软错误的冗余可靠性。多个数据集上的多个广泛使用的CNN的测试表明，以Monte Carlo（MC）模型的结果为参考，该方法的可靠性计算精度约为0.9834，比位级逻辑翻转（LFBL）方法高3.0%。此外，该方法对CNN中冗余节点的识别准确性优于HRank方法和$1 \times N$剪枝方法，平均因子分别约为14.32和1.55\times $。此外，该方法设计的冗余保护和自适应层几何中心搜索机制是有效的。因此，所提出的方法可以有效地实现CNN的冗余度和可靠性之间的一一对应。
Redundancy-reliability calculation for deep neural networks is of great importance for their lightweight operation. This article proposes a method for calculating the redundancy-reliability of convolutional neural networks (CNNs) based on strip-median attributes. First, based on the common cause principle, an input-sample-oriented reliability calculation method with automatic boundary detection is constructed to measure the network reliability differences caused by normal classification boundary changes. Next, the adaptive clustering algorithm is used and combined with strip-median attributes to identify the redundant strips of each convolutional layer (CONV) in the networks. Then, based on the central limit theorem and the Pauta criterion, a fault injection method with adaptive convergence ability is constructed to effectively calculate the soft-error-oriented redundancy-reliability of CNNs. The tests of multiple widely used CNNs with multiple datasets showed that with the results of the Monte Carlo (MC) model as the references, the reliability calculation accuracy of the proposed method was approximately 0.9834 and was 3.0% higher than that of the logical flips at bit level (LFBL) method. Additionally, the identification accuracy of the proposed method for redundant nodes in CNNs was better than that of the HRank method and the $1 \times N$ pruning method, by average factors of approximately 14.32 and $1.55\times $ , respectively. Furthermore, the redundant protection and the adaptive layer geometric center searching mechanism designed by the proposed method were effective. Thus, the proposed method can effectively achieve a one-to-one correspondence between the redundancy and reliability of CNNs.