文献类型：期刊文献

英文题名：CKNA: Kernel Hyperparameters Optimization Method for Group-Wise CNNs

作者：Huang, Rongjin[1]; Qu, Shifeng[2]; Yang, Hai[1]; Wang, Zhanquan[1]

机构：[1] Institute of Information Science and Engineering, East China University of Science and Technology, Shanghai, 200237, China; [2] Key Laboratory of Smart Manufacturing in Energy Chemical Process, Ministry of Education, East China University of Science and Technology, Shanghai, 200237, China

年份：2023

卷号：14263 LNCS

起止页码：74

外文期刊名：Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

收录：EI(收录号：20234314952576)

语种：英文

外文关键词：Image enhancement

摘要：Traditional CNNs lack specific interpretable indicators for configuring kernel hyperparameters. Practitioners typically tune model performance via tedious trial and error experiments, resulting in longer training time and expensive computational costs. To address this issue, a novel systematic method, namely CKNA, that considers comprehensive kernel hyperparameters to optimize the network architecture is proposed. Firstly, we innovatively define channel-dispersion and pseudo-kernel to characterize behaviors differentiated by kernel values across model layers. Based on these concepts, the kernel scaling algorithm and the flat wave algorithm are proposed to automatically adjust kernel size and numbers, respectively. Furthermore, we decompose the channel-expand and down-sample operations into adjacent layers to break the bottleneck effect existing in kernel elements. CKNA is applied to classic group-wise CNNs, and experimental results on ImageNet demonstrate that the optimized models obtain at least a 5.73% improvement in classification accuracy. ? 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.