动态权重证据推理规则的CNN超参数质量评估  

Quality Assessment of CNN Hyper-parameters Based on Dynamic Weight Evidential Reasoning Rule

作　　者：李红宇[1] 黄志鹏 张广玲[1] 贺维 LI Hong-yu;HUANG Zhi-peng;ZHANG Guang-ling;HE Wei(College of Computer Science and Information Engineering,Harbin Normal University,Harbin 150025,China;Rocket Force University of Engineering,Xi′an 710025,China)

机构地区：[1]哈尔滨师范大学计算机科学与信息工程学院,哈尔滨150025 [2]中国人民解放军火箭军工程大学,西安710025

出　　处：《小型微型计算机系统》2021年第5期1015-1021,共7页Journal of Chinese Computer Systems

基　　金：黑龙江省自然科学基金项目(F2018023)资助;哈尔滨师范大学博士科研启动金项目(XKB201905)资助;哈尔滨师范大学计算机科学与信息工程学院自然科学基金项目(JKYKYY2020002)资助.

摘　　要：卷积神经网络(Convolutional Neural Networks,CNN)在计算机视觉、自然语言处理为代表的数据处理领域应用广泛.在CNN训练过程中,超参数设置依赖于先验知识,不恰当的超参数会影响模型的训练效率和精度.为了解决这一问题,本文提出基于证据推理规则(Evidential Reasoning Rule,ER Rule)的CNN超参数质量评估模型,可在训练前提供高质量的超参数选择,从而提高模型的训练效率.首先,描述了CNN超参数评估的过程,然后构建了CNN的超参数质量评估模型,运用了奖罚策略实现指标权重的动态适应并给出模型的推理过程.最后,通过图像分类实验验证了模型的实际效果,实验结果表明本文提出的CNN超参数评估模型在多个数据集上预测准确率均高于90%,能够有效建立超参数与CNN模型效果的映射关系,从而过滤低质量的超参数.Convolutional Neural Networks(CNN)are widely used in data processing fields represented by computer vision and natural language processing.In the CNN training process,hyper-parameter setting depends on prior knowledge,and inappropriate hyper-parameters will affect the training efficiency and accuracy of the model.In order to solve this problem,this paper proposes a CNN hyper-parameter quality assessment model based on the Evidential Reasoning Rule(ER Rule),which can provide high-quality hyper-parameter selection before training and can prove the training efficiency of the model.First,the process of CNN hyper-parameter assessment is described,then CNN′s hyper-parameter quality evaluation model is constructed,and reward and penalty strategies are used to achieve the dynamic adaptation of indicator weights and the reasoning process of the model is given.Finally,the actual effect of the model was verified through image classification experiments.The experimental results show that the CNN hyper-parameter assessment model proposed in this paper has a prediction accuracy rate of more than 90%on multiple data sets,which can effectively establish a mapping between hyper-parameters and CNN model effects and can filter the low-quality hyper-parameters.

关 键 词：卷积神经网络(CNN) 超参数评估 证据推理规则(ER Rule) 动态权重 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]