Article(id=1329381598404915242, tenantId=1146029695717560320, journalId=1324759914146529478, issueId=1329381463562236291, articleNumber=null, orderNo=null, doi=10.7544/issn1000-1239.202220803, pmid=null, cstr=null, oa=null, hot=null, price=null, onlineType=0, articleFormat=0, articleType=null, articleTypeStr=null, receivedDate=1663084800000, receivedDateStr=2022-09-14, revisedDate=1676995200000, revisedDateStr=2023-02-22, acceptedDate=null, acceptedDateStr=null, onlineDate=1794878871775, onlineDateStr=2026-11-17, pubDate=1704124800000, pubDateStr=2024-01-02, doiRegisterDate=null, doiRegisterDateStr=null, onlineIssueDate=1794878871775, onlineIssueDateStr=2026-11-17, onlineJustAcceptDate=null, onlineJustAcceptDateStr=null, onlineFirstDate=null, onlineFirstDateStr=null, sourceXml=null, magXml=null, createTime=1794878871775, creator=13701087609, updateTime=1794878871775, updator=13701087609, issue=Issue{id=1329381463562236291, tenantId=1146029695717560320, journalId=1324759914146529478, year='2024', volume='61', issue='1', pageStart='1', pageEnd='270', issueExtLink='null', onlineDate='null', pubDate='1704124800000', pubDateStr='2024-01-02', beforeIssueId=null, nextIssueId=null, price=null, status=1, issueComplete=1, articleOrder=1, issueType=-1, specialIssue=null, createTime=1794878839626, creator='13701087609', updateTime=1794899256120, updator='13701087609', preIssue=null, nextIssue=null, articleTotal=null, ext={EN=IssueExt(id=1329467096993477032, tenantId=1146029695717560320, journalId=1324759914146529478, issueId=1329381463562236291, language=EN, specialIssueTitle=, coverIllustrator=null, specialIssueEditor=, specialIssueAbout=), CN=IssueExt(id=1329467096997671337, tenantId=1146029695717560320, journalId=1324759914146529478, issueId=1329381463562236291, language=CN, specialIssueTitle=, coverIllustrator=null, specialIssueEditor=, specialIssueAbout=)}, issueFiles=null, downloadFileDto=null}, startPage=209, endPage=220, ext={EN=ArticleExt(id=1329381598606241835, articleId=1329381598404915242, tenantId=1146029695717560320, journalId=1324759914146529478, language=EN, title=Optimizing Deep Neural Network Based on Interpretability Analysis, columnId=1325489891789910943, journalTitle=Journal of Computer Research and Development, columnName=Artificial Intelligence, runingTitle=null, highlight=null, articleAbstract=

In recent years, deep neural networks (DNN) have been widely used in many fields, even replacing human to make decisions in some safety-critical systems, such as autonomous driving and smart healthcare, which requires higher reliability of DNN. It is difficult to understand the internal prediction mechanism and debug because of the complex multi-layer nonlinear network structure of DNN. The existing DNN debugging work mainly improves the performance by adjusting the parameters and augments the training set to optimize DNN. However, it is difficult to control the modification range of adjusting parameters directly, and probably make the model lose the ability of fitting the training set. And unguided augmentation of training set will dramatically increase training costs. To address this problem, a DNN optimization method OptDIA (optimizing DNN based on interpretability analysis) is proposed. Interpretability analysis is conducted on the training process and the decision-making behavior of DNN. According to the interpretability analysis results, the original training data is split into different partitions to evaluate their influences on decision-making results of DNN. After that, the partitions of original training data are transformed with different probabilities to generate new training data, which are used to retrain DNN to improve the performance of the model. The experiments on nine DNNs trained by three datasets shows that OptDIA can improve the accuracy of DNNs by 0.39% to 2.15% and F1-score of DNNs by 0.11% to 2.03%.

, authors=Huanhuan Wu1, Ruilin Xie1, Yuanxin Qiao1, Xiang Chen2, Zhanqi Cui1, *, authorsList=Huanhuan Wu, Ruilin Xie, Yuanxin Qiao, Xiang Chen, Zhanqi Cui, authorCompany=null, correspAuthors=Zhanqi Cui, authorNote=

born in 1995. Master. Student member of CCF. Her main research interests include intelligent software engineering and trustworthy artificial intelligence

born in 1996. Master candidate. Student member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

born in 1999. Master candidate. Student member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

born in 1980. PhD, associate professor, master supervisor. Senior member of CCF. His main research interests include software testing, software maintenance, empirical software engineering, and mining software repository

born in 1984. PhD, professor, master supervisor. Senior member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

, correspAuthorsNote=null, copyrightStatement=Copyright © 2024 Journal of Computer Research and Development. All rights reserved., copyrightOwner=null, extLink=null, articleAbsUrl=null, sourceXml=null, magXml=null, pdfUrl=null, pdf=null, pdfFileSize=null, pdfExtLink=null, richHtmlUrl=null, mobilePdfUrl=null, reviewReport=null, pdfFirstPage=null, abstractGraph=null, abstractGraphContent=null, abstractVideo=null, citation=null, cebUrl=null, magXmlContent=null, mapNumber=null, fund=null), CN=ArticleExt(id=1329381600736948285, articleId=1329381598404915242, tenantId=1146029695717560320, journalId=1324759914146529478, language=CN, title=基于可解释性分析的深度神经网络优化方法, columnId=1325489892104483745, journalTitle=计算机研究与发展, columnName=人工智能, runingTitle=null, highlight=null, articleAbstract=

近年来,深度神经网络(deep neural network,DNN)在众多领域取得了广泛应用,甚至在安全关键系统中已经可以代替人类作出决策,如自动驾驶和智慧医疗等,这对DNN的可靠性提出了更高的要求. 由于DNN具有复杂的多层非线性网络结构,难以理解其内部预测机制,也很难对其进行调试. 现有的DNN调试工作主要通过修改DNN的参数和扩增数据集提升DNN性能,以达到优化的目的. 然而直接调整参数难以控制修改的幅度,甚至使模型失去对训练数据的拟合能力;而无指导地扩增训练数据则会大幅增加训练成本. 针对此问题,提出了一种基于可解释性分析的DNN优化方法(optimizing DNN based on interpretability analysis,OptDIA). 对DNN的训练过程及决策行为进行解释分析,根据解释分析结果,将原始数据中对DNN决策行为产生不同程度影响的部分以不同概率进行数据变换以生成新训练数据,并重训练DNN,以提升模型性能达到优化DNN的目的. 在使用3个数据集训练的9个DNN模型上的实验结果表明,OptDIA可以将DNN的准确率提升0.39~2.15个百分点,F1-score提升0.11~2.03个百分点.

, authors=吴欢欢1, 谢瑞麟1, 乔塬心1, 陈翔2, 崔展齐1, *, authorsList=吴欢欢, 谢瑞麟, 乔塬心, 陈翔, 崔展齐, authorCompany=null, correspAuthors=崔展齐, authorNote=

1995年生. 硕士. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

1996年生. 硕士研究生. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

1999年生. 硕士研究生. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

1980年生. 博士,副教授,硕士生导师. CCF高级会员. 主要研究方向为软件测试、软件维护、经验软件工程、软件仓库挖掘

1984年生. 博士,教授,硕士生导师. CCF高级会员. 主要研究方向为智能软件工程、可信人工智能

, correspAuthorsNote=
崔展齐()
, copyrightStatement=版权所有 © 2024 计算机研究与发展编辑部, copyrightOwner=null, extLink=null, articleAbsUrl=null, sourceXml=803bqdWDsDf9xuMqkhx5GQ==, magXml=tAucN0HObgsYfPPon+PU+A==, pdfUrl=null, pdf=Fq2MqUxpCc4/ndNbBhhjdQ==, pdfFileSize=1234475, pdfExtLink=null, richHtmlUrl=null, mobilePdfUrl=null, reviewReport=null, pdfFirstPage=null, abstractGraph=95CABN7Hc7iniVHtYyC3Xw==, abstractGraphContent=null, abstractVideo=null, citation=null, cebUrl=null, magXmlContent=luYm8SPIhbUoUTEwfYtOMQ==, mapNumber=null, fund=null)}, authors=[Author(id=1329381601101852742, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, orderNo=0, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=wuhuanhuan@bistu.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329381601168961608, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601101852742, language=EN, stringName=Huanhuan Wu, firstName=Huanhuan, middleName=null, lastName=Wu, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101, bio={"content":"

born in 1995. Master. Student member of CCF. Her main research interests include intelligent software engineering and trustworthy artificial intelligence

"}, bioImg=null, bioContent=

born in 1995. Master. Student member of CCF. Her main research interests include intelligent software engineering and trustworthy artificial intelligence

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329381601236070473, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601101852742, language=CN, stringName=吴欢欢, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1北京信息科技大学计算机学院 北京 100101, bio={"img":"4U+2cB9uc4nLEoJdMP2GdA==","content":"

1995年生. 硕士. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

"}, bioImg=4U+2cB9uc4nLEoJdMP2GdA==, bioContent=

1995年生. 硕士. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329381600934080574, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, xref=1, ext=[AuthorCompanyExt(id=1329381600942469183, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101), AuthorCompanyExt(id=1329381600950857792, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1北京信息科技大学计算机学院 北京 100101)])]), Author(id=1329381601303179339, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, orderNo=1, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329381601399648333, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601303179339, language=EN, stringName=Ruilin Xie, firstName=Ruilin, middleName=null, lastName=Xie, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101, bio={"content":"

born in 1996. Master candidate. Student member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

"}, bioImg=null, bioContent=

born in 1996. Master candidate. Student member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329381601475145806, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601303179339, language=CN, stringName=谢瑞麟, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1北京信息科技大学计算机学院 北京 100101, bio={"img":"j/qIfD7roeASmMV+BaVvlA==","content":"

1996年生. 硕士研究生. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

"}, bioImg=j/qIfD7roeASmMV+BaVvlA==, bioContent=

1996年生. 硕士研究生. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329381600934080574, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, xref=1, ext=[AuthorCompanyExt(id=1329381600942469183, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101), AuthorCompanyExt(id=1329381600950857792, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1北京信息科技大学计算机学院 北京 100101)])]), Author(id=1329381601546448976, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, orderNo=2, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329381601617752146, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601546448976, language=EN, stringName=Yuanxin Qiao, firstName=Yuanxin, middleName=null, lastName=Qiao, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101, bio={"content":"

born in 1999. Master candidate. Student member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

"}, bioImg=null, bioContent=

born in 1999. Master candidate. Student member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329381601684861011, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601546448976, language=CN, stringName=乔塬心, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1北京信息科技大学计算机学院 北京 100101, bio={"img":"hox6gwjwheZmqsNW27pP6w==","content":"

1999年生. 硕士研究生. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

"}, bioImg=hox6gwjwheZmqsNW27pP6w==, bioContent=

1999年生. 硕士研究生. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329381600934080574, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, xref=1, ext=[AuthorCompanyExt(id=1329381600942469183, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101), AuthorCompanyExt(id=1329381600950857792, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1北京信息科技大学计算机学院 北京 100101)])]), Author(id=1329381601760358485, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, orderNo=3, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329381601840050263, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601760358485, language=EN, stringName=Xiang Chen, firstName=Xiang, middleName=null, lastName=Chen, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=2, address=2School of Information Science and Technology, Nantong University, Nantong, Jiangsu 226019, bio={"content":"

born in 1980. PhD, associate professor, master supervisor. Senior member of CCF. His main research interests include software testing, software maintenance, empirical software engineering, and mining software repository

"}, bioImg=null, bioContent=

born in 1980. PhD, associate professor, master supervisor. Senior member of CCF. His main research interests include software testing, software maintenance, empirical software engineering, and mining software repository

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329381601902964824, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601760358485, language=CN, stringName=陈翔, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=2, address=2南通大学信息科学技术学院 江苏南通 226019, bio={"img":"HYYcBdhrd70RfEud2JLlBQ==","content":"

1980年生. 博士,副教授,硕士生导师. CCF高级会员. 主要研究方向为软件测试、软件维护、经验软件工程、软件仓库挖掘

"}, bioImg=HYYcBdhrd70RfEud2JLlBQ==, bioContent=

1980年生. 博士,副教授,硕士生导师. CCF高级会员. 主要研究方向为软件测试、软件维护、经验软件工程、软件仓库挖掘

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329381601017966658, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, xref=2, ext=[AuthorCompanyExt(id=1329381601022160963, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381601017966658, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2School of Information Science and Technology, Nantong University, Nantong, Jiangsu 226019), AuthorCompanyExt(id=1329381601030549572, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381601017966658, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2南通大学信息科学技术学院 江苏南通 226019)])]), Author(id=1329381601965879386, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, orderNo=4, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=czq@bistu.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=1, authorType=1, ext={EN=AuthorExt(id=1329381602037182556, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601965879386, language=EN, stringName=Zhanqi Cui, firstName=Zhanqi, middleName=null, lastName=Cui, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, *, address=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101, bio={"content":"

born in 1984. PhD, professor, master supervisor. Senior member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

"}, bioImg=null, bioContent=

born in 1984. PhD, professor, master supervisor. Senior member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329381602104291421, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601965879386, language=CN, stringName=崔展齐, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, *, address=1北京信息科技大学计算机学院 北京 100101, bio={"img":"YLuHKVYjEyvD+eeqtz3xuw==","content":"

1984年生. 博士,教授,硕士生导师. CCF高级会员. 主要研究方向为智能软件工程、可信人工智能

"}, bioImg=YLuHKVYjEyvD+eeqtz3xuw==, bioContent=

1984年生. 博士,教授,硕士生导师. CCF高级会员. 主要研究方向为智能软件工程、可信人工智能

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329381600934080574, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, xref=1, ext=[AuthorCompanyExt(id=1329381600942469183, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101), AuthorCompanyExt(id=1329381600950857792, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1北京信息科技大学计算机学院 北京 100101)])])], keywords=[Keyword(id=1329381602209149022, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=EN, orderNo=1, keyword=deep neural network), Keyword(id=1329381602276257887, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=EN, orderNo=2, keyword=DNN optimization), Keyword(id=1329381602347561056, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=EN, orderNo=3, keyword=DNN defect), Keyword(id=1329381602414669921, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=EN, orderNo=4, keyword=DNN performance), Keyword(id=1329381602485973090, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=EN, orderNo=5, keyword=interpretability analysis), Keyword(id=1329381602548887651, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=CN, orderNo=1, keyword=深度神经网络), Keyword(id=1329381602624385124, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=CN, orderNo=2, keyword=DNN优化), Keyword(id=1329381602733437029, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=CN, orderNo=3, keyword=DNN缺陷), Keyword(id=1329381602821517414, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=CN, orderNo=4, keyword=DNN性能), Keyword(id=1329381602897014887, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=CN, orderNo=5, keyword=可解释性分析)], refs=[Reference(id=1329381604797034616, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=1, rfOrder=0, authorNames=null, journalName=null, refType=null, unstructuredReference=Esteva A, Chou K, Yeung S, et al. Deep learning-enabled medical computer vision[J]. NPJ Digital Medicine, 2021, 4(1): 1−9, articleTitle=null, refAbstract=null), Reference(id=1329381604859949177, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=2, rfOrder=1, authorNames=null, journalName=null, refType=null, unstructuredReference=Malik M, Malik M K, Mehmood K, et al. Automatic speech recognition: A survey[J]. Multimedia Tools and Applications, 2021, 80(6): 9411−9457, articleTitle=null, refAbstract=null), Reference(id=1329381604931252346, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=3, rfOrder=2, authorNames=null, journalName=null, refType=null, unstructuredReference=郑海斌,陈晋音,章燕,等. 面向自然语言处理的对抗攻防与鲁棒性分析综述[J]. 计算机研究与发展,2021,58(8):1727−1750, articleTitle=null, refAbstract=null), Reference(id=1329381605010944123, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=3, rfOrder=3, authorNames=null, journalName=null, refType=null, unstructuredReference=Zheng Haibin, Chen Jinyin, Zhang Yan, et al. Survey of adversarial attack, defense and robustness analysis for natural language processing[J]. Journal of Computer Research and Development, 2021, 58(8): 1727−1750 (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1329381605107413116, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=4, rfOrder=4, authorNames=null, journalName=null, refType=null, unstructuredReference=Fang Peng, Wang Zecong, Zhang Xiangyun. Vehicle automatic driving system based on embedded and machine learning[C]//Proc of Int Conf on Computer Vision, Image and Deep Learning. Piscataway, NJ: IEEE, 2020: 281−284, articleTitle=null, refAbstract=null), Reference(id=1329381605174521981, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=5, rfOrder=5, authorNames=null, journalName=null, refType=null, unstructuredReference=Bakator M, Radosav D. Deep learning and medical diagnosis: A review of literature[J]. Multimodal Technologies and Interaction, 2018, 2(3): 47, articleTitle=null, refAbstract=null), Reference(id=1329381605258408062, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=6, rfOrder=6, authorNames=null, journalName=null, refType=null, unstructuredReference=Ghenescu V, Mihaescu R E, Carata S V, et al. Face detection and recognition based on general purpose DNN object detector[C]//Proc of Int Symp on Electronics and Telecommunications. Piscataway, NJ: IEEE, 2018: 1−4, articleTitle=null, refAbstract=null), Reference(id=1329381605333905535, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=7, rfOrder=7, authorNames=null, journalName=null, refType=null, unstructuredReference=Tulbure A A, Tulbure A A, Dulf E H. A review on modern defect detection models using DCNNs-Deep convolutional neural networks[J]. Journal of Advanced Research, 2022, 35: 33−48, articleTitle=null, refAbstract=null), Reference(id=1329381605396820096, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=8, rfOrder=8, authorNames=null, journalName=null, refType=null, unstructuredReference=界面新闻. 新能源汽车评论[EB/OL]. 2022 [2023-02-16].https://www.jiemian.com/article/7530592.html, articleTitle=null, refAbstract=null), Reference(id=1329381605463928961, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=8, rfOrder=9, authorNames=null, journalName=null, refType=null, unstructuredReference=Interface News. Review of new energy vehicles[EB/OL]. 2022 [2023-02-16].https://www.jiemian.com/article/7530592.html (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1329381605556203650, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=9, rfOrder=10, authorNames=null, journalName=null, refType=null, unstructuredReference=Goues C L, Nguyen T V, Forrest S, et al. GenProg: A generic method for automatic software repair[J]. IEEE Transactions on Software Engineering, 2011, 38(1): 54−72, articleTitle=null, refAbstract=null), Reference(id=1329381605631701123, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=10, rfOrder=11, authorNames=null, journalName=null, refType=null, unstructuredReference=Baudry B, Fleurey F, Traon Y L. Improving test suites for efficient fault localization[C]//Proc of the 28th Int Conf on Software Engineering. New York: ACM, 2006: 82−91, articleTitle=null, refAbstract=null), Reference(id=1329381605703004292, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=11, rfOrder=12, authorNames=null, journalName=null, refType=null, unstructuredReference=Zeller A. Yesterday, my program worked. Today, it does not. Why?[J]. ACM SIGSOFT Software Engineering Notes, 1999, 24(6): 253−267, articleTitle=null, refAbstract=null), Reference(id=1329381605761724549, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=12, rfOrder=13, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhang Xiangyu, Gupta N, Gupta R. Locating faults through automated predicate switching[C]//Proc of the 28th Int Conf on Software Engineering. New York: ACM, 2006: 272−281, articleTitle=null, refAbstract=null), Reference(id=1329381605837222022, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=13, rfOrder=14, authorNames=null, journalName=null, refType=null, unstructuredReference=Ma Lei, Xu Felix Juefei, Zhang Fuyuan, et al. Deepgauge: Multi-granularity testing criteria for deep learning systems[C]//Proc of the 33rd ACM/IEEE Int Conf on Automated Software Engineering. New York: ACM, 2018: 120−131, articleTitle=null, refAbstract=null), Reference(id=1329381605946273927, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=14, rfOrder=15, authorNames=null, journalName=null, refType=null, unstructuredReference=Ma Shiqing, Liu Yingqi, Lee W C, et al. MODE: Automated neural network model debugging via state differential analysis and input selection[C]//Proc of the 26th ACM Joint Meeting on European Software Engineering Conf and Symp on the Foundations of Software Engineering. New York: ACM, 2018: 175−186, articleTitle=null, refAbstract=null), Reference(id=1329381606030160008, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=15, rfOrder=16, authorNames=null, journalName=null, refType=null, unstructuredReference=王赞,闫明,刘爽,等. 深度神经网络测试研究综述[J]. 软件学报,2020,31(5):1255−1275, articleTitle=null, refAbstract=null), Reference(id=1329381606122434697, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=15, rfOrder=17, authorNames=null, journalName=null, refType=null, unstructuredReference=Wang Zan, Yan Ming, Liu Shuang, et al. A survey on testing of deep neural network. [J] Journal of Software, 2020, 31(5): 1255−1275 (in Chinese), articleTitle=null, refAbstract=null), Reference(id=1329381606202126474, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=16, rfOrder=18, authorNames=null, journalName=null, refType=null, unstructuredReference=Pei Kexin, Cao Yinzhi, Yang Junfen, et al. Deepxplore: Automated whitebox testing of deep learning systems[C]//Proc of the 26th Symp on Operating Systems Principles. New York: ACM, 2017: 1−18, articleTitle=null, refAbstract=null), Reference(id=1329381606277623947, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=17, rfOrder=19, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhang Mengshi, Zhang Yuqun, Zhang Lingming, et al. DeepRoad: GAN-based metamorphic testing and input validation framework for autonomous driving systems[C]//Proc of the 33rd IEEE/ACM Int Conf on Automated Software Engineering. Piscataway, NJ: IEEE, 2018: 132−142, articleTitle=null, refAbstract=null), Reference(id=1329381606378287244, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=18, rfOrder=20, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhang Hao, Chan W K. Apricot: A weight-adaptation approach to fixing deep learning models[C]//Proc of the 34th IEEE/ACM Int Conf on Automated Software Engineering. Piscataway, NJ: IEEE, 2019: 376−387, articleTitle=null, refAbstract=null), Reference(id=1329381606470561933, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=19, rfOrder=21, authorNames=null, journalName=null, refType=null, unstructuredReference=Selvaraju R R, Cogswell M, Das A, et al. Grad-CAM: Visual explanations from deep networks via gradient-based localization[C]//Proc of the IEEE Int Conf on Computer Vision. Piscataway, NJ: IEEE, 2017: 618−626, articleTitle=null, refAbstract=null), Reference(id=1329381606546059406, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=20, rfOrder=22, authorNames=null, journalName=null, refType=null, unstructuredReference=Simonyan K, Vedaldi A, Zisserman A. Deep inside convolutional networks: Visualising image classification models and saliency maps[J]. arXiv preprint, arXiv: 1312. 6034, 2013, articleTitle=null, refAbstract=null), Reference(id=1329381606617362575, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=21, rfOrder=23, authorNames=null, journalName=null, refType=null, unstructuredReference=Du Mengnan, Liu Ninghao, Song Qinghao, et al. Towards explanation of DNN-based prediction with guided feature inversion[C]//Proc of the 24th ACM SIGKDD Int Conf on Knowledge Discovery & Data Mining. New York: ACM, 2018: 1358−1367, articleTitle=null, refAbstract=null), Reference(id=1329381606692860048, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=22, rfOrder=24, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhou Bolei, Khosla A, Lapedriza A, et al. Learning deep features for discriminative localization[C]//Proc of the IEEE Conf on Computer Vision and Pattern Recognition. Piscataway, NJ: IEEE, 2016: 2921−2929, articleTitle=null, refAbstract=null), Reference(id=1329381606759968913, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=23, rfOrder=25, authorNames=null, journalName=null, refType=null, unstructuredReference=CIFAR official website. Dataset[EB/OL]. [2023-02-16]. http://www.cs.toronto.edu/~kriz/cifar.html, articleTitle=null, refAbstract=null), Reference(id=1329381606818689170, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=24, rfOrder=26, authorNames=null, journalName=null, refType=null, unstructuredReference=GitHub. Fashion-MNIST[EB/OL]. 2017[2023-02-16].https://github.com/zalandoresearch/fashion-mnist, articleTitle=null, refAbstract=null), Reference(id=1329381606885798035, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=25, rfOrder=27, authorNames=null, journalName=null, refType=null, unstructuredReference=Mikołajczyk A, Grochowski M. Data augmentation for improving deep learning in image classification problem[C]//Proc of 2018 Int Interdisciplinary PhD Workshop (IIPhDW). Piscataway, NJ: IEEE, 2018: 117−122, articleTitle=null, refAbstract=null), Reference(id=1329381606961295508, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=26, rfOrder=28, authorNames=null, journalName=null, refType=null, unstructuredReference=Taylor L, Nitschke G. Improving deep learning with generic data augmentation[C]// Proc of IEEE Symp Series on Computational Intelligence. Piscataway, NJ: IEEE, 2018: 1542−1547, articleTitle=null, refAbstract=null), Reference(id=1329381607028404373, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=27, rfOrder=29, authorNames=null, journalName=null, refType=null, unstructuredReference=Kang Guoliang, Dong Xuanyi, Zheng Liang, et al. Patchshuffle regularization[J]. arXiv preprint, arXiv: 1707. 07103, 2017, articleTitle=null, refAbstract=null), Reference(id=1329381607171010710, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=28, rfOrder=30, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhong Zhun, Zheng Liang, Kang Guoliang, et al. Random erasing data augmentation[C]//Proc of the AAAI Conf on Artificial Intelligence. Palo Alto, CA: AAAI, 2020: 13001−13008, articleTitle=null, refAbstract=null), Reference(id=1329381607301034135, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=29, rfOrder=31, authorNames=null, journalName=null, refType=null, unstructuredReference=Zajac M, Zołna K, Rostamzadeh N, et al. Adversarial framing for image and video classification[C]//Proc of the AAAI Conf on Artificial Intelligence. Palo Alto, CA: AAAI, 2019: 10077−10078, articleTitle=null, refAbstract=null), Reference(id=1329381607376531608, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=30, rfOrder=32, authorNames=null, journalName=null, refType=null, unstructuredReference=Johnson J, Alahi A, Li Feifei. Perceptual losses for real-time style transfer and super-resolution[C]//Proc of the 14th European Conf on Computer Vision. Berlin: Springer, 2016: 694−711, articleTitle=null, refAbstract=null), Reference(id=1329381607447834777, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=31, rfOrder=33, authorNames=null, journalName=null, refType=null, unstructuredReference=He Kaiming, Zhang Xiangyu, Ren Shaoqing, et al. Deep residual learning for image recognition[C]//Proc of the IEEE Conf on Computer Vision and Pattern Recognition. Piscataway, NJ: IEEE, 2016: 770−778, articleTitle=null, refAbstract=null), Reference(id=1329381607506555034, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=32, rfOrder=34, authorNames=null, journalName=null, refType=null, unstructuredReference=GitHub. Grad-CAM[EB/OL]. 2022[2023-02-16].https://github.com/jacobgil/pytorch-grad-cam, articleTitle=null, refAbstract=null), Reference(id=1329381607582052507, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=33, rfOrder=35, authorNames=null, journalName=null, refType=null, unstructuredReference=Chen Lei, Chen Jianhui, Hajimirsadeghi H, et al. Adapting Grad-CAM for embedding networks[C]//Proc of the IEEE/CVF Winter Conf on Applications of Computer Vision. Piscataway, NJ: IEEE, 2020: 2794−2803, articleTitle=null, refAbstract=null), Reference(id=1329381607649161372, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=34, rfOrder=36, authorNames=null, journalName=null, refType=null, unstructuredReference=GitHub. Random-Erasing[EB/OL]. 2019[2023-02-16].https://github.com/zhunzhong07/Random-Erasing, articleTitle=null, refAbstract=null), Reference(id=1329381607720464541, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=35, rfOrder=37, authorNames=null, journalName=null, refType=null, unstructuredReference=GitHub. Torchvision[EB/OL]. 2022[2023-02-16].https://github.com/pytorch/vision, articleTitle=null, refAbstract=null), Reference(id=1329381607783379102, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=36, rfOrder=38, authorNames=null, journalName=null, refType=null, unstructuredReference=Tian Yuchi, Ray B. Automatically diagnosing and repairing error handling bugs in C[C]//Proc of the 11th Joint Meeting on Foundations of Software Engineering. New York: ACM, 2017: 752−762, articleTitle=null, refAbstract=null), Reference(id=1329381607858876575, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=37, rfOrder=39, authorNames=null, journalName=null, refType=null, unstructuredReference=Nguyen H D T, Qi Dawei, Roychoudhury A, et al. Semfix: Program repair via semantic analysis[C]//Proc of the 35th Int Conf on Software Engineering. Piscataway, NJ: IEEE, 2013: 772−781, articleTitle=null, refAbstract=null), Reference(id=1329381607917596832, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=38, rfOrder=40, authorNames=null, journalName=null, refType=null, unstructuredReference=Jeffrey D, Feng Min, Gupta N, et al. BugFix: A learning-based tool to assist developers in fixing bugs[C]//Proc of the 17th IEEE Int Conf on Program Comprehension. Piscataway, NJ: IEEE, 2009: 70−79, articleTitle=null, refAbstract=null), Reference(id=1329381608005677217, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=39, rfOrder=41, authorNames=null, journalName=null, refType=null, unstructuredReference=Gao Xiang, Saha R K, Prasad M R, et al. Fuzz testing based data augmentation to improve robustness of deep neural networks[C]//Proc of the 42nd IEEE/ACM Int Conf on Software Engineering. Piscataway, NJ: IEEE, 2020: 1147−1158, articleTitle=null, refAbstract=null), Reference(id=1329381608076980386, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=40, rfOrder=42, authorNames=null, journalName=null, refType=null, unstructuredReference=Liu Meixi, Hong Weijiang, Pan Weiyu, et al. Styx: A data-oriented mutation framework to improve the robustness of DNN[C]//Proc of the 35th IEEE/ACM Int Conf on Automated Software Engineering. Piscataway, NJ: IEEE, 2020: 1260−1261, articleTitle=null, refAbstract=null), Reference(id=1329381608156672163, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=41, rfOrder=43, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhang Hao, Chan W K. Plum: Exploration and prioritization of model repair strategies for fixing deep learning models[C]//Proc of the 8th Int Conf on Dependable Systems and Their Applications. Piscataway, NJ: IEEE, 2021: 140−151, articleTitle=null, refAbstract=null), Reference(id=1329381608232169636, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=42, rfOrder=44, authorNames=null, journalName=null, refType=null, unstructuredReference=Wu Huanhuan, Li Zheng, Cui Zhanqi, et al. GenMuNN: A mutation-based approach to repair deep neural network models[J]. International Journal of Modeling, Simulation, and Scientific Computing, 2022, 13(2): 2341008, articleTitle=null, refAbstract=null), Reference(id=1329381608303472805, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=43, rfOrder=45, authorNames=null, journalName=null, refType=null, unstructuredReference=Zhang Xiaoyu, Zhai Juan, Ma Shiqing, et al. AUTOTRAINER: An automatic DNN training problem detection and repair system[C]//Proc of the 43rd IEEE/ACM Int Conf on Software Engineering. Piscataway, NJ: IEEE, 2021: 359−371, articleTitle=null, refAbstract=null), Reference(id=1329381608357998758, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, doi=null, pmid=null, pmcid=null, year=null, volume=null, issue=null, pageStart=null, pageEnd=null, url=null, language=null, rfNumber=44, rfOrder=46, authorNames=null, journalName=null, refType=null, unstructuredReference=Sun Bing, Sun Jun, Pham L H, et al. Causality-based neural network repair[C]//Proc of the 44th Int Conf on Software Engineering. New York: ACM, 2022: 338−349, articleTitle=null, refAbstract=null)], funds=null, companyList=[AuthorCompany(id=1329381600934080574, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, xref=1, ext=[AuthorCompanyExt(id=1329381600942469183, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101), AuthorCompanyExt(id=1329381600950857792, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1北京信息科技大学计算机学院 北京 100101)]), AuthorCompany(id=1329381601017966658, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, xref=2, ext=[AuthorCompanyExt(id=1329381601022160963, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381601017966658, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2School of Information Science and Technology, Nantong University, Nantong, Jiangsu 226019), AuthorCompanyExt(id=1329381601030549572, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381601017966658, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2南通大学信息科学技术学院 江苏南通 226019)])], figs=[ArticleFig(id=1329381603039621224, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=EN, label=Figure 1, caption=The framework of OptDIA, figureFileSmall=jpUPheGF6YdTcofu+aGJSw==, figureFileBig=95CABN7Hc7iniVHtYyC3Xw==, tableContent=null), ArticleFig(id=1329381603140284521, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=CN, label=图1, caption=基于可解释性分析的DNN优化方法框架, figureFileSmall=jpUPheGF6YdTcofu+aGJSw==, figureFileBig=95CABN7Hc7iniVHtYyC3Xw==, tableContent=null), ArticleFig(id=1329381603324833898, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=EN, label=Figure 2, caption=Example of gradient-weighted class activation map and heat map on different datasets, figureFileSmall=vzSoBueiKiYM6SPC4bxoCQ==, figureFileBig=Cix3RVNnuSj9dIh3to7tcQ==, tableContent=null), ArticleFig(id=1329381603404525675, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=CN, label=图2, caption=不同数据集的梯度加权类激活图和热力图示例, figureFileSmall=vzSoBueiKiYM6SPC4bxoCQ==, figureFileBig=Cix3RVNnuSj9dIh3to7tcQ==, tableContent=null), ArticleFig(id=1329381603500994668, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=EN, label=Figure 3, caption=Example of training data generated based on interpretability analysis on different datasets, figureFileSmall=LPO03vnAD6vMu3c8czhIjg==, figureFileBig=ZFZJo/f1eOWXlt1cO0/wQg==, tableContent=null), ArticleFig(id=1329381603618435181, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=CN, label=图3, caption=不同数据集基于可解释性分析的训练数据生成示例, figureFileSmall=LPO03vnAD6vMu3c8czhIjg==, figureFileBig=ZFZJo/f1eOWXlt1cO0/wQg==, tableContent=null), ArticleFig(id=1329381603740069998, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=EN, label=Table 1, caption=

The Statistics of Datasets

, figureFileSmall=null, figureFileBig=null, tableContent=
数据集总图像数总类别数训练集图像数测试集图像数
CIFAR-1060000105000010000
CIFAR-100600001005000010000
Fashion-MNIST70000106000010000
), ArticleFig(id=1329381603891064943, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=CN, label=表1, caption=

数据集信息

, figureFileSmall=null, figureFileBig=null, tableContent=
数据集总图像数总类别数训练集图像数测试集图像数
CIFAR-1060000105000010000
CIFAR-100600001005000010000
Fashion-MNIST70000106000010000
), ArticleFig(id=1329381603962368112, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=EN, label=Table 2, caption=

Accuracy Comparison of Different DNNs Before and After Optimization by OptDIA %

, figureFileSmall=null, figureFileBig=null, tableContent=
模型CIFAR-10CIFAR-100Fashion-MNIST
Org50Org300OptDIAOrg50Org300OptDIAOrg50Org300OptDIA
ResNet-1885.5685.8386.0949.9549.7850.5092.6093.1093.49
ResNet-3484.5984.5185.3955.2259.1356.6192.9893.5693.51
ResNet-5083.9684.1785.1955.6656.0257.8192.1892.3193.48
), ArticleFig(id=1329381604046254193, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=CN, label=表2, caption=

不同DNN使用OptDIA优化前后的准确率比较

, figureFileSmall=null, figureFileBig=null, tableContent=
模型CIFAR-10CIFAR-100Fashion-MNIST
Org50Org300OptDIAOrg50Org300OptDIAOrg50Org300OptDIA
ResNet-1885.5685.8386.0949.9549.7850.5092.6093.1093.49
ResNet-3484.5984.5185.3955.2259.1356.6192.9893.5693.51
ResNet-5083.9684.1785.1955.6656.0257.8192.1892.3193.48
), ArticleFig(id=1329381604188860530, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=EN, label=Table 3, caption=

Performance Improvements of Optimizing Different DNNs by OptDIA- and OptDIA

, figureFileSmall=null, figureFileBig=null, tableContent=
模型评价指标CIFAR-10CIFAR-100Fashion-MNIST
OptDIA-/%OptDIA/%OptDIA-/%OptDIA/%OptDIA-/%OptDIA/%
ResNet-18Accuracy0.200.530.330.550.530.60
F1-score0.170.410.020.240.110.69
ResNet-34Accuracy0.150.801.121.390.320.39
F1-score0.780.870.720.160.530.11
ResNet-50Accuracy1.261.232.362.151.211.32
F1-score0.160.142.992.031.231.5
平均值Accuracy0.540.851.211.360.690.77
F1-score0.850.901.240.900.821.24
), ArticleFig(id=1329381604314689651, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=CN, label=表3, caption=

不同DNN使用OptDIA-和OptDIA优化性能提升比较

, figureFileSmall=null, figureFileBig=null, tableContent=
模型评价指标CIFAR-10CIFAR-100Fashion-MNIST
OptDIA-/%OptDIA/%OptDIA-/%OptDIA/%OptDIA-/%OptDIA/%
ResNet-18Accuracy0.200.530.330.550.530.60
F1-score0.170.410.020.240.110.69
ResNet-34Accuracy0.150.801.121.390.320.39
F1-score0.780.870.720.160.530.11
ResNet-50Accuracy1.261.232.362.151.211.32
F1-score0.160.142.992.031.231.5
平均值Accuracy0.540.851.211.360.690.77
F1-score0.850.901.240.900.821.24
), ArticleFig(id=1329381604390187124, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=EN, label=Table 4, caption=

Time Costs Comparison of Different DNNs by Using OptDIA- and OptDIA to Generate Training Data

, figureFileSmall=null, figureFileBig=null, tableContent=
模型数据集OptDIA-OptDIA
ResNet-18CIFAR-101 min 05 s2 min 03 s
CIFAR-1001 min 05 s1 min 59 s
Fashion-MNIST1 min 01 s2 min 05 s
ResNet-34CIFAR-101 min 07 s2 min 36 s
CIFAR-1001 min 06 s2 min 34 s
Fashion-MNIST1 min 01 s2 min 41 s
ResNet-50CIFAR-101 min 07 s3 min 09 s
CIFAR-1001 min 06 s3 min 08 s
Fashion-MNIST1 min 03 s3 min 30 s
平均值1 min 04 s2 min 38 s
), ArticleFig(id=1329381604541182069, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=CN, label=表4, caption=

不同DNN使用OptDIA-和OptDIA生成训练数据的时间消耗比较

, figureFileSmall=null, figureFileBig=null, tableContent=
模型数据集OptDIA-OptDIA
ResNet-18CIFAR-101 min 05 s2 min 03 s
CIFAR-1001 min 05 s1 min 59 s
Fashion-MNIST1 min 01 s2 min 05 s
ResNet-34CIFAR-101 min 07 s2 min 36 s
CIFAR-1001 min 06 s2 min 34 s
Fashion-MNIST1 min 01 s2 min 41 s
ResNet-50CIFAR-101 min 07 s3 min 09 s
CIFAR-1001 min 06 s3 min 08 s
Fashion-MNIST1 min 03 s3 min 30 s
平均值1 min 04 s2 min 38 s
), ArticleFig(id=1329381604608290934, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=EN, label=Table 5, caption=

Performance Comparison of Different DNNs Trained by Using RE and Before and After Optimization by OptDIA

, figureFileSmall=null, figureFileBig=null, tableContent=
模型评价指标CIFAR-10CIFAR-100Fashion-MNIST
RE/%OptDIA(本文)/%RE/%OptDIA(本文)/%RE/%OptDIA(本文)/%
ResNet-18Accuracy87.2987.5753.0253.6493.9593.95
F1-score87.2887.4153.3453.7393.9493.94
ResNet-34Accuracy87.6987.8057.3058.4194.0794.07
F1-score87.6687.9457.4658.1294.0593.89
ResNet-50Accuracy85.8386.1556.0258.8193.5993.84
F1-score85.8486.3156.2558.6093.5993.74
), ArticleFig(id=1329381604700565623, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, language=CN, label=表5, caption=

不同DNN使用RE和OptDIA优化前后性能比较

, figureFileSmall=null, figureFileBig=null, tableContent=
模型评价指标CIFAR-10CIFAR-100Fashion-MNIST
RE/%OptDIA(本文)/%RE/%OptDIA(本文)/%RE/%OptDIA(本文)/%
ResNet-18Accuracy87.2987.5753.0253.6493.9593.95
F1-score87.2887.4153.3453.7393.9493.94
ResNet-34Accuracy87.6987.8057.3058.4194.0794.07
F1-score87.6687.9457.4658.1294.0593.89
ResNet-50Accuracy85.8386.1556.0258.8193.5993.84
F1-score85.8486.3156.2558.6093.5993.74
)], attaches=null, journal=Journal(id=1324759606947315908, delFlag=0, nameCn=计算机研究与发展, nameEn=Journal of Computer Research and Development, nameHistory1=null, nameHistory2=null, issn=1000-1239, eissn=null, cn=11-1777/TP, coden=null, periodic=0, language=CN, oaType=null, ccby=null, superviseOffice=null, ownerOffice=null, pubOffice=null, editorOffice=null, officeType=null, aims=null, clcCode=null, officeProv=null, officeCity=null, officeAddr=null, officeZip=null, officeEmail=null, officePhone=null, editDirector=null, officeDirector=null, officeDirectorPhone=null, officeStaffNum=null, officeEmpNum=null, coverPicUrl=k2l8zgqwstuxxevFX2qxVA==, journalPrice=null, startedYear=null, abbrevIsoEn=Journal of Computer Research and Development, journalRemark=null, publicationField=null, createdTime=1793776903141, updatedTime=1793850698612, createdBy=18614031015, updatedBy=13701087609, firstLetterCn=J, firstLetterEn=J, subjectCode=Electronic Information, subjectName=null, subjectCodeEn=Electronic Information, subjectNameEn=null, picCn=k2l8zgqwstuxxevFX2qxVA==, picEn=r6k0Ry+RD+c/VXj5QarmDw==, jcr=null, cjcr=null, exts=[JournalExt(id=1325069127710286688, language=CN, name=计算机研究与发展, nameHistory1=null, nameHistory2=null, managedBy=, sponsoredBy=, publishedBy=, editorOffice=, officeProv=null, officeCity=null, officeAddr=, officeZip=, editDirector=, officeDirector=null, officePhone=null, coverPicUrl=null, journalRemark=, submitArticleUrl=null, websiteUrl=, createdTime=1793850698640, updatedTime=1793850698640, createdBy=13701087609, updatedBy=13701087609, submissionGuidelinesUrl=, submissionAuthorUrl=https://crad.ict.ac.cn/journalx/authorLogOn.action, submissionEditorUrl=https://crad.ict.ac.cn/journalx/editorLogOn.action, submissionReviewUrl=https://crad.ict.ac.cn/journalx/expertLogOn.action, submissionCeEditorUrl=https://crad.ict.ac.cn/journalx/editorInChiefLogOn.action, submissionAeEditorUrl=, option={"copyright":""}), JournalExt(id=1325069127777395553, language=EN, name=Journal of Computer Research and Development, nameHistory1=null, nameHistory2=null, managedBy=, sponsoredBy=, publishedBy=, editorOffice=, officeProv=null, officeCity=null, officeAddr=, officeZip=, editDirector=, officeDirector=null, officePhone=null, coverPicUrl=null, journalRemark=, submitArticleUrl=null, websiteUrl=, createdTime=1793850698656, updatedTime=1793850698656, createdBy=13701087609, updatedBy=13701087609, submissionGuidelinesUrl=, submissionAuthorUrl=https://crad.ict.ac.cn/journalx/authorLogOn.action, submissionEditorUrl=https://crad.ict.ac.cn/journalx/editorLogOn.action, submissionReviewUrl=https://crad.ict.ac.cn/journalx/expertLogOn.action, submissionCeEditorUrl=https://crad.ict.ac.cn/journalx/editorInChiefLogOn.action, submissionAeEditorUrl=, option={"copyright":""})], databaseList=null, tenantJournalId=1324759914146529478, websiteList=[Website(id=1325070300139892771, webName=null, webTitle=null, webDomain=null, webCopyrigh=null, webIpcNo=null, seoTitle=null, seoKeywords=null, seoDescription=null, tenantJournalId=null, journalId=1324759914146529478, journalNameCn=null, journalNameEn=null, grayFlag=null, tenantId=1146029695717560320, platformId=null, journalGroupId=null, journalGroupNameCn=null, journalGroupNameEn=null, type=1, domain=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN, language=CN, createTime=1793850978169, createBy=18614031015, updateTime=1793851040259, updateBy=18614031015, name=计算机研究与发展-中文, tplId=1146099689490845704, title=计算机研究与发展, delFlag=0, indexPage=/home, props=[WebsiteProps(id=1325070893751349360, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=articleTextType, value=kx, createTime=1793851119697, updateTime=1793851119697, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893730377837, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=banner, value=null, createTime=1793851119692, updateTime=1793851119692, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893772320883, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=grayFlag, value=0, createTime=1793851119702, updateTime=1793851119702, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893721989228, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=logo, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/file/pic?fileId=U27tDsy717YF7XSPA+Vwdw==, createTime=1793851119690, updateTime=1793851119690, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893784903797, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=minRunFlag, value=0, createTime=1793851119705, updateTime=1793851119705, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893742960751, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=picServerUrl, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/file/pic, createTime=1793851119695, updateTime=1793851119695, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893780709492, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=silenceFlag, value=0, createTime=1793851119704, updateTime=1793851119704, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893738766446, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=staticResourcePath, value=https://castjournals.cast.org.cn/joweb/cast_kjdb_cn_619/, createTime=1793851119694, updateTime=1793851119694, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893755543665, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=themeColor, value=null, createTime=1793851119698, updateTime=1793851119698, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070893763932274, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300139892771, code=themeStyle, value=null, createTime=1793851119700, updateTime=1793851119700, creator=18614031015, updator=18614031015)]), Website(id=1325070300580294698, webName=null, webTitle=null, webDomain=null, webCopyrigh=null, webIpcNo=null, seoTitle=null, seoKeywords=null, seoDescription=null, tenantJournalId=null, journalId=1324759914146529478, journalNameCn=null, journalNameEn=null, grayFlag=null, tenantId=1146029695717560320, platformId=null, journalGroupId=null, journalGroupNameCn=null, journalGroupNameEn=null, type=1, domain=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN, language=EN, createTime=1793850978274, createBy=18614031015, updateTime=1793851036638, updateBy=18614031015, name=计算机研究与发展-英文, tplId=1146101810881728533, title=Journal of Computer Research and Development, delFlag=0, indexPage=/home, props=[WebsiteProps(id=1325070925288321148, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=articleTextType, value=kx, createTime=1793851127216, updateTime=1793851127216, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925263155321, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=banner, value=null, createTime=1793851127210, updateTime=1793851127210, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925309292671, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=grayFlag, value=0, createTime=1793851127221, updateTime=1793851127221, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925254766712, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=logo, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/file/pic?fileId=U27tDsy717YF7XSPA+Vwdw==, createTime=1793851127208, updateTime=1793851127208, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925326069889, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=minRunFlag, value=0, createTime=1793851127225, updateTime=1793851127225, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925279932539, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=picServerUrl, value=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/file/pic, createTime=1793851127214, updateTime=1793851127214, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925317681280, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=silenceFlag, value=0, createTime=1793851127223, updateTime=1793851127223, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925271543930, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=staticResourcePath, value=https://castjournals.cast.org.cn/joweb/cast_kjdb_en_623/, createTime=1793851127212, updateTime=1793851127212, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925292515453, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=themeColor, value=null, createTime=1793851127217, updateTime=1793851127217, creator=18614031015, updator=18614031015), WebsiteProps(id=1325070925300904062, tenantId=1146029695717560320, journalId=null, journalGroupId=null, siteId=1325070300580294698, code=themeStyle, value=null, createTime=1793851127219, updateTime=1793851127219, creator=18614031015, updator=18614031015)])], journalTitle=计算机研究与发展, weixinUrl=null, journalUrl=https://crad.ict.ac.cn/, iacademicId=null, status=1, seqNo=null, journalTitleEn=Journal of Computer Research and Development, journalPhotoCn=k2l8zgqwstuxxevFX2qxVA==, journalPhotoEn=r6k0Ry+RD+c/VXj5QarmDw==, journalFirstLetter=J, journalRecommend=null, journalNew=null, journalCollection=null, jcrJf=null, cjcrJf=null, jcrJfStr=null, cjcrJfStr=null, submissionFirstDecision=null, sciSubjectClassification=null, casSubjectClassification=null, citeScore=null, totalCitationFrequency=null, icpCode=null, psCode=null, advertisingLicenseCode=null, copyrightInformation=null, country=null, option=, provinceCode=null, provinceName=null, collectFlag=false, interPubPlatform=, interPubPlatformUrl=null), detailUrlCn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/10.7544/issn1000-1239.202220803, detailUrlEn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/10.7544/issn1000-1239.202220803, pdfUrlCn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/PDF/10.7544/issn1000-1239.202220803, pdfUrlEn=https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/PDF/10.7544/issn1000-1239.202220803, aliStartDate=0, aliEndDate=0, collectionFlag=false, citedCount=null, citedUrl=null, previewStatus=0, delFlag=0, hasFullText=1, orderTime=1704124800000, fullTextJson=null, articleText=null, reference=null)
收藏切换
基于可解释性分析的深度神经网络优化方法
吴欢欢 1 , 谢瑞麟 1 , 乔塬心 1 , 陈翔 2 , 崔展齐 1, *
计算机研究与发展 | 人工智能 2024,61(1): 209-220
收起
收藏切换
计算机研究与发展 |人工智能 2024 , 61 (1) : 209 -220
基于可解释性分析的深度神经网络优化方法
全屏
[Author(id=1329381601101852742, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, orderNo=0, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=wuhuanhuan@bistu.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329381601168961608, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601101852742, language=EN, stringName=Huanhuan Wu, firstName=Huanhuan, middleName=null, lastName=Wu, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101, bio={"content":"

born in 1995. Master. Student member of CCF. Her main research interests include intelligent software engineering and trustworthy artificial intelligence

"}, bioImg=null, bioContent=

born in 1995. Master. Student member of CCF. Her main research interests include intelligent software engineering and trustworthy artificial intelligence

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329381601236070473, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601101852742, language=CN, stringName=吴欢欢, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1北京信息科技大学计算机学院 北京 100101, bio={"img":"4U+2cB9uc4nLEoJdMP2GdA==","content":"

1995年生. 硕士. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

"}, bioImg=4U+2cB9uc4nLEoJdMP2GdA==, bioContent=

1995年生. 硕士. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329381600934080574, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, xref=1, ext=[AuthorCompanyExt(id=1329381600942469183, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101), AuthorCompanyExt(id=1329381600950857792, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1北京信息科技大学计算机学院 北京 100101)])]), Author(id=1329381601303179339, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, orderNo=1, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329381601399648333, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601303179339, language=EN, stringName=Ruilin Xie, firstName=Ruilin, middleName=null, lastName=Xie, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101, bio={"content":"

born in 1996. Master candidate. Student member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

"}, bioImg=null, bioContent=

born in 1996. Master candidate. Student member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329381601475145806, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601303179339, language=CN, stringName=谢瑞麟, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1北京信息科技大学计算机学院 北京 100101, bio={"img":"j/qIfD7roeASmMV+BaVvlA==","content":"

1996年生. 硕士研究生. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

"}, bioImg=j/qIfD7roeASmMV+BaVvlA==, bioContent=

1996年生. 硕士研究生. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329381600934080574, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, xref=1, ext=[AuthorCompanyExt(id=1329381600942469183, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101), AuthorCompanyExt(id=1329381600950857792, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1北京信息科技大学计算机学院 北京 100101)])]), Author(id=1329381601546448976, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, orderNo=2, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329381601617752146, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601546448976, language=EN, stringName=Yuanxin Qiao, firstName=Yuanxin, middleName=null, lastName=Qiao, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101, bio={"content":"

born in 1999. Master candidate. Student member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

"}, bioImg=null, bioContent=

born in 1999. Master candidate. Student member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329381601684861011, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601546448976, language=CN, stringName=乔塬心, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, address=1北京信息科技大学计算机学院 北京 100101, bio={"img":"hox6gwjwheZmqsNW27pP6w==","content":"

1999年生. 硕士研究生. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

"}, bioImg=hox6gwjwheZmqsNW27pP6w==, bioContent=

1999年生. 硕士研究生. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329381600934080574, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, xref=1, ext=[AuthorCompanyExt(id=1329381600942469183, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101), AuthorCompanyExt(id=1329381600950857792, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1北京信息科技大学计算机学院 北京 100101)])]), Author(id=1329381601760358485, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, orderNo=3, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=null, emailSecond=null, emailThird=null, correspondingAuthor=0, authorType=1, ext={EN=AuthorExt(id=1329381601840050263, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601760358485, language=EN, stringName=Xiang Chen, firstName=Xiang, middleName=null, lastName=Chen, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=2, address=2School of Information Science and Technology, Nantong University, Nantong, Jiangsu 226019, bio={"content":"

born in 1980. PhD, associate professor, master supervisor. Senior member of CCF. His main research interests include software testing, software maintenance, empirical software engineering, and mining software repository

"}, bioImg=null, bioContent=

born in 1980. PhD, associate professor, master supervisor. Senior member of CCF. His main research interests include software testing, software maintenance, empirical software engineering, and mining software repository

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329381601902964824, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601760358485, language=CN, stringName=陈翔, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=2, address=2南通大学信息科学技术学院 江苏南通 226019, bio={"img":"HYYcBdhrd70RfEud2JLlBQ==","content":"

1980年生. 博士,副教授,硕士生导师. CCF高级会员. 主要研究方向为软件测试、软件维护、经验软件工程、软件仓库挖掘

"}, bioImg=HYYcBdhrd70RfEud2JLlBQ==, bioContent=

1980年生. 博士,副教授,硕士生导师. CCF高级会员. 主要研究方向为软件测试、软件维护、经验软件工程、软件仓库挖掘

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329381601017966658, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, xref=2, ext=[AuthorCompanyExt(id=1329381601022160963, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381601017966658, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2School of Information Science and Technology, Nantong University, Nantong, Jiangsu 226019), AuthorCompanyExt(id=1329381601030549572, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381601017966658, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=2南通大学信息科学技术学院 江苏南通 226019)])]), Author(id=1329381601965879386, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, orderNo=4, firstName=null, middleName=null, lastName=null, nameCn=null, orcid=null, stid=null, country=null, authorPic=null, dead=0, email=czq@bistu.edu.cn, emailSecond=null, emailThird=null, correspondingAuthor=1, authorType=1, ext={EN=AuthorExt(id=1329381602037182556, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601965879386, language=EN, stringName=Zhanqi Cui, firstName=Zhanqi, middleName=null, lastName=Cui, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, *, address=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101, bio={"content":"

born in 1984. PhD, professor, master supervisor. Senior member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

"}, bioImg=null, bioContent=

born in 1984. PhD, professor, master supervisor. Senior member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

, aboutCorrespAuthor=null), CN=AuthorExt(id=1329381602104291421, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, authorId=1329381601965879386, language=CN, stringName=崔展齐, firstName=null, middleName=null, lastName=null, prefix=null, suffix=null, authorComment=null, nameInitials=null, affiliation=null, department=null, xref=1, *, address=1北京信息科技大学计算机学院 北京 100101, bio={"img":"YLuHKVYjEyvD+eeqtz3xuw==","content":"

1984年生. 博士,教授,硕士生导师. CCF高级会员. 主要研究方向为智能软件工程、可信人工智能

"}, bioImg=YLuHKVYjEyvD+eeqtz3xuw==, bioContent=

1984年生. 博士,教授,硕士生导师. CCF高级会员. 主要研究方向为智能软件工程、可信人工智能

, aboutCorrespAuthor=null)}, companyList=[AuthorCompany(id=1329381600934080574, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, xref=1, ext=[AuthorCompanyExt(id=1329381600942469183, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=EN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101), AuthorCompanyExt(id=1329381600950857792, tenantId=1146029695717560320, journalId=1324759914146529478, articleId=1329381598404915242, companyId=1329381600934080574, language=CN, country=null, province=null, city=null, postcode=null, companyName=null, departmentName=null, remark=1北京信息科技大学计算机学院 北京 100101)])])]
吴欢欢1 , 谢瑞麟1, 乔塬心1, 陈翔2, 崔展齐1, *
作者信息
  • 1北京信息科技大学计算机学院 北京 100101
  • 2南通大学信息科学技术学院 江苏南通 226019
通讯作者:
崔展齐()
作者简介:

1995年生. 硕士. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

1996年生. 硕士研究生. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

1999年生. 硕士研究生. CCF学生会员. 主要研究方向为智能软件工程、可信人工智能

1980年生. 博士,副教授,硕士生导师. CCF高级会员. 主要研究方向为软件测试、软件维护、经验软件工程、软件仓库挖掘

1984年生. 博士,教授,硕士生导师. CCF高级会员. 主要研究方向为智能软件工程、可信人工智能

Optimizing Deep Neural Network Based on Interpretability Analysis
Huanhuan Wu1 , Ruilin Xie1, Yuanxin Qiao1, Xiang Chen2, Zhanqi Cui1, *
Affiliations
  • 1School of Computer Science, Beijing Information Science and Technology University, Beijing 100101
  • 2School of Information Science and Technology, Nantong University, Nantong, Jiangsu 226019
  • born in 1995. Master. Student member of CCF. Her main research interests include intelligent software engineering and trustworthy artificial intelligence

    born in 1996. Master candidate. Student member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

    born in 1999. Master candidate. Student member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

    born in 1980. PhD, associate professor, master supervisor. Senior member of CCF. His main research interests include software testing, software maintenance, empirical software engineering, and mining software repository

    born in 1984. PhD, professor, master supervisor. Senior member of CCF. His main research interests include intelligent software engineering and trustworthy artificial intelligence

出版时间: 2024-01-02 doi: 10.7544/issn1000-1239.202220803
文章导航
收藏切换

近年来,深度神经网络(deep neural network,DNN)在众多领域取得了广泛应用,甚至在安全关键系统中已经可以代替人类作出决策,如自动驾驶和智慧医疗等,这对DNN的可靠性提出了更高的要求. 由于DNN具有复杂的多层非线性网络结构,难以理解其内部预测机制,也很难对其进行调试. 现有的DNN调试工作主要通过修改DNN的参数和扩增数据集提升DNN性能,以达到优化的目的. 然而直接调整参数难以控制修改的幅度,甚至使模型失去对训练数据的拟合能力;而无指导地扩增训练数据则会大幅增加训练成本. 针对此问题,提出了一种基于可解释性分析的DNN优化方法(optimizing DNN based on interpretability analysis,OptDIA). 对DNN的训练过程及决策行为进行解释分析,根据解释分析结果,将原始数据中对DNN决策行为产生不同程度影响的部分以不同概率进行数据变换以生成新训练数据,并重训练DNN,以提升模型性能达到优化DNN的目的. 在使用3个数据集训练的9个DNN模型上的实验结果表明,OptDIA可以将DNN的准确率提升0.39~2.15个百分点,F1-score提升0.11~2.03个百分点.

深度神经网络  /  DNN优化  /  DNN缺陷  /  DNN性能  /  可解释性分析

In recent years, deep neural networks (DNN) have been widely used in many fields, even replacing human to make decisions in some safety-critical systems, such as autonomous driving and smart healthcare, which requires higher reliability of DNN. It is difficult to understand the internal prediction mechanism and debug because of the complex multi-layer nonlinear network structure of DNN. The existing DNN debugging work mainly improves the performance by adjusting the parameters and augments the training set to optimize DNN. However, it is difficult to control the modification range of adjusting parameters directly, and probably make the model lose the ability of fitting the training set. And unguided augmentation of training set will dramatically increase training costs. To address this problem, a DNN optimization method OptDIA (optimizing DNN based on interpretability analysis) is proposed. Interpretability analysis is conducted on the training process and the decision-making behavior of DNN. According to the interpretability analysis results, the original training data is split into different partitions to evaluate their influences on decision-making results of DNN. After that, the partitions of original training data are transformed with different probabilities to generate new training data, which are used to retrain DNN to improve the performance of the model. The experiments on nine DNNs trained by three datasets shows that OptDIA can improve the accuracy of DNNs by 0.39% to 2.15% and F1-score of DNNs by 0.11% to 2.03%.

deep neural network  /  DNN optimization  /  DNN defect  /  DNN performance  /  interpretability analysis
吴欢欢, 谢瑞麟, 乔塬心, 陈翔, 崔展齐. 基于可解释性分析的深度神经网络优化方法. 计算机研究与发展, 2024 , 61 (1) : 209 -220 . DOI: 10.7544/issn1000-1239.202220803
Huanhuan Wu, Ruilin Xie, Yuanxin Qiao, Xiang Chen, Zhanqi Cui. Optimizing Deep Neural Network Based on Interpretability Analysis[J]. Journal of Computer Research and Development, 2024 , 61 (1) : 209 -220 . DOI: 10.7544/issn1000-1239.202220803
近年来,深度神经网络(deep neural network,DNN)被成功应用于计算机视觉[1]、语音识别[2]、自然语言处理[3]等领域. 在自动驾驶[4]和智慧医疗[5]等场景下,DNN甚至可以代替人类完成决策. 例如汽车制造商的自动驾驶技术[4]、监控系统常用的人脸识别技术[6]、工业制造使用的缺陷检测技术[7]等. 随着DNN应用领域的不断扩大,人们对DNN的质量提出了更高要求. 尤其在一些安全攸关领域,如果DNN的质量达不到要求,或将引发灾难性后果. 据NHTSA统计[8],在2016—2021年,仅与特斯拉自动驾驶系统有关的8起交通事故中就导致了10人死亡. 因此,如何有效保障DNN的质量已成为智能软件领域的一个重要研究课题.
与传统软件类似,DNN也不可避免地含有缺陷. 对于传统软件的质量保障问题,已有大量工作围绕缺陷检测[9]、缺陷定位[10]、增量调试[11]和自动错误修复[12]等技术展开研究,并取得了良好的效果. 但上述质量保障技术无法直接适用于DNN. 传统软件通常根据软件需求设计编写程序的内部逻辑,DNN则是在定义好模型结构后使用数据集对模型进行训练,使其学习输入数据到预测输出的高维映射关系. DNN的内部行为逻辑取决于其结构、训练数据及训练得出的神经元权重等[13]. 因此,需要根据DNN的特点研究有针对性的质量保障方法.
在DNN质量保障研究工作中,通常认为准确率低于理想值就称为DNN存在缺陷[14]. 目前已有许多工作对DNN缺陷检测技术进行研究[15]. 例如Pei等人[16]根据DNN的结构特点提出了一种白盒差分测试技术,在测试用例生成过程中将系统行为差异和高神经元覆盖率作为优化目标,取得了较高的神经元覆盖率,并在Udacity自动驾驶汽车挑战赛数据集中发现了撞击护栏等数千个错误行为. Zhang等人[17]提出了基于对抗式生成网络(generative adversarial network,GAN)的测试数据生成方法DeepRoad,通过生成不同天气条件下的驾驶场景来检测基于DNN的自动驾驶系统的不一致行为.
在检测到DNN中存在的缺陷后,更为重要的是针对所发现的缺陷优化DNN,以提升其性能. 现有的DNN优化工作主要通过扩增训练数据和优化DNN参数2种方式进行. 扩增训练数据是对训练数据进行扩增,使用扩增后的数据训练DNN以提升性能. 例如Ma等人[14]提出了一种基于状态差分分析和输入选择的自动神经网络调试技术MODE,该方法可帮助识别有缺陷的神经元,测量它们的重要性并使用GAN指导生成具有较高真实性的训练数据,以提高模型性能. 优化DNN参数则直接调整DNN中的权重等参数以提升DNN性能. 例如,Zhang等人[18]提出了一种DNN修复方法Apricot,其思想是使用部分原始数据集训练生成一组简化的DNN,然后利用它们提供搜索方向,以调整待修复DNN的权重值. 上述2种方法均达到了较好的优化效果,如Apricot方法在所有模型上的准确率平均提升了1.08个百分点,提升的范围是1.00~2.15个百分点. MODE方法的过程较为复杂且需要较高的训练成本,而Apricot方法的修复过程需要构建训练多个DNN来提供搜索方向,这将提高修复成本,且修复过程不具有可解释性.
针对上述问题,本文提出了一种基于可解释性分析的DNN优化方法(optimizing DNN based on interpretability analysis,OptDIA). 首先,利用DNN解释方法计算不同区域对DNN预测结果的贡献值,获取不同区域对DNN预测结果产生影响的重要程度. 然后,对重要区域以不同概率使用遮挡、变异等方式进行数据变换以扩增训练数据集. 其中,重要程度越低的区域进行数据变换的概率越高. 最后,使用扩增后训练数据集再训练DNN以进行优化. 实验结果表明,OptDIA可以将DNN的准确率提升0.39~2.15个百分点,F1-score提升0.11~2.03个百分点.
本文的主要贡献有2个方面:
1) 提出了一种基于可解释性分析的DNN优化方法OptDIA,使用DNN解释方法指导扩增训练数据,并通过再训练优化DNN,以达到提升其性能的目的.
2) 为验证OptDIA的有效性,本文实现了原型工具,对9个DNN模型在3个开源数据集上进行实验,并对OptDIA优化模型性能的能力进行验证.
目前,缺乏可解释性已经成为制约机器学习算法研究、开发和实现的一个关键问题. 随着DNN模型越来越多地用于自动驾驶、智慧医疗等安全攸关领域,人们对使用解释方法分析DNN的需求越来越大. 虽然已经提出了许多DNN解释方法,但使用解释方法分析DNN是否达到预期效果及提升DNN性能的研究相对较少. 由于训练数据会影响DNN性能,若使用解释方法对DNN预测过程进行分析,以理解其预测正确或错误的原因,可获取每条训练数据不同区域对DNN预测结果的影响;并进行有针对性的数据变换,可增加训练数据的多样性,提升模型性能. 为此,本文提出了基于可解释性分析的DNN优化方法OptDIA.
OptDIA的框架图如图1所示,主要包括2个部分. 首先使用解释方法对DNN的训练过程和决策行为进行分析以生成梯度加权类激活图[19],梯度加权类激活图可显示原始数据中不同区域对DNN预测结果的重要程度. 若某块区域重要程度越高,则表示该区域所对应原始数据中的特征对预测结果影响越大(将在1.2节中详细介绍);随后是基于可解释性分析的训练数据生成阶段,该阶段按一定概率选取训练数据,根据梯度加权类激活图进行数据变换操作生成新训练数据,对DNN再训练并记录训练过程中的性能,在达到停止条件时输出性能最佳的DNN,并将其作为优化后的DNN(将在1.3节中详细介绍).
由于DNN具有复杂的多层非线性网络结构,难以全面理解其预测机制,因此常被当作黑盒使用,这也导致难以确认DNN产生缺陷的原因. 对DNN预测过程的内部工作机制进行深入分析,度量不同特征对预测结果的影响程度,才能有针对性地优化DNN. 因此,OptDIA中的一个关键步骤是对DNN的预测过程进行解释分析,以了解训练数据不同区域影响DNN行为的重要性差异.
目前,已有多种解释方法可用于理解DNN的内部工作机制,例如基于反向传播的Grad[20]、基于特征反演的Guide Inversion[21]、基于类激活映射的CAM(class activation mapping)[22]和基于梯度加权类激活映射的Grad-CAM(gradient-weighted CAM)[19]等. 相比于其他DNN解释方法,将梯度信息与特征映射相结合的Grad-CAM无需修改网络架构或重训练模型,可避免导致DNN准确率下降,并适用于不同任务以及多种结构的DNN模型. 因此,OptDIA采用Grad-CAM方法对DNN进行解释分析和可视化.
OptDIA首先会对训练数据进行解释分析. 具体地:在DNN的正向传播过程中,将最后一个卷积层作为Grad-CAM的目标层,并得到预测结果未经过Softmax激活之前的预测值,再通过反向传播计算目标层所输出的各个特征图对于预测值的梯度,将梯度全局平均池化后得到特征图对于预测结果类别的贡献值. 目标层特征图对于预测值的贡献值计算方法如式(1)[19]:
$ \alpha _k^c = \frac{1}{Z}\sum\limits_i {\sum\limits_j {\dfrac{{\partial {y^{\;c}}}}{{\partial A_{ij}^k}}} } . $
其中,c表示类别,$y^{\;c} $是在经过Softmax激活之前DNN对类别c的预测值,Ak表示目标层的第k个特征图,i,j分别为特征图Ak的单个像素值在横轴和纵轴的索引,Z为特征图Ak的长和宽的乘积. $ a_k^c $即第k个特征图对于类别c的贡献值.
接下来,将特征图对于预测类别的贡献值作为权重,对目标层的所有特征图进行加权求和,并通过函数ReLU将所有负值取0,消除一些与目标类别无关的干扰,从而获得一个粗粒度的梯度加权类激活图,以用于定位每条训练数据中具有类判别性的重要区域. 其中,类判别性指的是训练数据中对DNN预测提供重要依据的特征. 类别c的梯度加权类激活图计算方法如式(2)所示[19]:
$ L_{{\mathrm{Grad {\text{-}} CAM}}}^c = ReLU(\sum\limits_k {a_k^c{A^k}} ) . $
在获取梯度加权类激活图后,可将其与原始数据叠加生成热力图,以更加直观地了解不同区域对DNN预测结果的重要性差异. 图2为解释方法Grad-CAM生成的梯度加权类激活图,以及与原始数据叠加生成的热力图示例. 图2中每一行的图片分别为来自CIFAR-10,CIFAR-100[23],Fashion-MNIST[24]数据集中的训练数据,标签分别为“bird”“man”“t-shirt”. 图2中第1列为原始数据;第2列为使用Grad-CAM所生成的梯度加权类激活图;第3列为将梯度加权类激活图与原始数据叠加生成的热力图,其中颜色越深的区域对DNN分类结果影响越大. 从图2可以看出;对于图片“bird”,对DNN分类结果起重要作用的部分为其纺锤形的体型特征;对于图片“man”,对DNN分类结果起重要作用的部分为其身体特征;对于图片“t-shirt”,对DNN分类结果起重要作用的部分为衣服袖口.
DNN性能受模型结构、超参数和训练数据等因素影响,在模型结构和超参数确定的情况下,训练数据的质量和多样性等因素也会影响DNN性能. 因此如何提高训练数据的质量及多样性是OptDIA的关键. 目前,几何变换[25]、颜色空间增强[26]、核滤波器[27]、随机擦除[28]、对抗训练[29]和神经风格转移[30]等数据增强方法常用于优化训练数据. 但是这些增强训练数据方法通常被研究者用来解决小样本训练问题. OptDIA将基于解释分析结果应用数据增强方法提高训练数据质量,以优化DNN,进而提升其性能.
基于可解释性分析的训练数据生成步骤如算法1所示.
算法1. 基于可解释性分析的训练数据生成.
输入:训练集trainSet,测试集testSet,原始数据索引imgId,待优化DNN M;
输出:优化后DNN M'.
① for each i in imgId
②  camMap = GradCAM(M, trainSet[i]);
③   areaRank = getImportantArea(camMap);
④  trainSet[i] = Augment(areaRank, trainSet[i]);
⑤ end for
⑥ models = reTrain(trainSet, testSet, M);
⑦ bestAcc = 0;
⑧ for each m in models
⑨  acc = validate(testSet, m);
⑩  if acc > bestAcc
⑪   bestAcc = acc;
⑫    M' = m;
⑬  end if
⑭ end for
⑮ return M'.
算法1中,首先将训练数据输入到DNN进行解释分析,以生成梯度加权类激活图camMap(行①②). 然后,根据camMap统计固定大小区域的热力值以获取原始数据中对预测结果影响重要性程度区域排序araRank(行③),其中,热力值指训练数据中不同区域对DNN预测结果的重要性程度,值越大越重要. 接着,根据原始数据中的重要性程度区域排序areaRank使用数据增强方法Augment对原始数据中的一个区域进行数据变换(其中重要程度排序越低的区域被选择进行数据变换的概率越高),并将所生成的数据替换训练数据集中的原始数据(行④),其中变换数据方法可以采用随机擦除[28]等数据增强技术. 训练数据替换完成后,对DNN进行重训练并验证每个训练轮次中DNN的准确率(行⑧⑨). 最后,在达到预设定的重训练轮数后停止优化,并从中选择准确率最高的DNN M'作为优化后的DNN输出(行⑩~⑮). 此算法的最坏时间复杂度为O(n),其中n为数据集规模.
基于算法1,通过使用数据增强方法对梯度加权类激活图中的不同区域按重要程度和不同概率选择进行变换生成新的训练数据,以提高训练数据的质量及多样性,并使用生成的数据再次训练DNN以提升其性能完成优化. 图3为OptDIA方法生成新数据示例. 其中,第1列为对来自CIFAR-10,CIFAR-100,Fashion-MNIST数据集中的原始数据使用Grad-CAM所生成的梯度加权类激活图,标签分别为“bird”“man”“t-shirt”. 第2列为根据梯度加权类激活图生成的新数据,用于再训练. 梯度加权类激活图用作变换数据范围的依据以生成新数据,将对DNN预测起重要作用的区域以较小的概率变换,而其他对DNN预测结果不重要的区域以较大的概率进行数据变换,来让DNN学习更多重要区域的特征. 例如对于“bird”,让DNN学习更多身体部位特征;对于“man”,让DNN学习更多人体形态特征;对于“t-shirt”,让DNN学习更多衣服特征. OptDIA通过解释分析对预测结果不重要的特征进行变换,来让DNN学习更多的重要特征,从而正确分类测试数据.
我们基于所提出的OptDIA方法,实现了DNN优化原型工具. 在原型工具中,通过解释分析对原始数据中的区域进行重要性排序,排序越高的区域被选择使用随机擦除方法[28]进行数据变换的概率越低. 其中,随机擦除指的是对被选择进行数据变换的区域用随机值像素替换以生成具有不同遮挡程度的训练图像,每条数据被选择擦除的概率沿用文献[28]中的0.1. 为确定实验所用的擦除区域大小,分别将图片区域的分割方式设置为64,16,4个大小相等的矩形区域,初步实验结果表明,不同区域大小对OptDIA的优化效果并无显著影响. 在实验中我们选择将图像等分为16个大小相等的矩形区域,CIFAR-10和CIFAR-100数据集的矩形大小为8×8,Fashion-MNIST数据集的矩形大小为7×7. 原型工具的开发和运行环境为Python3.6和Pytorch1.12.
本文通过3个研究问题来验证OptDIA的有效性.
问题1. OptDIA是否能优化DNN以提升DNN性能.
OptDIA是一种基于可解释性分析的DNN优化方法,为了评估OptDIA的有效性,设计了问题1验证OptDIA在不同数据集训练的不同DNN模型上提升DNN性能的有效性.
问题2. 解释方法的引入是否提高了OptDIA优化DNN的有效性.
解释方法的引入是OptDIA区别于其他DNN优化方法的重要步骤,为了评估解释方法的引入对DNN优化能力的影响,设计了问题2验证OptDIA在DNN优化过程中,使用解释方法选择数据变换区域和使用随机方法选择数据变换区域对优化效果的影响.
问题3. OptDIA能否在已使用数据增强方法进行训练的模型上继续提升DNN性能.
目前,数据增强方法已经能有效地在训练过程中对训练数据进行扩增,以提高模型的泛化能力并达到较高的性能. 问题3验证OptDIA是否能进一步提升应用了数据增强方法进行训练的DNN性能.
为回答2.1节的3个问题,实验将OptDIA在使用CIFAR-10,CIFAR-100,Fashion-MNIST数据集训练的DNN上进行验证和分析. 所使用的数据集被广泛应用于提升DNN性能的工作中[14,18,28]. 其中,CIFAR-10数据集由10类共60000张尺寸为32×32的3通道彩色图像组成,其中包含50000张训练图像和10000张测试图像. CIFAR-100具有和CIFAR-10相同的数据格式和规模,2个数据集的区别在于类别数不同,CIFAR-10具有10种分类类别,而CIFAR-100具有100种分类类别. Fashion-MNIST由10类共70000张尺寸为28×28的单通道灰度图像组成,其中包含60000张训练图像和10000张测试图像. 数据集的详细信息如表1所示. 数据集已经被划分好训练集及测试集,实验使用其中的训练集训练,测试集用来测试及评估优化效果.
我们使用CIFAR-10,CIFAR-100,Fashion-MNIST数据集训练了3种模型:分别使用18层、34层和50层结构的ResNet[31]. 优化器学习率从0.01开始每训练20轮下降3倍,在第50轮停止训练. 将这些完成训练的DNN作为待优化模型. 为减少随机性的影响,所有实验均重复运行3次后取平均值.
在实验中使用准确率Accuracy和F1-score来评价优化效果,以反映DNN的拟合能力和预测未知样本的能力.
对于二分类问题,其中正样本被正确预测称为TP,负样本被错误预测称为FP,负样本被正确预测称为TN,正样本被错误预测称为FN. 式(3)为准确率Accuracy的计算方式,即所有样本中正确预测为正样本占所有样本的比例;式(4)为精确率Precision的定义,即预测的正样本中实际是正样本的个数占被预测为正样本的比例;式(5)为召回率Recall的定义,即所有样本中被预测为正样本占所有正样本的比例;式(6)为F1-score的定义,同时考虑精确率和召回率. 实验中使用的CIFAR-10,CIFAR-100,Fashion-MNIST数据分别有10,100,10个类别. 对于多分类问题,可将每个数据集中的类两两组合,分别计算每个类和其他类之间的Accuracy和F1-score,再计算平均值即可.
$ Accuracy = \frac{{TP + TN}}{{TP + TN + FN + FP}} , $
$ {{Precision}} = \frac{{TP}}{{TP + FP}} , $
$ {{Recall = }}\frac{{TP}}{{TP + FN}} , $
$ {{F1 {\text{-}} score}} = \frac{{2 \times {{Precision}} \times {{Recall}}}}{{{{Precision}} + {{Recall}}}} . $
为了评估OptDIA的有效性,在实验中使用固定大小的滑动窗口扫描原始数据,对每个矩形区域统计其在加权梯度类激活图中的贡献值,并根据贡献值排序,贡献值越低的区域进行数据变换的概率越高. 实验将分别在3个数据集上预训练50轮的9个DNN模型作为待优化模型,再使用OptDIA重训练50轮进行优化,并将预训练300轮的DNN模型的准确率与优化后DNN模型的准确率进行对比.
实验结果如表2所示,其中Org50和Org300分别表示优化前预训练50轮和300轮DNN的准确率,OptDIA表示优化后DNN的准确率. 实验结果表明,OptDIA可提升使用3个数据集训练的9个DNN的准确率以进行优化. 其中,对于使用CIFAR-10训练的ResNet-50模型,OptDIA将准确率提升了1.23个百分点;对于使用CIFAR-100训练的ResNet-50模型,OptDIA将准确率提升了2.15个百分点;对于使用Fashion-MNIST训练的ResNet-50模型,OptDIA将准确率提升了1.32个百分点. 实验结果表明,在3个数据集上训练的9个模型中,有7个使用OptDIA优化50轮后所提升的准确率比直接预训练300轮模型的准确率高. 其中,对于使用CIFAR-10训练的ResNet-50模型,优化50轮后的准确率要比预训练300轮模型准确率高1.02个百分点;对于使用CIFAR-100训练的ResNet-50模型,优化50轮后的准确率要比预训练300轮模型准确率高1.79个百分点;对于使用Fashion-MNIST训练的ResNet-18模型,优化50轮后的准确率要比预训练300轮模型准确率高0.39个百分点. 其中,对于初始准确率较低的模型,准确率提升空间比较大,OptDIA提升准确率的幅度也相对较大. 如使用CIFAR-100训练的ResNet-50模型,初始准确率55.66%,OptDIA将其准确率提升了2.15个百分点;对于初始准确率较高的模型,其准确率提升空间较小,OptDIA提升准确率的幅度也相对较小. 如使用Fashion-MNIST训练的ResNet-34模型,初始准确率已达92.98%,OptDIA仅将其准确率提升了0.53个百分点.
从表2可以看出,对于3种DNN模型,使用OptDIA均可提升其准确率,甚至相比较预训练300轮的DNN模型,预训练50轮后使用OptDIA优化的DNN模型准确率更高. 对实验结果进行分析发现,这是因为OptDIA使用解释方法对DNN的预测过程进行分析,获取每条训练数据中不同区域特征对DNN结果预测的重要程度,并对越不重要的区域以越大的概率进行变换生成新数据,增加训练数据的多样性,从而提升DNN模型的性能.
对问题1的回答:对于使用不同数据集和结构训练的DNN,OptDIA均可有效提升其准确率,提升幅度为0.39~2.15个百分点,所有模型的准确率平均提升了1.04个百分点,即使与训练300轮的DNN模型相比,使用OptDIA优化的模型准确率也更高.
为分析OptDIA中解释方法对DNN模型优化效果的影响,实验在不同结构的DNN和数据集上分别使用OptDIA-和OptDIA优化方法,并对所提升的性能进行比较. 其中,OptDIA-指没有使用Grad-CAM对数据进行解释分析,而是直接对训练数据进行随机擦除操作以扩增训练数据. 实验在预训练50轮的待优化DNN基础上,分别用OptDIA-和OptDIA再训练50轮,并对比所提升性能的差异.
表3给出了OptDIA-和OptDIA使用3个数据集训练的9个DNN模型上提升性能的情况. 如表3所示,除ResNet-50模型在CIFAR-100数据集上使用OptDIA-的优化效果优于OptDIA外,其他模型使用OptDIA的优化效果均优于比OptDIA-,准确率提升了0.39~2.15个百分点,F1-score提升了0.11~2.03个百分点. 其中,对于使用CIFAR-10训练的ResNet-18,OptDIA可以提升0.53个百分点的准确率和0.41个百分点的F1-score,而OptDIA-只能提升0.20个百分点的准确率和0.17个百分点的F1-score,OptDIA所提升的准确率和F1-score分别为OptDIA-的2.65倍和2.41倍;对于使用CIFAR-10训练的ResNet-34,OptDIA可以提升0.80个百分点的准确率和0.87个百分点的F1-score,而OptDIA-只能提升0.15个百分点的准确率和0.78个百分点的F1-score,OptDIA所提升的准确率和F1-score分别为OptDIA-的5.33倍和1.12倍.
就所有模型的平均优化效果而言,对使用CIFAR-10数据集训练的DNN,OptDIA和OptDIA-平均提升的准确率为0.85个百分点和0.54个百分点,提升的F1-score为0.90个百分点和0.85个百分点,OptDIA平均提升的准确率和F1-score分别是OptDIA-的1.57倍和1.06倍;对使用CIFAR-100数据集训练的DNN,OptDIA和OptDIA-平均提升的准确率分别为1.36个百分点和1.21个百分点,提升的F1-score分别为0.90个百分点和1.24个百分点,OptDIA平均提升的准确率是OptDIA-的1.12倍,平均提升的F1-score与OptDIA-相近;对使用Fashion-MNIST数据集训练的DNN,OptDIA和OptDIA-平均提升的准确率为0.77个百分点和0.69个百分点,提升的F1-score为1.24个百分点和0.82个百分点,OptDIA平均提升的准确率和F1-score分别是OptDIA-的1.12倍和1.47倍.
从表3可以看出,使用解释方法指导变换数据再训练DNN所提升的性能高于未使用解释方法指导变换数据再训练DNN所提升的性能. 对实验结果进行分析发现,这是因为解释方法可更好地帮助理解DNN模型的预测过程. 使用解释方法将DNN训练过程透明化,以掌握训练数据中部分区域特征的重要程度,并将此作为变换数据以增加其多样性的依据. 而未使用解释方法指导数据变换则有较强随机性,难以使DNN模型学习多样化的特征.
OptDIA使用解释方法分析DNN的预测行为,并根据分析结果对原始数据进行数据变换以生成新数据. 解释方法的引入将消耗更高的时间成本,过高的额外时间成本将影响OptDIA在现实场景中的实用性. 为研究OptDIA在有效提高DNN准确率的前提下额外消耗的时间成本,在计算表3中的实验数据时,我们还统计了对不同结构的DNN和数据集分别使用OptDIA-和OptDIA生成新数据所消耗的时间成本,如表4所示.
如表4所示,使用解释方法Grad-CAM选择擦除区域的OptDIA,与直接对训练数据进行随机擦除的OptDIA-相比,解释方法的引入平均额外消耗了1 min 34 s,即平均每生成一张新训练数据额外消耗约0.016 s(以擦除概率为0.1,每个数据集生成约6000条新数据计算). OptDIA只对被选择进行数据变换的数据进行解释分析,算法的时间复杂度为O(n),且分析和数据变换区域的选择均可由GPU执行,具有较高的执行效率. 与DNN动辄数小时的训练耗时相比,OptDIA额外消耗的时间成本并不会明显降低其实用性.
对问题2的回答:与随机选择数据变换区域的方法相比,OptDIA能更有效优化DNN以提升其性能,且消耗额外时间较少.
为验证OptDIA是否能在应用现有数据增强方法进行训练的DNN上进行优化,首先将Zhong等人[28]提出的数据增强方法,即随机擦除(random erasing,RE),应用于DNN训练,然后再对使用RE训练的DNN尝试使用OptDIA进行优化. 实验使用RE方法在3个数据集上分别训练了300轮的9个DNN模型,并在其基础上再使用OptDIA继续训练50轮.
实验结果如表5所示,其中RE表示不同结构的DNN使用随机擦除方法进行数据增强训练所得到DNN的准确率,OptDIA表示在使用RE方法训练的基础上再使用OptDIA继续训练所得到DNN的准确率. 实验结果表明,与直接训练相比,RE训练的模型准确率提升了0.51~3.24个百分点,OptDIA则进一步提升使用RE训练的6个模型的准确率和7个模型的F1-score. 其中,对使用CIFAR-10数据集训练的模型,OptDIA将RE训练的ResNet-50模型的准确率和F1-score分别进一步提升了0.32个百分点和0.97个百分点,对使用CIFAR-100数据集训练的模型,OptDIA将RE训练模型的ResNet-50模型的准确率和F1-score进一步提升了2.72个百分点和2.35个百分点,对使用Fashion-MNIST数据集训练的模型,其准确率和F1-score已较高,OptDIA将RE训练的ResNet-50模型的准确率和F1-score分别进一步提升了0.25个百分点和0.15个百分点.
从表5可以看出,对于经使用数据增强方法预训练的DNN模型,OptDIA仍可进一步提升其准确率,即达到优化DNN的效果. 对实验结果进行分析发现,因为利用DNN的解释方法指导进行数据变换能有效增加训练样本的多样性,使DNN学到更多的特征.
对问题3的回答:OptDIA可进一步提升使用数据增强方法训练的DNN模型的准确率和F1-score,提升幅度分别为0.11~2.72个百分点和0.13~2.35个百分点.
本文方法的内部有效性分析主要在于确实影响实验正确性的因素. 首先,实验中使用的DNN解释方法为Grad-CAM的开源项目pytorch-grad-cam[32],该开源项目也被其他相关工作使用[33];其次,我们参考开源项目Random-Erasing的源代码[34]实现了实验所用的数据增强方法;最后,实验构建和训练DNN模型的代码来自Torchvision[35]的开源实现,此外,还对代码进行了多次内部交叉检查,以尽量确保实现的正确性.
基于可解释性分析的深度神经网络优化方法的外部有效性主要在于实验中所用的DNN模型和数据集是否具有代表性以及OptDIA是否具有可扩展性. 首先,本文实验所使用的ResNet模型以及CIFAR-10,CIFAR-100,Fashion-MNIST数据集在优化DNN的工作中被广泛使用[14,18,28],具有一定代表性,但不能确保OptDIA在其他待优化DNN和数据集上的有效性. 其次,在实验中只使用了擦除方法对数据进行扩增,不能确保OptDIA能在裁剪等其他数据扩增方法上的有效性. 此外,虽然任何准确率的提升都是对模型有意义的优化[14],但由于DNN模型具有内在不确定性,因此无法保证准确率的提升幅度.
本文方法的构造有效性影响主要在于优化DNN的评价指标. 实验中使用测试集的准确率和F1-score来评价实验效果,这2项指标常被用于评价DNN性能[14]. 此外,为避免实验结果的随机性,我们使用相同参数进行多次实验求得平均值作为实验结果.
传统软件缺陷修复方法通过基于启发式搜索[9]、人工模板[36]、语义约束[37]和统计分析[38]等技术自动生成针对特定缺陷的程序补丁,并将补丁自动添加到软件以修复缺陷,或将补丁作为提示信息来帮助开发者继续优化软件[12]. 与传统软件不同,OptDIA主要关注基于DNN的智能软件,其内部具体权重参数通过训练获取且结构较为复杂,这使其输出结果缺乏传统软件的解释性,因此传统方法难以直接应用于优化DNN. 目前对DNN进行优化的研究主要从训练集和模型参数2个切入点进行,本节将对这2类方法分别进行介绍.
一类方法是通过扩增训练集优化DNN以提升其性能. Ma等人[14]提出了DNN调试方法MODE,该方法首先通过分析神经网络状态生成差分热图(differential heat map),并结合DNN模型找出对于分类影响较大的神经元,分别对在训练过程中存在欠拟合或过拟合的DNN进行调试,然后指导GAN生成具有较高真实性的训练数据,从而提升模型的性能以优化DNN. Gao等人[39]提出了一种基于模糊测试的数据扩充方法,该方法结合模糊测试和遗传算法扩增数据,将优化问题转化为搜索问题,以提高模型的健壮性,从而达到优化DNN的目的. Liu等人[40]提出了一个面向数据的变异框架Styx,该框架首先通过对原始数据进行轻微变异生成新的训练集,并将其用于重新训练DNN,以提高模型健壮性,从而达到优化的目的. 文献[14, 39-40]所述的方法可在一定程度上优化DNN,然而依然存在扩充数据集的过程缺乏针对性,需要引入新的DNN,增加数据多样性的能力有限且过程难以理解等问题.
另一类方法是通过修改模型参数优化DNN以提升其性能. Zhang等人[18]提出了一种权重调整方法Apricot,该方法通过在原始训练集的许多不同子集上进行训练得到简化深度学习模型(reduced deep learning model,rDLM),使用其权重来辅助修改原始模型中的权重,从而提升模型的准确率以优化DNN. 随后他们又提出一种超启发式方法Plum[41],该方法通过应用不同的修复策略生成一组候选DNN模型,然后根据候选修复策略的整体优化效果对修复策略进行评估和排序,并使用最佳优化策略生成优化后的DNN. 基于变异的DNN优化方法GenMuNN[42]量化神经元权重对DNN预测结果的影响,并根据量化结果对各层神经元权重进行排序,然后使用遗传算法对权重进行变异以生成变异体,直到出现准确率较高的DNN,即可达到优化的目的. Zhang等人[43]提出了一种DNN训练监控与自动修复方法AUTOTRAINER,该方法可检测梯度消失、梯度爆炸、损失振荡和收敛缓慢等常见训练问题,并自动采用如替换激活函数等内置修复方案提高DNN性能,以达到优化DNN的目的. Sun等人[44]提出了基于因果关系的神经网络修复方法CARE,该方法通过最小限度地调整给定DNN的权重参数,通过执行基于因果关系的缺陷定位以识别并优化“有罪(guilty)”神经元,通过构建一个满足目标属性(如公平性、无后门等)的神经网络来提高其性能. 然而,直接对DNN参数进行优化也会有一定风险,例如在调整权重时,直接调整参数难以控制修改幅度,甚至可能会影响模型准确地拟合训练数据.
文献[18, 41-44]所述的工作主要通过扩增数据集或修改模型参数的方式优化DNN,由于缺乏解释性导致其修复过程大多难以理解. 本文所提出的基于可解释性分析的深度神经网络优化方法OptDIA使用DNN解释方法获取训练数据中对预测结果影响较大的区域,并通过对其进行数据变换生成新的数据再训练DNN,以提升DNN性能,从而达到优化DNN的目的. 相较之前的工作,OptDIA引入了DNN解释方法,提高了数据扩增的针对性,使优化DNN的过程更容易被理解.
为优化DNN以提升其性能,本文提出了一种基于可解释性分析的DNN优化方法OptDIA. 使用解释方法获取原始数据中对DNN决策具有较高重要性的区域,并将其作为变换数据生成新数据的依据,再重新训练DNN以提升其性能,从而达到优化DNN的目的. 在使用3个数据集训练的9个DNN模型上进行的实验结果表明,OptDIA可以有效提升DNN性能以进行优化,并可进一步优化使用现有数据增强方法训练的DNN. 在下一步工作中,我们计划在更大规模数据集上引入更多数据增强方法验证OptDIA的优化效果,并计划尝试将本文方法引入语音、文本处理等领域,如利用解释方法获取词语对情感分析结果的重要性程度,将其作为变换数据的依据,使用变换后的数据重新训练模型,以提高其性能.
作者贡献声明:吴欢欢负责相关工作调研、论文撰写及修订等工作;谢瑞麟完成实验并修改论文;乔塬心协助完成实验;陈翔提出指导意见并修改论文;崔展齐指导论文选题、确定论文整体框架设计并修改论文.

参考文献 引证文献
排序方式:
1
Esteva A, Chou K, Yeung S, et al. Deep learning-enabled medical computer vision[J]. NPJ Digital Medicine, 2021, 4(1): 1−9
2
Malik M, Malik M K, Mehmood K, et al. Automatic speech recognition: A survey[J]. Multimedia Tools and Applications, 2021, 80(6): 9411−9457
3
郑海斌,陈晋音,章燕,等. 面向自然语言处理的对抗攻防与鲁棒性分析综述[J]. 计算机研究与发展,2021,58(8):1727−1750
Zheng Haibin, Chen Jinyin, Zhang Yan, et al. Survey of adversarial attack, defense and robustness analysis for natural language processing[J]. Journal of Computer Research and Development, 2021, 58(8): 1727−1750 (in Chinese)
4
Fang Peng, Wang Zecong, Zhang Xiangyun. Vehicle automatic driving system based on embedded and machine learning[C]//Proc of Int Conf on Computer Vision, Image and Deep Learning. Piscataway, NJ: IEEE, 2020: 281−284
5
Bakator M, Radosav D. Deep learning and medical diagnosis: A review of literature[J]. Multimodal Technologies and Interaction, 2018, 2(3): 47
6
Ghenescu V, Mihaescu R E, Carata S V, et al. Face detection and recognition based on general purpose DNN object detector[C]//Proc of Int Symp on Electronics and Telecommunications. Piscataway, NJ: IEEE, 2018: 1−4
7
Tulbure A A, Tulbure A A, Dulf E H. A review on modern defect detection models using DCNNs-Deep convolutional neural networks[J]. Journal of Advanced Research, 2022, 35: 33−48
8
界面新闻. 新能源汽车评论[EB/OL]. 2022 [2023-02-16].https://www.jiemian.com/article/7530592.html
Interface News. Review of new energy vehicles[EB/OL]. 2022 [2023-02-16].https://www.jiemian.com/article/7530592.html (in Chinese)
9
Goues C L, Nguyen T V, Forrest S, et al. GenProg: A generic method for automatic software repair[J]. IEEE Transactions on Software Engineering, 2011, 38(1): 54−72
10
Baudry B, Fleurey F, Traon Y L. Improving test suites for efficient fault localization[C]//Proc of the 28th Int Conf on Software Engineering. New York: ACM, 2006: 82−91
11
Zeller A. Yesterday, my program worked. Today, it does not. Why?[J]. ACM SIGSOFT Software Engineering Notes, 1999, 24(6): 253−267
12
Zhang Xiangyu, Gupta N, Gupta R. Locating faults through automated predicate switching[C]//Proc of the 28th Int Conf on Software Engineering. New York: ACM, 2006: 272−281
13
Ma Lei, Xu Felix Juefei, Zhang Fuyuan, et al. Deepgauge: Multi-granularity testing criteria for deep learning systems[C]//Proc of the 33rd ACM/IEEE Int Conf on Automated Software Engineering. New York: ACM, 2018: 120−131
14
Ma Shiqing, Liu Yingqi, Lee W C, et al. MODE: Automated neural network model debugging via state differential analysis and input selection[C]//Proc of the 26th ACM Joint Meeting on European Software Engineering Conf and Symp on the Foundations of Software Engineering. New York: ACM, 2018: 175−186
15
王赞,闫明,刘爽,等. 深度神经网络测试研究综述[J]. 软件学报,2020,31(5):1255−1275
Wang Zan, Yan Ming, Liu Shuang, et al. A survey on testing of deep neural network. [J] Journal of Software, 2020, 31(5): 1255−1275 (in Chinese)
16
Pei Kexin, Cao Yinzhi, Yang Junfen, et al. Deepxplore: Automated whitebox testing of deep learning systems[C]//Proc of the 26th Symp on Operating Systems Principles. New York: ACM, 2017: 1−18
17
Zhang Mengshi, Zhang Yuqun, Zhang Lingming, et al. DeepRoad: GAN-based metamorphic testing and input validation framework for autonomous driving systems[C]//Proc of the 33rd IEEE/ACM Int Conf on Automated Software Engineering. Piscataway, NJ: IEEE, 2018: 132−142
18
Zhang Hao, Chan W K. Apricot: A weight-adaptation approach to fixing deep learning models[C]//Proc of the 34th IEEE/ACM Int Conf on Automated Software Engineering. Piscataway, NJ: IEEE, 2019: 376−387
19
Selvaraju R R, Cogswell M, Das A, et al. Grad-CAM: Visual explanations from deep networks via gradient-based localization[C]//Proc of the IEEE Int Conf on Computer Vision. Piscataway, NJ: IEEE, 2017: 618−626
20
Simonyan K, Vedaldi A, Zisserman A. Deep inside convolutional networks: Visualising image classification models and saliency maps[J]. arXiv preprint, arXiv: 1312. 6034, 2013
21
Du Mengnan, Liu Ninghao, Song Qinghao, et al. Towards explanation of DNN-based prediction with guided feature inversion[C]//Proc of the 24th ACM SIGKDD Int Conf on Knowledge Discovery & Data Mining. New York: ACM, 2018: 1358−1367
22
Zhou Bolei, Khosla A, Lapedriza A, et al. Learning deep features for discriminative localization[C]//Proc of the IEEE Conf on Computer Vision and Pattern Recognition. Piscataway, NJ: IEEE, 2016: 2921−2929
23
CIFAR official website. Dataset[EB/OL]. [2023-02-16]. http://www.cs.toronto.edu/~kriz/cifar.html
24
GitHub. Fashion-MNIST[EB/OL]. 2017[2023-02-16].https://github.com/zalandoresearch/fashion-mnist
25
Mikołajczyk A, Grochowski M. Data augmentation for improving deep learning in image classification problem[C]//Proc of 2018 Int Interdisciplinary PhD Workshop (IIPhDW). Piscataway, NJ: IEEE, 2018: 117−122
26
Taylor L, Nitschke G. Improving deep learning with generic data augmentation[C]// Proc of IEEE Symp Series on Computational Intelligence. Piscataway, NJ: IEEE, 2018: 1542−1547
27
Kang Guoliang, Dong Xuanyi, Zheng Liang, et al. Patchshuffle regularization[J]. arXiv preprint, arXiv: 1707. 07103, 2017
28
Zhong Zhun, Zheng Liang, Kang Guoliang, et al. Random erasing data augmentation[C]//Proc of the AAAI Conf on Artificial Intelligence. Palo Alto, CA: AAAI, 2020: 13001−13008
29
Zajac M, Zołna K, Rostamzadeh N, et al. Adversarial framing for image and video classification[C]//Proc of the AAAI Conf on Artificial Intelligence. Palo Alto, CA: AAAI, 2019: 10077−10078
30
Johnson J, Alahi A, Li Feifei. Perceptual losses for real-time style transfer and super-resolution[C]//Proc of the 14th European Conf on Computer Vision. Berlin: Springer, 2016: 694−711
31
He Kaiming, Zhang Xiangyu, Ren Shaoqing, et al. Deep residual learning for image recognition[C]//Proc of the IEEE Conf on Computer Vision and Pattern Recognition. Piscataway, NJ: IEEE, 2016: 770−778
32
GitHub. Grad-CAM[EB/OL]. 2022[2023-02-16].https://github.com/jacobgil/pytorch-grad-cam
33
Chen Lei, Chen Jianhui, Hajimirsadeghi H, et al. Adapting Grad-CAM for embedding networks[C]//Proc of the IEEE/CVF Winter Conf on Applications of Computer Vision. Piscataway, NJ: IEEE, 2020: 2794−2803
34
GitHub. Random-Erasing[EB/OL]. 2019[2023-02-16].https://github.com/zhunzhong07/Random-Erasing
35
GitHub. Torchvision[EB/OL]. 2022[2023-02-16].https://github.com/pytorch/vision
36
Tian Yuchi, Ray B. Automatically diagnosing and repairing error handling bugs in C[C]//Proc of the 11th Joint Meeting on Foundations of Software Engineering. New York: ACM, 2017: 752−762
37
Nguyen H D T, Qi Dawei, Roychoudhury A, et al. Semfix: Program repair via semantic analysis[C]//Proc of the 35th Int Conf on Software Engineering. Piscataway, NJ: IEEE, 2013: 772−781
38
Jeffrey D, Feng Min, Gupta N, et al. BugFix: A learning-based tool to assist developers in fixing bugs[C]//Proc of the 17th IEEE Int Conf on Program Comprehension. Piscataway, NJ: IEEE, 2009: 70−79
39
Gao Xiang, Saha R K, Prasad M R, et al. Fuzz testing based data augmentation to improve robustness of deep neural networks[C]//Proc of the 42nd IEEE/ACM Int Conf on Software Engineering. Piscataway, NJ: IEEE, 2020: 1147−1158
40
Liu Meixi, Hong Weijiang, Pan Weiyu, et al. Styx: A data-oriented mutation framework to improve the robustness of DNN[C]//Proc of the 35th IEEE/ACM Int Conf on Automated Software Engineering. Piscataway, NJ: IEEE, 2020: 1260−1261
41
Zhang Hao, Chan W K. Plum: Exploration and prioritization of model repair strategies for fixing deep learning models[C]//Proc of the 8th Int Conf on Dependable Systems and Their Applications. Piscataway, NJ: IEEE, 2021: 140−151
42
Wu Huanhuan, Li Zheng, Cui Zhanqi, et al. GenMuNN: A mutation-based approach to repair deep neural network models[J]. International Journal of Modeling, Simulation, and Scientific Computing, 2022, 13(2): 2341008
43
Zhang Xiaoyu, Zhai Juan, Ma Shiqing, et al. AUTOTRAINER: An automatic DNN training problem detection and repair system[C]//Proc of the 43rd IEEE/ACM Int Conf on Software Engineering. Piscataway, NJ: IEEE, 2021: 359−371
44
Sun Bing, Sun Jun, Pham L H, et al. Causality-based neural network repair[C]//Proc of the 44th Int Conf on Software Engineering. New York: ACM, 2022: 338−349
2024年第61卷第1期
68
32
0
引用本文
BibTeX
文章信息
doi: 10.7544/issn1000-1239.202220803
  • 接收时间:2022-09-14
  • 首发时间:2026-11-17
  • 出版时间:2024-01-02
补充材料
相关文章
文章信息
作者
出版历史
  • 收稿日期:2022-09-14
  • 修回日期:2023-02-22
基金
作者信息
    1北京信息科技大学计算机学院 北京 100101
    2南通大学信息科学技术学院 江苏南通 226019

通讯作者:

崔展齐()
参考文献
分享链接
https://castjournals.cast.org.cn/joweb/jsjyjyfz/CN/10.7544/issn1000-1239.202220803
分享至
全文二维码

扫描看全文

引用本文
BibTeX
本文的引用情况
2种不同金属材料的力学参数
科
Family
属数
Number of
genus
种数
Number of
species
占总种数比例
Percentage of
total species (%)
属
Genus
种数
Number of
species
占总种数比例
Percentage of total
species (%)
鹅膏菌科Amanitaceae 2 11 5.26 鹅膏菌属 Amanita 10 4.78
小菇科 Mycenaceae 2 12 5.74 丝盖伞属 Inocybe 5 2.39
多孔菌科 Polyporaceae 8 14 6.70 蜡蘑属 Laccaria 5 2.39
红菇科 Russulaceae 3 23 11.00 小皮伞属 Marasmius 6 2.87
小菇属 Mycena 11 5.26
光柄菇属 Pluteus 5 2.39
红菇属 Russula 17 8.13
栓菌属 Trametes 5 2.39
关闭全屏