[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-124863-en":3,"doc-seo-124863-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":27,"seo_description":14,"update_tm":28,"read_time":29},124863,549758252649,"Ivy","https://ap-avatar.wpscdn.com/avatar/8000253669c5317157?_k=1778319167496531819",8,"Research & Report","Enhancing And Evaluating Interpretability In Machine Learning Through Theory And Practice","The field of Explainable AI (XAI) seeks to explain decisions made by machine learning models, yet many approaches still lack rigorous theoretical foundations. This dissertation investigates explainability methods both theoretically and empirically, starting with a history of XAI. It proposes a theoretically motivated attribution method using an absolute frame of reference in bits per pixel and evaluates it against 11 baselines. Further work analyzes why modified backward-propagation attributions often fail sanity checks, examines limitations of Deep Taylor Decomposition, studies user benefits with a structured experiment, and presents an explainable k-nearest neighbors regression with higher-order information.","Enhancing And Evaluating Interpretability In Machine Learning Through Theory And Practice  \nDissertation  \nzur Erlangung des Grades eines  \nDoktors der Naturwissenschaften  \n(Dr. rer. nat.)  \nam Fachbereich Informatik und Mathematik  \nder Freie Universität Berlin  \nvorgelegt von  \nLeon Sixt  \nBerlin, 2023  \nErstgutachter: Prof. Dr. Tim Landgraf  \nZweitgutachter: Prof. Dr. Oisin Mac Aodha  \nTag der Disputation: 16.02.2024  \nEnhancing And Evaluating Interpretability In Machine Learning Through The ory And Practice, Leon Sixt  \n© 2023 Berlin  \nContents  \nAbstract viii  \nPublication Of The Thesis ix  \nDeclaration Of Authorship xi  \n1 Introduction 1  \n1.1 Why Should We Explain Deep Neural Networks? ....... 1  \n1.2 Challenges of Explainable AI .................. 3  \n1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 4  \n1.4 Structure of the Thesis . . . . . . . . . . . . . . . . . . . . . 6  \n1.5 Contributions To Individual Publications . . . . . . . . . . . . 7  \n2 History Of The Field 9  \n3 Restricting the Flow: Information Bottlenecks for Attribution 13  \n4 When Explanations Lie: Why Many Modified BP Attributions Fail 33  \n5 A Rigorous Study Of The Deep Taylor Decomposition 55  \n6 Do Users Benefit From Interpretable Vision? A User Study,  \nBaseline, And Dataset 79  \n7 DNNR: Differential Nearest Neighbor Regression 105  \n8 Conclusion 129  \n8. 1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129  \n8.2 Future Prospects ......................... 130  \nBibliography 131  \nThis solution of the problem of induction gives rise to a new theory of the method of science, to an analysis of the critical method, the method of trial and error: the method of proposing bold hypotheses, and exposing them to the severest criticism, in order to detect where we have erred.  \n—Karl Popper  \nAcknowledgements  \nFirst, I want to thank my supervisor, Prof. Dr. Tim Landgraf, for his encouragement and dedicated support throughout my Ph.D. studies.  \nI had the pleasure of collaborating with great people during my studies. In particular, I would like to thank Karl Schulz, Maximilian Granz, Martin Schüßler, Oana-Iuliana Popescu, and Youssef Nader.  \nDuring my Ph.D., I had the opportunity to apply my research during two internships at Google Research and Angsa Robotics. I thank my supervisors Dr. Been Kim, Dr. Martin Maas, and Lukas Wießmeier.  \nI would like to thank the Elsa Neumann Scholarship issued by the state of Berlin for their financial support and the Freie Universität Berlin for providing me with free access to their HPC cluster.  \nFinally, I am grateful to my partner, family, and friends for their support and encouragement.  \nAbstract  \nThe field of Explainable AI (XAI) aims to explain the decisions made by machine learning models. Recently, there have been calls for more rigorous and theoretically grounded approaches to explainability. In my thesis, I respond to this call by investigating the properties of explainability methods theoretically and empirically. As an introduction, I provide a brief overview of the history of XAI. The first contribution is a novel theoretically motivated attribution method that estimates the importance of each input feature in bits per pixel, an absolute frame of reference. The method is evaluated against 11 baselinesin several benchmarks. In my next publication, the limitations of modified backward propagation methods are examined. It is found that many methods fail the weight-randomization sanity check, and the reasons for these failures are analyzed in detail. In a follow-up publication, the limitations of one particular method, Deep Taylor Decomposition (DTD), are further analyzed. DTD has been cited as the theoretical basis for many other attribution methods. However, it is found to be either under-constrained or reduced to the simpler gradient 􀀂 input method. In the next contribution, a user study design is presented to evaluate the helpfulness of XAI to users in practice. In","cbCaivU2btM9cRxm","https://ap.wps.com/l/cbCaivU2btM9cRxm","pdf",15230706,1,153,"English","en",105,"# Introduction\n## Why Should We Explain Deep Neural Networks?\n## Challenges of Explainable AI\n## Contributions\n## Structure of the Thesis\n## Contributions To Individual Publications\n# History Of The Field\n# Restricting the Flow: Information Bottlenecks for Attribution\n# When Explanations Lie: Why Many Modified BP Attributions Fail\n# A Rigorous Study Of The Deep Taylor Decomposition\n# Do Users Benefit From Interpretable Vision?\n## A User Study, Baseline, And Dataset\n# DNNR: Differential Nearest Neighbor Regression\n# Conclusion\n## Discussion\n## Future Prospects\n# Bibliography","[{\"question\":\"What is the main focus of this dissertation on Explainable AI?\",\"answer\":\"It investigates properties of explainability methods using both theoretical and empirical analyses, aiming to improve interpretability and reliability.\"},{\"question\":\"How is the proposed attribution method evaluated?\",\"answer\":\"The method estimates feature importance in bits per pixel and is evaluated against 11 baselines across several benchmarks.\"},{\"question\":\"What are the findings about modified backward-propagation attribution methods and Deep Taylor Decomposition?\",\"answer\":\"Many modified BP attributions fail the weight-randomization sanity check, and Deep Taylor Decomposition is found to be under-constrained or reduced to the simpler gradient×input method.\"}]","Enhancing And Evaluating Interpretability In Machine Learning Through Theory And Practice | PDF",1785895095,386,{"code":4,"msg":31,"data":32},"ok",{"site_id":24,"language":23,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"enhancing-and-evaluating-interpretability-in-machine-learning-through-theory-and-practice","",{"@graph":36,"@context":85},[37,54,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":50},"https://docshare.wps.com/document/research-report/",3,{"item":52,"name":13,"@type":43,"position":53},"https://docshare.wps.com/document/enhancing-and-evaluating-interpretability-in-machine-learning-through-theory-and-practice/124863/",4,{"url":52,"name":13,"@type":55,"author":56,"headline":13,"publisher":58,"fileFormat":61,"inLanguage":23,"description":14,"dateModified":62,"datePublished":62,"encodingFormat":61,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":57},"Person",{"url":41,"name":59,"@type":60},"DocShare","Organization","application/pdf","2026-08-05",true,{"@type":65,"interactionType":66,"userInteractionCount":4},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What is the main focus of this dissertation on Explainable AI?","Question",{"text":75,"@type":76},"It investigates properties of explainability methods using both theoretical and empirical analyses, aiming to improve interpretability and reliability.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How is the proposed attribution method evaluated?",{"text":80,"@type":76},"The method estimates feature importance in bits per pixel and is evaluated against 11 baselines across several benchmarks.",{"name":82,"@type":73,"acceptedAnswer":83},"What are the findings about modified backward-propagation attribution methods and Deep Taylor Decomposition?",{"text":84,"@type":76},"Many modified BP attributions fail the weight-randomization sanity check, and Deep Taylor Decomposition is found to be under-constrained or reduced to the simpler gradient×input method.","https://schema.org",{"og:url":52,"og:type":87,"og:title":13,"og:site_name":59,"og:description":14},"article",{"robots":89,"canonical":52},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":53,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":46,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":46,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]