[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-137721-105":3,"detail-sidebar-cat-0-en-105":81,"doc-detail-137721-en":130},{"code":4,"msg":5,"data":6},0,"ok",{"site_id":7,"language":8,"slug":9,"title":10,"keywords":11,"description":12,"schema_data":13,"social_meta":74,"head_meta":76,"extra_data":78,"updated_unix":80},105,"en","on-the-potential-and-limits-of-zero-shot-out-of-distribution-detection","On the Potential and Limits of Zero-Shot Out-of-Distribution Detection","","The thesis addresses uncertainty in AI systems deployed beyond the closed-world settings in which they are designed and tested, focusing on Out-of-Distribution Detection (OOD) and its zero-shot variant, Z-OOD detection. It studies image classification using multi-modal architectures such as CLIP, evaluating two published approaches (MCM and ZOC) across 12 datasets with strong semantic shifts and additional stress conditions like image corruption. The work links method robustness to difficulty metrics, proposes UC-CLP for split comparability, and develops T-MCM and T-ZOC for domain-adapted few-shot setups.",{"@graph":14,"@context":73},[15,34,56],{"@type":16,"itemListElement":17},"BreadcrumbList",[18,23,27,31],{"item":19,"name":20,"@type":21,"position":22},"https://docshare.wps.com","Home","ListItem",1,{"item":24,"name":25,"@type":21,"position":26},"https://docshare.wps.com/document/","Document",2,{"item":28,"name":29,"@type":21,"position":30},"https://docshare.wps.com/document/research-report/","Research & Report",3,{"item":32,"name":10,"@type":21,"position":33},"https://docshare.wps.com/document/on-the-potential-and-limits-of-zero-shot-out-of-distribution-detection/137721/",4,{"url":32,"name":10,"@type":35,"image":36,"author":41,"headline":10,"publisher":44,"fileFormat":47,"inLanguage":8,"description":12,"dateModified":48,"datePublished":49,"encodingFormat":47,"isAccessibleForFree":50,"interactionStatistic":51},"DigitalDocument",{"url":37,"@type":38,"width":39,"height":40},"https://docshare.wps.com/thumbnails/on-the-potential-and-limits-of-zero-shot-out-of-distribution-detection/137721.png","ImageObject",300,407,{"name":42,"@type":43},"Đào","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-09-19","2026-08-22",true,{"@type":52,"interactionType":53,"userInteractionCount":55},"InteractionCounter",{"@type":54},"ViewAction",6,{"@type":57,"mainEntity":58},"FAQPage",[59,65,69],{"name":60,"@type":61,"acceptedAnswer":62},"What problem does Z-OOD detection target in AI systems?","Question",{"text":63,"@type":64},"Z-OOD detection targets the uncertainty and risk that arise when AI models encounter inputs from distributions not covered during training and testing in closed-world environments.","Answer",{"name":66,"@type":61,"acceptedAnswer":67},"How is Z-OOD detection evaluated in this thesis?",{"text":68,"@type":64},"The thesis runs large-scale benchmarks on 12 datasets with strong semantic shifts, tests robustness under conditions like image corruption, and compares two published methods (MCM and ZOC) with an additional smaller semantic shift setting.",{"name":70,"@type":61,"acceptedAnswer":71},"What challenges are observed for Z-OOD detection?",{"text":72,"@type":64},"Z-OOD detection is generally effective, especially in far-OOD scenarios, but it becomes challenging in near-OOD cases where the underlying CLIP model faces difficulties in classification.","https://schema.org",{"og:url":32,"og:type":75,"og:title":10,"og:site_name":45,"og:description":12},"article",{"robots":77,"canonical":32},"index,follow",{"doc_id":79,"site_id":7},137721,1787438954,{"code":4,"msg":82,"data":83},"success",[84,88,92,96,101,105,110,114,119,122,126],{"id":22,"doc_module":4,"doc_module_name":25,"category_name":85,"show_sort_weight":86,"slug":87},"Story & Novel",90,"story-novel",{"id":26,"doc_module":4,"doc_module_name":25,"category_name":89,"show_sort_weight":90,"slug":91},"Literature",80,"literature",{"id":33,"doc_module":4,"doc_module_name":25,"category_name":93,"show_sort_weight":94,"slug":95},"Exam",70,"exam",{"id":97,"doc_module":4,"doc_module_name":25,"category_name":98,"show_sort_weight":99,"slug":100},5,"Comic",60,"comic",{"id":55,"doc_module":4,"doc_module_name":25,"category_name":102,"show_sort_weight":103,"slug":104},"Technology",50,"technology",{"id":106,"doc_module":4,"doc_module_name":25,"category_name":107,"show_sort_weight":108,"slug":109},7,"Healthcare",40,"healthcare",{"id":111,"doc_module":4,"doc_module_name":25,"category_name":29,"show_sort_weight":112,"slug":113},8,30,"research-report",{"id":115,"doc_module":4,"doc_module_name":25,"category_name":116,"show_sort_weight":117,"slug":118},9,"Religion & Spirituality",20,"religion-spirituality",{"id":117,"doc_module":4,"doc_module_name":25,"category_name":120,"show_sort_weight":117,"slug":121},"World Cup","world-cup",{"id":123,"doc_module":4,"doc_module_name":25,"category_name":124,"show_sort_weight":123,"slug":125},10,"Lifestyle","lifestyle",{"id":127,"doc_module":4,"doc_module_name":25,"category_name":128,"show_sort_weight":97,"slug":129},19,"General","general",{"code":4,"msg":82,"data":131},{"doc_id":79,"user_id":132,"nickname":42,"user_avatar":133,"doc_module":4,"category_id":111,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":55,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":139,"language":140,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":141,"faqs":142,"seo_title":143,"seo_description":12,"update_tm":80,"read_time":144},1374402968488,"https://ap-avatar.wpscdn.com/davatar_29158cc5080c5b710cf443261637dec0","M A S T E R T H E S I S  \nOn the Potential and Limits of Zero-Shot Out-of-Distribution Detection  \nsubmitted by  \nFabian Meyer  \nMIN-Faculty  \nDepartment of Informatics  \nCourse of studies: Master Informatics  \nMatrikelnummer: 6816480  \nSubmission date: 16.04.2023  \nFirst examiner: Prof. Dr. Chris Biemann  \nSecond examiner: Dr. Florian Wilhelm  \nSupervisors: Sven Müller, Florian Schneider, Xintong Wang  \nFabian Meyer:  \nOn the Potential and Limits of  \nZero-Shot Out-of-Distribution Detection Master Thesis, Informatics Universität Hamburg  \nAbstract  \nThe growing prevalence of artificial intelligent (AI) systems in nearly all aspects of everyday life has also led to their integration into critical domains, e.g. in nuclear powerplants, autonomous vehicles, and the detection of fatal diseases. Given that these systems are initially designed and tested within controlled, closed-world environments, they may face unanticipated inputs when deployed in real-world scenarios, leading to uncertainty in their interpretation and response. To mitigate the risk of incorrect decision-making, Out-of-Distribution Detection (OOD detection) techniques ensure that AI systems make decisions only for data that originates from familiar distributions. Zero-Shot Out-ofDistribution Detection (Z-OOD detection) is a special case recently introduced, which builds on the zero-shot classification paradigm.  \nIn this thesis, we explore the potential and limitations of Z-OOD detection for image classification by leveraging the capabilities of recent multi-modal architectures, such as the Clip model. To test the generalizability of the approach, we conduct large-scale benchmarks on 12 datasets with strong semantic shifts in the data using the two published Z-OOD detection methods, Maximum Concept Matching (MCM) and Zero-Shot Out-of-Distribution Detection based on Clip (ZOC), followed by a challenging comparison with a smaller semantic shift. The robustness of the methods is tested under different conditions, such as image corruption, and attempts are made to determine the lower bound of task difficulty of the methods. We investigate correlations with difficulty metrics from OOD detection and assess their predictive power.  \nThe thesis also aims to understand whether advancements in domain adaptation methods can be transferred to OOD detection. To accomplish this, we test the methodology in a few-shot setup and compare it against benchmark results. Our findings indicate that Z-OOD detection is generally effective, especially in far-OOD scenarios. However, challenges arise in near-OOD cases, where the underlying Clip model faces difficulties in classification. We propose the Universal Clip-based Confusion Log Probability (UC-CLP) as a universal indicator of the difficulty of selected In-Distribution/OOD splits, improving comparability within the field.  \nFinally, we propose T-MCM and T-ZOC as domain-adapted few-shot OOD detection methodologies, with T-MCM demonstrating a lightweight, fast-adapting approach. The performance of these methods depends on the success of domain adaptation, showing potential for improvement.  \nAcknowledgements  \nI certainly would not have been able to complete this research project on my own, so I would like to take this opportunity to thank people who helped me to achieve this. First, I would like to thank Dr. Florian Wilhelm for the opportunity to write my thesis withinovex GmbH. Then, I would like to thank Prof. Dr. Chris Biemann for allowing me to write the master thesis in his research group.  \nSpecial thanks go to Florian Schneider, Sven Müller, Lars Engellandt, and Xintong Wang, who accompanied me throughout the entire process from finding the topic to the final revision. They provided me with fruitful discussions, advice, had an open ear for any amount of confused thoughts, and offered any support I needed. It was a pleasure to work with you. I would also like to thank all friends and my colleagues in Hamburg, where I spent most","cbCaityiDLmCtT15","https://ap.wps.com/l/cbCaityiDLmCtT15","pdf",1893908,111,"English","# Introduction\n## Motivation\n## Approach\n## Research Questions\n## Structure of this Work\n# Background\n## Learning Theory\n### Supervised Learning\n### Binary Classification\n## Deep Learning Architectures\n### Classification\n### Natural Language Processing\n### Computer Vision\n### CLIP\n### Adapters\n## Out-of-Distribution Detection\n### Differentiation from Related Topics\n### Zero-Shot Out-of-Distribution Detection\n# Related Work\n## Out-of-Distribution Detection\n### Evaluation Protocols\n### Out-of-Distribution Detection without Outlier Exposure\n### Few-Shot Out-of-Distribution Detection\n### Zero-Shot Out-of-Distribution Detection\n### OOD Detection Difficulty\n## Zero-shot Transfer\n### Vision Language Models\n### Task and Domain Adaption","[{\"question\":\"What problem does Z-OOD detection target in AI systems?\",\"answer\":\"Z-OOD detection targets the uncertainty and risk that arise when AI models encounter inputs from distributions not covered during training and testing in closed-world environments.\"},{\"question\":\"How is Z-OOD detection evaluated in this thesis?\",\"answer\":\"The thesis runs large-scale benchmarks on 12 datasets with strong semantic shifts, tests robustness under conditions like image corruption, and compares two published methods (MCM and ZOC) with an additional smaller semantic shift setting.\"},{\"question\":\"What challenges are observed for Z-OOD detection?\",\"answer\":\"Z-OOD detection is generally effective, especially in far-OOD scenarios, but it becomes challenging in near-OOD cases where the underlying CLIP model faces difficulties in classification.\"}]","On the Potential and Limits of Zero-Shot Out-of-Distribution Detection | PDF",280]