{"id":{"repo_id":"rowan","oai_identifier":"oai:rdw.rowan.edu:etd-4551"},"canonical_url":"https://search.ndltd.org/etd/rowan/oai:rdw.rowan.edu:etd-4551","repository":{"repo_id":"rowan","name":"Rowan University","base_url":"https://rdw.rowan.edu/do/oai/"},"display":{"title":"A Holistic Framework for Enhancing Machine Learning Reliability: From Model Architecture and Algorithm Design to Human Interpretability","abstract":"Modern machine learning systems can perform well on benchmarks yet fail in ways that are difficult for users to anticipate, interpret, or manage. This dissertation treats reliability as more than accuracy: a reliable system performs when it can, signals when it should not be trusted, and supports abstention, fallback, escalation, or human review. Unlike robustness, which emphasizes performance under shift, corruption, adversarial perturbation, or drift, reliability also requires recognizing limited competence, preserving useful semantic structure when fine-grained certainty is unjustified, and exposing mechanisms that make failures understandable. The dissertation develops this view through four contributions. Uncertainty-aware decision heads provide portable trust signals for classifiers and detectors. A modular perception-to-risk pipeline combines detection, tracking, intention estimation, and interpretable risk fusion for aerial threat prioritization. Hierarchy-aware classification enables semantic fallback by returning meaningful parent concepts when an exact leaf class is unavailable, unsupported, or withheld. Training-time reliability control is framed as learning to reallocate effort toward difficult or under-served groups. Together, these studies argue that trust improves when systems can state when they should not be trusted, explain why decisions are made, and adapt learning where standard optimization is weakest, yielding a unified view of reliability across model architecture, algorithm design, human interpretability, and decision support.","abstract_html":"Modern machine learning systems can perform well on benchmarks yet fail in ways that are difficult for users to anticipate, interpret, or manage. This dissertation treats reliability as more than accuracy: a reliable system performs when it can, signals when it should not be trusted, and supports abstention, fallback, escalation, or human review. Unlike robustness, which emphasizes performance under shift, corruption, adversarial perturbation, or drift, reliability also requires recognizing limited competence, preserving useful semantic structure when fine-grained certainty is unjustified, and exposing mechanisms that make failures understandable. The dissertation develops this view through four contributions. Uncertainty-aware decision heads provide portable trust signals for classifiers and detectors. A modular perception-to-risk pipeline combines detection, tracking, intention estimation, and interpretable risk fusion for aerial threat prioritization. Hierarchy-aware classification enables semantic fallback by returning meaningful parent concepts when an exact leaf class is unavailable, unsupported, or withheld. Training-time reliability control is framed as learning to reallocate effort toward difficult or under-served groups. Together, these studies argue that trust improves when systems can state when they should not be trusted, explain why decisions are made, and adapt learning where standard optimization is weakest, yielding a unified view of reliability across model architecture, algorithm design, human interpretability, and decision support.","abstract_has_math":false,"creators":["Naddeo, Kyle"],"institution":null,"degree_name":"Ph.D. Electrical and Computer Engineering","degree_level":"Dissertation","degree_discipline":"Electrical and Computer Engineering","degree_department":null,"school":null,"contributors":["Nidhal Bouaynaya, Ph.D.","Umashanger Thayasivam, Ph.D.","Ying Tang, Ph.D."],"advisors":[],"committee_chairs":[],"committee_members":[],"year":null,"date_issued":"","date_published":null,"updated_at":"2026-11-19T09:00:05Z","subjects":[],"languages":[],"rights":[],"rights_urls":[],"identifier_entries":[],"person_identifier_entries":[]},"links":{"outbound_url":"https://rdw.rowan.edu/etd/3545","outbound_label":"Repository record","outbound_source":"dc:identifier","record_url":"https://rdw.rowan.edu/etd/3545","record_label":"Repository record","record_kind":"record","record_source":"dc:identifier","download_url":null,"download_label":null,"download_source":null,"access_status":null,"embargo_until":null,"access_note":null},"source_record":{"url":"https://rdw.rowan.edu/do/oai/?verb=GetRecord&metadataPrefix=oai_etdms&identifier=oai%3Ardw.rowan.edu%3Aetd-4551","prefix":"oai_etdms"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["Nidhal Bouaynaya, Ph.D.","Umashanger Thayasivam, Ph.D.","Ying Tang, Ph.D."]},{"key":"dc:creator","label":"Author","values":["Naddeo, Kyle"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.available","label":"Dc Date Available","values":["2026-06-23T07:00:00Z"]},{"key":"dc:date.created","label":"Dc Date Created","values":["2026-06-23T07:00:00Z"]},{"key":"thesis:degree_discipline","label":"Discipline","values":["Electrical and Computer Engineering"]},{"key":"thesis:degree_level","label":"Degree Level","values":["Dissertation"]},{"key":"thesis:degree_name","label":"Degree Name","values":["Ph.D. Electrical and Computer Engineering"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["https://rdw.rowan.edu/etd/3545"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["Modern machine learning systems can perform well on benchmarks yet fail in ways that are difficult for users to anticipate, interpret, or manage. This dissertation treats reliability as more than accuracy: a reliable system performs when it can, signals when it should not be trusted, and supports abstention, fallback, escalation, or human review. Unlike robustness, which emphasizes performance under shift, corruption, adversarial perturbation, or drift, reliability also requires recognizing limited competence, preserving useful semantic structure when fine-grained certainty is unjustified, and exposing mechanisms that make failures understandable. The dissertation develops this view through four contributions. Uncertainty-aware decision heads provide portable trust signals for classifiers and detectors. A modular perception-to-risk pipeline combines detection, tracking, intention estimation, and interpretable risk fusion for aerial threat prioritization. Hierarchy-aware classification enables semantic fallback by returning meaningful parent concepts when an exact leaf class is unavailable, unsupported, or withheld. Training-time reliability control is framed as learning to reallocate effort toward difficult or under-served groups. Together, these studies argue that trust improves when systems can state when they should not be trusted, explain why decisions are made, and adapt learning where standard optimization is weakest, yielding a unified view of reliability across model architecture, algorithm design, human interpretability, and decision support."]},{"key":"dc:title","label":"Title","values":["A Holistic Framework for Enhancing Machine Learning Reliability: From Model Architecture and Algorithm Design to Human Interpretability"]}]}],"canonical_facts":{"dc:contributor":["Nidhal Bouaynaya, Ph.D.","Umashanger Thayasivam, Ph.D.","Ying Tang, Ph.D."],"dc:creator":["Naddeo, Kyle"],"dc:date.available":["2026-06-23T07:00:00Z"],"dc:date.created":["2026-06-23T07:00:00Z"],"dc:description.abstract":["Modern machine learning systems can perform well on benchmarks yet fail in ways that are difficult for users to anticipate, interpret, or manage. This dissertation treats reliability as more than accuracy: a reliable system performs when it can, signals when it should not be trusted, and supports abstention, fallback, escalation, or human review. Unlike robustness, which emphasizes performance under shift, corruption, adversarial perturbation, or drift, reliability also requires recognizing limited competence, preserving useful semantic structure when fine-grained certainty is unjustified, and exposing mechanisms that make failures understandable. The dissertation develops this view through four contributions. Uncertainty-aware decision heads provide portable trust signals for classifiers and detectors. A modular perception-to-risk pipeline combines detection, tracking, intention estimation, and interpretable risk fusion for aerial threat prioritization. Hierarchy-aware classification enables semantic fallback by returning meaningful parent concepts when an exact leaf class is unavailable, unsupported, or withheld. Training-time reliability control is framed as learning to reallocate effort toward difficult or under-served groups. Together, these studies argue that trust improves when systems can state when they should not be trusted, explain why decisions are made, and adapt learning where standard optimization is weakest, yielding a unified view of reliability across model architecture, algorithm design, human interpretability, and decision support."],"dc:identifier":["https://rdw.rowan.edu/etd/3545"],"dc:title":["A Holistic Framework for Enhancing Machine Learning Reliability: From Model Architecture and Algorithm Design to Human Interpretability"],"thesis:degree_discipline":["Electrical and Computer Engineering"],"thesis:degree_level":["Dissertation"],"thesis:degree_name":["Ph.D. Electrical and Computer Engineering"]},"updated_at":"2026-11-19T09:00:05Z"}