{"id":{"repo_id":"stellenbosch","oai_identifier":"oai:scholar.sun.ac.za:10019.1/135655"},"canonical_url":"https://search.dev.ndltd.org/etd/stellenbosch/oai:scholar.sun.ac.za:10019.1/135655","repository":{"repo_id":"stellenbosch","name":"Stellenbosch University","base_url":"https://scholar.sun.ac.za/server/oai/request"},"display":{"title":"The Removal of False Signals from Convolutional Neural Networks","abstract":"Convolutional neural networks have achieved performance comparable to human experts in various computer vision tasks. However, despite this comparable performance, the decision-making process differs significantly between convolutional neural networks and human experts. Whereas human decisions are guided by causal inference, convolutional neural networks rely on associations that may not reflect causal relationships. When the decision of a convolutional neural network is based on false associations, learned from confounding features in the training data, the model will produce incorrect predictions when the false associations do not hold. To prevent failures in critical domains such as healthcare, false relationships must be identified and removed. This dissertation critically evaluates, compares and improves methods to identify and remove false associations from convolutional neural networks. A taxonomy of common confounding features is introduced along with a second taxonomy that categorises confounders by their key characteristics, to support the principled selection of appropriate removal strategies. A general evaluation framework is further developed to quantify the effectiveness of con-founder removal methods across a wide range of confounding features. The framework is applied to both established and newly proposed methods which allowed the relative effectiveness of confounding removal methods to be established. The results highlight significant variation in method effectiveness, demonstrate practical shortcomings of current methods, and show that targeted refinements can yield measurable improvements. Overall, the dissertation provides a structured foundation for the development and reliable evaluation of confounder robust convolutional neural networks.","abstract_html":"Convolutional neural networks have achieved performance comparable to human experts in various computer vision tasks. However, despite this comparable performance, the decision-making process differs significantly between convolutional neural networks and human experts. Whereas human decisions are guided by causal inference, convolutional neural networks rely on associations that may not reflect causal relationships. When the decision of a convolutional neural network is based on false associations, learned from confounding features in the training data, the model will produce incorrect predictions when the false associations do not hold. To prevent failures in critical domains such as healthcare, false relationships must be identified and removed. This dissertation critically evaluates, compares and improves methods to identify and remove false associations from convolutional neural networks. A taxonomy of common confounding features is introduced along with a second taxonomy that categorises confounders by their key characteristics, to support the principled selection of appropriate removal strategies. A general evaluation framework is further developed to quantify the effectiveness of con-founder removal methods across a wide range of confounding features. The framework is applied to both established and newly proposed methods which allowed the relative effectiveness of confounding removal methods to be established. The results highlight significant variation in method effectiveness, demonstrate practical shortcomings of current methods, and show that targeted refinements can yield measurable improvements. Overall, the dissertation provides a structured foundation for the development and reliable evaluation of confounder robust convolutional neural networks.","abstract_has_math":false,"creators":["Burger, Leon Eldon"],"institution":"Stellenbosch : Stellenbosch University","degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":["Engelbrecht, A. P."],"committee_chairs":[],"committee_members":[],"year":2026,"date_issued":"2026-03","date_published":"2026-03","updated_at":"2026-07-24T04:40:09Z","subjects":[],"languages":["en"],"rights":[],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"https://scholar.sun.ac.za/handle/10019.1/135655","outbound_label":"Repository record","outbound_source":"dc:identifier.uri"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["Engelbrecht, A. P."]},{"key":"dc:contributor.other","label":"Dc Contributor Other","values":["Stellenbosch University. Faculty of Engineering. Dept. of Industrial Engineering."]},{"key":"dc:creator","label":"Author","values":["Burger, Leon Eldon"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.accessioned","label":"Dc Date Accessioned","values":["2026-04-07T09:01:36Z"]},{"key":"dc:date.available","label":"Dc Date Available","values":["2026-04-07T09:01:36Z"]},{"key":"dc:date.issued","label":"Date","values":["2026-03"]},{"key":"dc:publisher","label":"Institution","values":["Stellenbosch : Stellenbosch University"]},{"key":"dc:type","label":"Dc Type","values":["Thesis"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language.iso","label":"Language (ISO)","values":["en"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://scholar.sun.ac.za/handle/10019.1/135655"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Thesis (PhD)--Stellenbosch University, 2026.","Burger, L. E. 2026. The Removal of False Signals from Convolutional Neural Networks. Unpublished doctoral dissertation. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/760a798b-c924-412a-9b44-da5aefe4a79e"]},{"key":"dc:description.abstract","label":"Abstract","values":["Convolutional neural networks have achieved performance comparable to human experts in various computer vision tasks. However, despite this comparable performance, the decision-making process differs significantly between convolutional neural networks and human experts. Whereas human decisions are guided by causal inference, convolutional neural networks rely on associations that may not reflect causal relationships. When the decision of a convolutional neural network is based on false associations, learned from confounding features in the training data, the model will produce incorrect predictions when the false associations do not hold. To prevent failures in critical domains such as healthcare, false relationships must be identified and removed. This dissertation critically evaluates, compares and improves methods to identify and remove false associations from convolutional neural networks. A taxonomy of common confounding features is introduced along with a second taxonomy that categorises confounders by their key characteristics, to support the principled selection of appropriate removal strategies. A general evaluation framework is further developed to quantify the effectiveness of con-founder removal methods across a wide range of confounding features. The framework is applied to both established and newly proposed methods which allowed the relative effectiveness of confounding removal methods to be established. The results highlight significant variation in method effectiveness, demonstrate practical shortcomings of current methods, and show that targeted refinements can yield measurable improvements. Overall, the dissertation provides a structured foundation for the development and reliable evaluation of confounder robust convolutional neural networks."]},{"key":"dc:title","label":"Title","values":["The Removal of False Signals from Convolutional Neural Networks"]}]}],"canonical_facts":{"dc:contributor.advisor":["Engelbrecht, A. P."],"dc:contributor.other":["Stellenbosch University. Faculty of Engineering. Dept. of Industrial Engineering."],"dc:creator":["Burger, Leon Eldon"],"dc:date.accessioned":["2026-04-07T09:01:36Z"],"dc:date.available":["2026-04-07T09:01:36Z"],"dc:date.issued":["2026-03"],"dc:description":["Thesis (PhD)--Stellenbosch University, 2026.","Burger, L. E. 2026. The Removal of False Signals from Convolutional Neural Networks. Unpublished doctoral dissertation. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/760a798b-c924-412a-9b44-da5aefe4a79e"],"dc:description.abstract":["Convolutional neural networks have achieved performance comparable to human experts in various computer vision tasks. However, despite this comparable performance, the decision-making process differs significantly between convolutional neural networks and human experts. Whereas human decisions are guided by causal inference, convolutional neural networks rely on associations that may not reflect causal relationships. When the decision of a convolutional neural network is based on false associations, learned from confounding features in the training data, the model will produce incorrect predictions when the false associations do not hold. To prevent failures in critical domains such as healthcare, false relationships must be identified and removed. This dissertation critically evaluates, compares and improves methods to identify and remove false associations from convolutional neural networks. A taxonomy of common confounding features is introduced along with a second taxonomy that categorises confounders by their key characteristics, to support the principled selection of appropriate removal strategies. A general evaluation framework is further developed to quantify the effectiveness of con-founder removal methods across a wide range of confounding features. The framework is applied to both established and newly proposed methods which allowed the relative effectiveness of confounding removal methods to be established. The results highlight significant variation in method effectiveness, demonstrate practical shortcomings of current methods, and show that targeted refinements can yield measurable improvements. Overall, the dissertation provides a structured foundation for the development and reliable evaluation of confounder robust convolutional neural networks."],"dc:identifier.uri":["https://scholar.sun.ac.za/handle/10019.1/135655"],"dc:language.iso":["en"],"dc:publisher":["Stellenbosch : Stellenbosch University"],"dc:title":["The Removal of False Signals from Convolutional Neural Networks"],"dc:type":["Thesis"]},"updated_at":"2026-07-24T04:40:09Z"}