2026-05-22T10:49:47Zhttps://keep.lib.asu.edu/oai/request

oai:keep.lib.asu.edu:node-2024082025-08-18T22:22:09Zoai_pmh:alloai_pmh:repo_items

202408 https://hdl.handle.net/2286/R.2.N.202408 http://rightsstatements.org/vocab/InC/1.0/ All Rights Reserved 2025 119 pages Doctoral Dissertation Academic theses en Kamboj, Payal Gupta, Sandeep Davulcu, Hasan Banerjee, Ayan Cescon, Marzia Lee, Kookjin Arizona State University Partial requirement for: Ph.D., Arizona State University, 2025 Field of study: Computer Science Rare events are occurrences that happen infrequently but have significant importance due to their unique characteristics and impactful consequences. Automated detection of such events poses a challenge due to the limited availability of data, which complicates the estimation of their distribution. Examples of rare events include critical pathological regions in medical domain and construction zones encountered during autonomous vehicle navigation. Furthermore, variability across data sources, sensing modalities, environmental conditions, and acquisition protocols introduces domain shifts that amplify inconsistencies in rare event characteristics. In fields such as medical, additional variability arises from biological and demographic factors such as age, gender, and comorbidities, which can alter the visual or physiological presentation of critical findings. This variability hinders the ability to generalize across different domains, thereby leading to the domain generalization (DG) problem. Current state-of-the-art deep learning (DL) techniques typically rely on large, annotated datasets to perform well in automated rare event detection. However, the scarcity of data on these rare events makes it difficult to accurately model their distribution even within a single domain and manage covariate shifts across different domains. To address these challenges in rare event detection and DG, this work explores and proposes integrating expert knowledge with DL methods. Expert knowledge can help mitigate data scarcity by offering additional information that often remains consistent across domains. By leveraging this expertise, models can potentially achieve effective DG, allowing them to apply learned knowledge to new, unseen domains. This work aims to enhance rare event detection in a structured manner, even with extremely limited data, by addressing challenges related to data scarcity, class imbalance, and high intra-class variability within rare classes in a multi-class setting. By combining expert knowledge with DL, this work proposes to significantly improve the detection of rare events and enable robust DG. Finally, we also demonstrate that DL and expert knowledge integration also has the potential to enable zero-shot learning for rare events when utilized with large language models (LLM)s and language vision models (LVM)s. Computer Science Rare Event Detection and Domain Generalization using Expert Knowledge and Deep Learning