PhD Dissertation Defense by ZHU Jiawen | Toward generalist anomaly detectors

ABOUT THE TALK

This dissertation investigates generalist anomaly detection, aiming to improve generalization across unseen anomaly types and diverse domains while progressively reducing dependence on target-domain data. It covers open-set supervised, few-shot, and zero-shot anomaly detection through four works. AHL models anomaly heterogeneity to improve generalization to unseen anomaly types. InCTRL introduces in-context residual learning with a few normal reference samples for cross-domain anomaly detection without target-domain retraining. InCTRL v2 further integrates semantic-guided abnormality and normality modeling to improve both anomaly detection and localization under large domain shifts. Finally, FAPrompt enables zero-shot anomaly detection by learning fine-grained abnormality prompts within vision-language models, requiring no target-domain data. Together, these works establish a coherent progression toward robust, scalable, and increasingly domain-agnostic anomaly detection systems.

ABOUT THE SPEAKER

Jiawen ZHU is pursuing her doctoral research in Computer Science at Singapore Management University, focusing on anomaly and deepfake detection, multimodal learning, and generalizable and trustworthy AI. Her dissertation investigates how anomaly detection systems can generalize across unseen anomaly types and diverse domains with limited or no target-domain supervision. Her work spans open-set, few-shot, and zero-shot anomaly detection, with publications at major AI conferences including CVPR and ICCV.