Title: Adapting to the Unknown: Universal Novelty Detection Across Datasets and Tasks
Speaker:
Mohammad Sabokrou
Staff Scientist
Machine Learning and Data Sciences Unit
Okinawa Institute of Science and Technology (OIST), Japan
Date: November 10, 2025, Monday
Time: 13:30
Place: EA 409
Abstract:
Modern AI systems often struggle when faced with data that differs from what they were trained on. In this talk, I will present our efforts toward universal and adaptable novelty detection, focusing on methods that generalize across new datasets, tasks, and unseen scenarios. I will discuss two complementary directions—automatic augmentation for out-of-distribution detection and zero-shot anomaly detection using vision–language models—that together move toward more flexible and robust AI systems capable of adapting to the unknown. I will conclude by sharing my vision and future research plans on AI trustworthiness and how these ideas connect to building more reliable intelligent systems.
Bio:
Mohammad Sabokrou is a Staff Research Scientist at the Okinawa Institute of Science and Technology (OIST), Japan. He received his PhD in Artificial Intelligence in 2017 and has since held research positions in Iran, Finland, and France. His research focuses on trustworthy AI, with core interests in anomaly and out-of-distribution detection and robust machine learning. He has published in top venues such as ICCV, CVPR, NeurIPS, and ICLR, and currently serves as an Area Chair for ICLR 2025 and ICLR 2026.