In a world increasingly reliant on high-quality data, how can we automatically find the anomalies, especially when we often only have examples of what is "normal"?

In this Upper Bound 2025 session, Amii Fellow and Canada CIFAR AI Chair Xingyu Li discusses the evolution of visual anomaly detection, a critical technology for industrial quality control and medical diagnosis.

Li explains the journey from traditional "one-class" models, which required separate training for every object category, to their innovative "reverse distillation" technique, which improves detection by forcing a "student" AI to infer logic rather than just mimic a "teacher." She then details their ultimate goal: a scalable, zero-shot, "one-for-all" system. By cleverly adapting large Vision-Language Models to focus on fine-grained details instead of just broad concepts, their latest work can identify defects in any new, unseen category without any retraining, paving the way for truly automated data curation.

Upper Bound 2025 is Amii's annual artificial intelligence conference, held in Edmonton, Alberta, Canada, bringing together researchers, industry, and policymakers. The conference focuses on accelerating AI excellence and innovation for good, emphasizing AI for critical infrastructure, health, industrial operations, responsible AI, and AI Literacy.