TagsAI Seminar 2025: Test-Time Updates: Next Steps for Robust and Efficient Vision, Evan Shelhamer
Abstract
The AI Seminar is a weekly meeting at the University of Alberta where researchers interested in artificial intelligence (AI) can share their research. Presenters include both local speakers from the University of Alberta and visitors from other institutions. Topics can be related in any way to artificial intelligence, from foundational theoretical work to innovative applications of AI techniques to new fields and problems.
Learning from data is essential for state-of-the-art vision, but learning once may not be enough. When the data shifts between training and testing the accuracy of predictions can degrade. The problem is that the data changes, but our systems remain the same. Does it have to stay this way? In this talk we will examine how visual recognition can adapt and generalize to new and different data during testing. We will cover natural shifts, such as image corruptions, to highlight opportunities to update models (by minimizing the entropy of predictions and by mixing parameters) when the data differs. Then we will take a critical look at adversarial shifts with a case study of adaptive defenses. As a last step, we will discuss test-time updates as a kind of multi-modeling—or how to apply multiple models at the same time—and summarize progress on efficiency, uncertainty, and adaptivity by multi-modeling.
Bio:
Evan Shelhamer is an assistant professor at UBC in Vancouver and a CIFAR AI Chair at the Vector Institute. He earned his PhD at UC Berkeley in 2019 advised by Prof. Trevor Darrell. While there he was the lead developer of the Caffe open-source deep learning framework from version 0.1 to 1.0. He then worked at Google DeepMind and Adobe before returning to academia. His research and service have received international awards, including for fully convolutional networks (best paper honorable mention at CVPR’15, test-of-time at CVPR’25), and for Caffe (the Mark Everingham award at ICCV’17, the open-source award at MM’14, and the test-of-time at MM’24). He likes to brew coffee and community, and his latest organizing efforts include the 2nd workshop on test-time adaptation at ICML’25 and the 3rd workshop on machine learning for remote sensing at ICLR’25. He is new to Canada, and welcomes your input and advice as he adapts and makes his own updates.