Math Modeling Seminar: Learning Beyond the Training Distribution: Toward Adaptive and Open-World Learning Systems

Math Modeling Seminar
Learning Beyond the Training Distribution: Toward Adaptive and Open-World Learning Systems

Zoom Link here

Dr. Tyler Hayes
Research Faculty and Research Scientist
Georgia Institute of Technology

Abstract:
Modern artificial intelligence (AI) and machine learning systems excel when deployed in environments that mirror their training data, yet the real world is open, dynamic, and continually evolving. This talk explores how we can build models that not only survive but thrive under distributional shifts, i.e., systems capable of adapting to the unknown. I will introduce two complementary paradigms for adaptation in neural networks: dynamic adaptation, where models continuously learn from incoming data on-the-fly, and static adaptation, where models are prepared in advance to recognize unseen concepts using external knowledge sources. I will present our recent work on PANDAS, a prototype-based approach for discovering and detecting novel object classes, and SHiNe, which leverages class taxonomies and large vision-language models to improve open-vocabulary object detection. Together, these methods illustrate how dynamic and static adaptation can be used to enable open-world systems that generalize beyond their training distribution. I will then conclude with my vision for the future of novel concept discovery and its broader applications.

Bio:
Tyler L. Hayes is a Research Faculty member and Research Scientist in the College of Computing at the Georgia Institute of Technology, where she conducts research at the intersection of artificial intelligence and scientific discovery. She earned her bachelor’s and master’s degrees in Applied and Computational Mathematics at RIT, and later a Ph.D. in Imaging Science, where she focused on efficient lifelong machine learning for computer vision under the supervision of Dr. Christopher Kanan. Her work spans online continual learning, novel class discovery, open-world learning, and explainable AI, with publications in top venues including CVPR, ECCV, TMLR, and BMVC. Before Georgia Tech, Tyler was a Research Scientist at NAVER LABS Europe, developing methods for novel class discovery and open-vocabulary detection using vision-language models.

Intended Audience: Beginners, undergraduates, graduates. Those with interest in the topic.

Interpreters have been requested.