NAVER LABS Europe seminars are open to the public. Please register to participate.
Date: 2nd October 2026, 3:00 pm (CEST)
About the speaker: Ramazan Gökberk Cinbiş is an Associate Professor in the Department of Computer Engineering at Middle East Technical University (METU). He received his BSc from Bilkent University in 2008, an M.A. from Boston University in 2010, and a PhD from the Université de Grenoble in 2014, following his doctoral research in the LEAR (now THOTH) team at INRIA. His research focuses on data-efficient learning at the intersection of machine learning, computer vision, and natural language modeling. His current interests span generative models, meta-learning, and learning with limited supervision, with applications bridging vision, language, and robotics.
Abstract: Vision systems often repeat work or overlook useful information already available to them. In this talk, I will present three ways to put that information to work. For streaming video perception, we carry a model’s representation from one frame to the next instead of recomputing it from scratch, enabling efficient segmentation, detection, and pose estimation with models trained in a single-frame setting, without temporal supervision. For knowledge distillation, we transfer knowledge from a teacher to a student, typically a smaller one, even across different architectures. A fixed set of random prototypes provides a shared reference space through which the student learns from the teacher, without requiring a learned pair-specific alignment space. Because these prototypes are independent of the task’s labels, they provide a useful learning signal even when there are only a few classes of interest. Finally, we improve a pretrained autoregressive image generator using feedback about both individual images and how well its overall output distribution matches real data. The resulting model improves image quality even without costly classifier-free guidance at sampling time.