Ayoub Karine
title
Green Deep Learning via Knowledge Distillation for Efficient Vision and Vision-Language Models
Abstract
This talk presents Green Deep Learning achieved through Knowledge Distillation for efficient vision and efficient vision-language models (VLMs). We address the trade-off between performance and computational efficiency. Applications include efficient semantic segmentation and point cloud quality assessment for vision, as well as efficient Visual Question Answering (VQA) on documents for VLMs. By transferring knowledge from large teacher models to smaller students, we demonstrate that deep learning can be both high-performing and resource-efficient.
Biography
Ayoub Karine received the M.Sc. degree from Mohammed V University, Rabat, Morocco, in 2014, and a joint Ph.D. degree from Mohammed V University and ENSTA Bretagne, France, in 2018. From 2017 to 2019, he was a Temporary Assistant and Researcher at Université de Haute Alsace, France. He then served as an Associate Professor at ISEN Yncréa Ouest, Nantes, France, from 2019 to 2024.
He is currently an Associate Professor at Université Paris Cité, and a member of the SIP team of the LIPADE laboratory. His research interests include machine learning, deep learning, and computer vision. He has served as a TPC member and reviewer for several leading international conferences and journals.