Professor Jang-Ho Kim's research team at the Department of Artificial Intelligence, Kookmin University (Master's student An Sang-ho, Master's graduate Kim Jin-woo, and joint research team from Hyundai Motor Company's Robotics Lab) presented a new framework called SERo for lightweighting Vision Transformer (ViT) models at the 41st UAI (Uncertainty in Artificial Intelligence) Conference held in Rio de Janeiro, Brazil, from July 22 to 24. (Sparse Structure Exploration and Re-optimization, a hardware-friendly structure exploration-based ViT compression technique)

SERo stands for Sparse Structure Exploration and Re-optimization, which is a new model lightweighting method that reduces the computation and memory costs of vision transformers by removing and re-optimizing the structure itself, rather than simply setting the parameters to zero.

SERo has been applied to various vision transformer architectures, including DeiT-Tiny, DeiT-Small, and DeiT-Base models, achieving outstanding results: it reduced computational complexity by approximately 69% compared to DeiT-Base, improved inference speed by 2.4 times, and only caused a 1.55% drop in accuracy.

Furthermore, SERo has been applied to various vision tasks, including image classification, object detection (DAMO-YOLO), face recognition (Faceptor), and face attribute analysis (CelebA, LaPa, etc.), demonstrating high compression efficiency and performance maintenance simultaneously.

The UAI (Uncertainty in Artificial Intelligence) conference is one of the world's leading academic conferences focusing on uncertainty inference, probabilistic modeling, Bayesian learning theory, and its applications in the field of artificial intelligence, held annually since 1985.

This research was conducted as a joint research project and industry-academia collaboration between Kookmin University and Hyundai Motor Company's Robotics Lab, and received support from the Ministry of Science and ICT's AI Star Fellowship Program (IITP).