Cho Hyun-jun and Ahn Sang-ho, master’s students in the Department of AI and Software Engineering at Kookmin University (President Jeong Seung Ryul) and members of Professor Kim Jangho’s research team, have developed “SharedKD,” a technology that efficiently streamlines 3D object detection models for autonomous driving through joint research with Hyundai Motor Company. The research findings were presented last July at DAC (Design Automation Conference) 2026, a world-renowned international academic conference in the field of design automation.

Unlike existing knowledge distillation methods that use separate teacher and student models, the SharedKD developed by the research team is characterized by utilizing the entire network within a single 3D object detection model as the teacher model, and the subnetworks created through pruning as the student model. In particular, during the training process, it dynamically selects important structures based on gradients, enabling the exploration of efficient, lightweight models while maintaining high accuracy.

This research is significant in that it goes beyond existing methods of retroactively reducing the size of pre-trained AI models; it enables the simultaneous exploration and training of lightweight models suitable for actual deployment within large-scale models. Furthermore, the model is designed so that the full model and the lightweight model assist each other’s training, enabling efficient model exploration and training.

In environments such as autonomous vehicles, where high recognition accuracy and real-time processing are required simultaneously, this approach can significantly reduce the computational load and execution time of 3D perception models, and is expected to be utilized in future automotive AI and edge AI systems.

This research aligns with the future specialization strategy that Kookmin University is pursuing through “KMU Vision 2035: EDGE.” In particular, it is significant as a research achievement that concretely demonstrates the university’s specialization direction—specifically at the intersection of “AI+X” and “Mobility,” two of the eight key specialization areas—by applying AI model optimization technology to the field of autonomous driving and conducting joint research with industry partners.

Meanwhile, the Design Automation Conference (DAC) is a world-renowned international academic conference representing the fields of electronic circuit and system design and design automation, where a wide range of research findings—from semiconductor design to artificial intelligence and machine learning systems, as well as efficient AI implementation—are presented.

△ Photo: Master’s student Jo Hyun-jun presenting a paper at DAC 2026