Kookmin University team develops AI technology to slim down autonomous driving object detection

Researchers at Kookmin University have pioneered a new technology, 'SharedKD,' designed to make artificial intelligence models for autonomous driving significantly more efficient. The breakthrough, developed in collaboration with Hyundai Motor, focuses on 'knowledge distillation' – a method for shrinking AI models.

“

The research team developed 'SharedKD' to efficiently reduce the size of 3D object detection models for autonomous driving.

Unlike traditional approaches that use separate 'teacher' and 'student' models, SharedKD integrates both within a single 3D object detection model. The entire network functions as the teacher, guiding a pruned, smaller network that acts as the student. This dynamic process uses gradients to identify and retain crucial structural elements, ensuring high accuracy while achieving substantial size reduction.

“

The new method uses a single 3D object detection model, where the entire network is the teacher model and a pruned part is the student model.

This innovative technique moves beyond simply shrinking already trained AI models. It allows for the simultaneous exploration and training of lightweight models suitable for real-world deployment directly within a larger model. The mutual learning process between the full and lightweight models enhances efficiency, making it a promising advancement for applications demanding both high recognition accuracy and real-time processing, such as in autonomous vehicles.

“

This research is meaningful in that it allows for the simultaneous exploration and learning of lightweight models suitable for actual deployment within a large-scale model, beyond existing methods of reducing the size of already trained AI models.

The research aligns with Kookmin University's 'KMU Vision 2035: EDGE' strategy, particularly in the 'AI+X' and 'Mobility' focus areas. By applying AI model optimization to autonomous driving and engaging in industry collaboration, the university demonstrates a concrete path for its specialization.

“

It is expected to be utilized in in-vehicle AI and edge AI systems in the future, as it can significantly reduce the amount of computation and execution time of 3D recognition models in environments that require both high recognition accuracy and real-time processing, such as autonomous driving vehicles.