A technology has been developed to create an AI that is as smart as if it had learned everything by learning only 1/25th of the amount it originally had to learn.

A research team led by Professor Jae-Young Shim of UNIST's Graduate School of Artificial Intelligence announced on the 1st that it has developed a data distillation technology that can maximize the learning efficiency of 3D artificial intelligence (AI) models, which are key to cutting-edge industries such as self-driving cars and robots.

This achievement has garnered global attention as it was officially accepted as a paper at the international artificial intelligence conference NeurIPS 2025.

3D point cloud data is a method of representing objects as points, and serves as the 'eyes' of self-driving cars, drones, and robots.

On the other hand, data distillation is difficult to apply because there is no set order in the arrangement of points and there is a large rotational variability.

Existing technologies inevitably suffered from performance degradation due to matching errors occurring during the process of comparing the characteristics of original data and summarized data.

Professor Shim Jae-young's team introduced two key techniques to solve these problems.

The SADM loss function automatically sorts point data with different order to match the semantic structure, and the Learnable Rotation techniqueThe AI was made to learn and optimize the rotation angle of the object on its own.

This fundamentally resolves the matching errors of existing technologies and allows the original performance to be maintained even during the data summarization process.

The research team verified performance using the representative 3D dataset ModelNet40.

Even though it was trained with summarized data that was reduced to 1/25th of the original data, it recorded 80.1% recognition accuracy.

This result is not significantly different from the 87.8% accuracy achieved when using the entire data, proving that learning efficiency and performance can be balanced even at high compression ratios.

Professor Shim Jae-young emphasized, “This technology fundamentally solves the existing matching errors caused by the disorderly structure and rotational uncertainty of 3D point data,” and “It will contribute to significantly reducing AI learning costs and time in fields that require the use of large-scale 3D data, such as autonomous driving, drones, robots, and digital twins.”

This research was conducted with the support of the National Research Foundation of Korea (NRF) under the Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA).

The research results have been accepted as an official paper at the Neural Information Processing Systems (NeurIPS) 2025, one of the world's top three artificial intelligence conferences. The conference will be held in San Diego, USA, from December 2nd to 7th, and UNIST's achievements are expected to garner significant attention from global AI researchers.