- About Us
- Welcome Message
- Vision
- Brief History
- Staff
- Contact Us
- Faculty / Research
- Faculty
- Research Labs
- Academics
- Curriculum
- Course Guide
- Graduate School
- Graduation Project
- Admissions
- Admissions
- Academic Affairs
- Academic Calendar
- Scholarships
- Graduation Requirements
- Student Clubs
- Academic Information
- Facility Reservations
- Community / News
- NEWS
- Notices
- Event
- Job Openings
- Forms
- KU Q&A
Community / News
NEWS
Research Teams Led by Prof. Sung Whan Yoon of UNIST and Prof. Joongheon Kim of Korea University Reduce AI Training Costs by 80% Using Quantum Computing
관리자 2026.07.11 Views 64
▲ Prof. Sung Whan Yoon of UNIST (Left) and Prof. Joongheon Kim of Korea University (Right)
The Ulsan National Institute of Science and Technology (UNIST) announced on July 6 that a research team led by Professor Sung Whan Yoon of its Graduate School of Artificial Intelligence, together with a team led by Professor Joongheon Kim of the School of Electrical Engineering at Korea University, has developed a quantum computing-based method that reduces the computational cost of artificial intelligence (AI) reinforcement learning to as little as one-fifth of that required by conventional approaches.
The study was accepted for presentation at the 2026 International Conference on Machine Learning (ICML), which is being held at COEX in Seoul from July 6 to 11. ICML is widely regarded as one of the world’s three leading AI conferences, alongside NeurIPS and ICLR.
This year, 6,352 papers were accepted from 23,918 submissions worldwide. The conference, expected to attract approximately 15,000 researchers and industry representatives, is being held on the largest scale in its history. Among the quantum AI papers accepted by ICML this year, this is the only study led by a Korean research institution.
The method developed by the research team is called Quantum Robust Inner Minimization (QRIM). Reinforcement learning is a method through which AI learns behavioral strategies by trial and error. However, its performance can deteriorate sharply when the real-world environment differs even slightly from the environment used during training. A typical example is an autonomous vehicle trained only under clear weather conditions failing to respond properly when it rains or snows.
To address this problem, researchers have studied robust reinforcement learning, which trains AI to identify and prepare for worst-case scenarios during the learning process. However, conventional approaches must examine possible scenarios individually, causing computational costs to increase rapidly as the number of scenarios grows.
QRIM resolves this bottleneck by using the quantum computing principle of superposition. Superposition is a quantum mechanical property that allows a quantum bit to exist in the states of both 0 and 1 simultaneously, enabling multiple possibilities to be represented in parallel.
For example, a conventional method would require 10,000 calculations to examine 10,000 scenarios, whereas QRIM can produce the same result with only 100 evaluations.
In experiments, QRIM achieved greater robustness while requiring only approximately 20–30% of the computational workload of conventional methods. The researchers also conducted verification experiments using an actual IBM quantum computer and confirmed that the method maintained its robustness despite the noise inherent in quantum hardware.
QRIM is also highly versatile, as it can be applied by replacing only the worst-case scenario search component with a quantum module while leaving the existing reinforcement learning algorithm unchanged.
Hyun Kyu Lee, the first author of the study, explained, “This technology introduces a newly designed structure that uses a quantum algorithm to accelerate the search for worst-case scenarios, which is the largest computational bottleneck in robust reinforcement learning.”
Professor Sung Whan Yoon said, “This study provides a concrete example of how quantum computing can complement the limitations of existing AI technologies. It could be applied to fields such as robotics and autonomous driving, where adapting to environmental changes is essential.”
This research was supported by the Mid-Career Researcher Program of the National Research Foundation of Korea and the AI Graduate School Program of the Institute of Information & Communications Technology Planning & Evaluation (IITP), funded by the Ministry of Science and ICT.
The study was accepted for presentation at the 2026 International Conference on Machine Learning (ICML), which is being held at COEX in Seoul from July 6 to 11. ICML is widely regarded as one of the world’s three leading AI conferences, alongside NeurIPS and ICLR.
This year, 6,352 papers were accepted from 23,918 submissions worldwide. The conference, expected to attract approximately 15,000 researchers and industry representatives, is being held on the largest scale in its history. Among the quantum AI papers accepted by ICML this year, this is the only study led by a Korean research institution.
The method developed by the research team is called Quantum Robust Inner Minimization (QRIM). Reinforcement learning is a method through which AI learns behavioral strategies by trial and error. However, its performance can deteriorate sharply when the real-world environment differs even slightly from the environment used during training. A typical example is an autonomous vehicle trained only under clear weather conditions failing to respond properly when it rains or snows.
To address this problem, researchers have studied robust reinforcement learning, which trains AI to identify and prepare for worst-case scenarios during the learning process. However, conventional approaches must examine possible scenarios individually, causing computational costs to increase rapidly as the number of scenarios grows.
QRIM resolves this bottleneck by using the quantum computing principle of superposition. Superposition is a quantum mechanical property that allows a quantum bit to exist in the states of both 0 and 1 simultaneously, enabling multiple possibilities to be represented in parallel.
For example, a conventional method would require 10,000 calculations to examine 10,000 scenarios, whereas QRIM can produce the same result with only 100 evaluations.
In experiments, QRIM achieved greater robustness while requiring only approximately 20–30% of the computational workload of conventional methods. The researchers also conducted verification experiments using an actual IBM quantum computer and confirmed that the method maintained its robustness despite the noise inherent in quantum hardware.
QRIM is also highly versatile, as it can be applied by replacing only the worst-case scenario search component with a quantum module while leaving the existing reinforcement learning algorithm unchanged.
Hyun Kyu Lee, the first author of the study, explained, “This technology introduces a newly designed structure that uses a quantum algorithm to accelerate the search for worst-case scenarios, which is the largest computational bottleneck in robust reinforcement learning.”
Professor Sung Whan Yoon said, “This study provides a concrete example of how quantum computing can complement the limitations of existing AI technologies. It could be applied to fields such as robotics and autonomous driving, where adapting to environmental changes is essential.”
This research was supported by the Mid-Career Researcher Program of the National Research Foundation of Korea and the AI Graduate School Program of the Institute of Information & Communications Technology Planning & Evaluation (IITP), funded by the Ministry of Science and ICT.
File
Prev
2026.07.05
Next
No next post.
Engineering Building #412, College of Engineering, 145 Anam-ro, Seongbuk-gu, Seoul, 02841, KoreaTEL : 02-3290-3210, 3211, 3213, 3214, 4592FAX : 02-921-0544
Copyrights ⓒ 2021 KOREA University. All Rights Reserved.
- Family Site