Domestic Researchers Present Achievements at ICRA
AI Learning Inspired by Brain Learning Methods for Faster and Easier Training
A new artificial intelligence (AI) learning method using Hyperdimensional Computing (HDC), known as the "dream AI," has been commercialized by domestic researchers. HDC is 15 times faster and consumes less power than the deep learning methods used in AI such as ChatGPT.
Professor Emeritus Seo Ilhong of Hanyang University and CEO of Koga Robotics (right) and Professor Kim Yesung of the Department of Electrical, Electronics, and Computer Engineering at DGIST attended the annual conference on the 14th in Yokohama, Japan, to present a paper at ICRA, the world's largest robotics conference. [Photo by Koga Robotics]
Kogarobotics announced on the 14th that the paper "Hyperdimensional Computing Inspired by the Human Brain," co-authored by a total of 12 researchers including CEO Ilhong Seo, an emeritus professor at Hanyang University, and Yesung Kim, a professor at Daegu Gyeongbuk Institute of Science and Technology (DGIST), was officially presented at the ICRA (International Conference for Robot & Automation), the world's largest robotics conference, held in Yokohama, Japan, after undergoing verification.
HDC is regarded as a next-generation learning method that can replace deep learning, which has been criticized for consuming enormous costs and power in AI training and inference. HDC is designed by mimicking the brain's computational method, storing information not in specific neurons but distributed across many neurons. It corresponds and combines all objects, concepts, and functions to hyperdimensional vectors represented by thousands of unique vectors, enabling fast inference results through simple calculations.
Software-based brain-inspired learning methods involve values input into artificial neural networks being processed through numerous synapses between multilayered nodes. Because this method involves a vast amount of matrix operations, the computational load increases exponentially as the neural network grows to improve AI performance. Ultimately, this leads to increased system construction costs such as expensive GPUs and higher power consumption.
Consequently, AI learning using HDC, which requires only minimal memory and computation, has emerged as a key solution. Although some research on HDC has been conducted overseas, it has remained at the theoretical research stage. This is the world's first case of practical application of HDC and its use in autonomous robot driving, with a paper presented at a global academic conference.
Applying HDC to indoor autonomous driving robots and comparing it with deep learning showed that it can achieve speeds 15 times faster at about 1/30th the cost. Additionally, power consumption is about 1/20th, making it suitable for lightweight, on-device AI implementation. Ilhong Seo, CEO of Kogarobotics, explained, "Existing deep learning-based AI algorithms show very high-quality learning results, but as the model size increases, the cost burden also rises. The significance lies in developing technology that performs both training and inference processes in an on-device robot environment using lightweight AI technology."
In particular, this research developed technology that can autonomously learn given goals without human intervention. It uses HDC to mimic the perception-action relationship that reads LiDAR data measuring distance in 360 degrees to control motors, and employs reinforcement learning functions that provide rewards during the learning process.
Hyuntaek Choi, president of the Robotics Society, said, "In a situation where global big tech companies almost monopolize core AI and robotics technologies, the paper presentation by Dr. Ilhong Seo and Professor Yesung Kim has allowed us to showcase the level of AI and robotics technology in Korea. This will be a milestone demonstrating that HDC technology can be applied to fields requiring on-site learning in the future."
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
