KAIST (President Lee Kwang-hyung) announced on the 23rd that the research team led by Professor Hoe-Jun Yoo of the Department of Electrical and Electronic Engineering has developed the world’s first high-accuracy artificial intelligence (AI) semiconductor by implementing real-time learning of artificial intelligence in a mobile device. Artificial intelligence semiconductor refers to a semiconductor that is equipped with artificial intelligence processing functions such as recognition, reasoning, learning, and judgment, and implemented with optimized technology based on super intelligence, ultra-low power, and super trust.
The artificial intelligence semiconductor developed by the research team can learn on mobile devices by applying low-bit learning and low-latency learning methods. In particular, this semiconductor chip successfully implemented real-time learning technology that can prevent unexpected performance degradation of artificial intelligence.
This research, in which Dong-Hyun Han, Ph.D., Ph.D. student, Department of Electrical and Electronic Engineering, KAIST, participated as the first author, was presented at the International Conference on Artificial Intelligence Circuits and Systems (AICAS) held from June 12 to 15 at Songdo Convensia, Yeonsu-gu, Incheon. was demonstrated at the site, and its excellence was widely publicized by winning both the Best Paper Award and the Best Demo Award. (Paper Title: A 0.95 mJ/frame DNN Training Processor for Robust Object Detection with Real-World Environmental Adaptation (Authors: Dong-Hyun Han, Dong-Seok Lim, Kwang-Tae Park, Young-Woo Kim, Seok-Chan Song, Lee Ju-Hyeong, Hoe-Jun Yoo))
‘AICAS 2022’, an international academic conference covering artificial intelligence (AI) semiconductor technology, is evaluated as the world’s most authoritative IEEE (American Institute of Electrical and Electronic Engineers) society in the field of artificial intelligence semiconductors. It is an event to share research results related to artificial intelligence semiconductor circuits and systems in all fields of artificial intelligence semiconductor circuits and systems, attended by renowned domestic and foreign companies and institutions such as ETRI, NVIDIA, and Cadence.
Existing artificial intelligence made inference only with pre-learned intelligence, so it was difficult to detect an object in a new environment or object that had not been learned. However, the real-time learning developed by Professor Hoi-Jun Yoo’s research team greatly raised the level of intelligence of artificial intelligence by providing a learning function to the existing mobile artificial intelligence semiconductor that only performed reasoning. The new artificial intelligence semiconductor showed high-accuracy object detection performance by utilizing the knowledge learned in advance and the knowledge learned during application execution. In particular, Professor Hoi-Jun Yoo’s research team automatically recognized the unexpected decrease in accuracy of artificial intelligence due to a broken lens or machine error, and corrected it through real-time learning to solve the problems of existing artificial intelligence.
In addition to the real-time learning function, Professor Yoo’s team proposes a low-bit artificial intelligence learning method and a direct error transcription-based low-latency learning method to enable low-power learning on a mobile device, and semiconductor (HNPU) and applications that can optimize it All systems have been developed. The following six core technologies were introduced in HNPU, a semiconductor dedicated to mobile artificial intelligence that can perform low-power, real-time learning. △Low-bit learning method using stochastic dynamic fixed-point (SDFXP: Stochastic Dynamic Fixed-point Representation) △ Automatic precision search algorithm and hardware for each layer (LAPS: Layer-wise Adaptive Precision Scaling) △ Utilizing input bit slice sparsity (ISS: Input Slice Skipping or Bit-slice Level Sparsity Exploitation) △ Intrinsic True Random Number Generator (iTRNG) △ High-speed learning algorithm and hardware through multi-learning step allocation (MLTA: Multi Learning Task Allocation & Backward Unlocking) △ Real-time artificial Development of intelligent learning-based automatic error detection function degradation correction system (Real-time DNN Training based Automatic Performance Monitor and Performance Recovery System). Using this technology, HNPU realizes low-power object detection, achieving 75% higher speed and 44% lower energy consumption compared to other mobile object detection systems, while developing high-accuracy object detection with real-time learning, attracting attention.