A sensor semiconductor able to ultra-high-speed, low-power operation of a synthetic intelligence (AI) vision system for motion recognition has been developed. The reason is that this makes it possible to cut back data traffic and power consumption in various on-device AI fields.
The Korea Advanced Institute of Science and Technology (KAIST, President Gwanghyung Lee) announced on the nineteenth that Professor Kyungmin Kim's research team within the Department of Materials Science and Engineering has succeeded in developing an intelligent motion recognition device that mimics the visual intelligence of the optic nerve of an insect by fusing various memristor devices.
The researchers focused on the incontrovertible fact that insects effectively process visual information through an optic nerve circuit called the fundamental motion detector to detect objects and recognize motion.
Nevertheless, it was identified that existing silicon integrated circuit (CMOS) technology required complex circuits to implement it in an AI vision system, making it difficult to fabricate into actual devices.
Subsequently, the research team said that they developed an intelligent motion recognition device able to highly efficient and ultra-fast motion recognition by integrating memristor devices with various functions.
Neurons within the optic ganglion of insects include Tm3, which delays signals coming from outside, Mi1, which transmits them immediately, and T4, which processes and responds. Tm3 and Mi1 each transmit external signals to be processed and responded to by T4, allowing insects to quickly recognize the movement of objects.
The research team implemented this principle with two sorts of memristor elements and resistors to effectively process motion information of moving objects, reduce data traffic between memory and computing devices, and efficiently provide direction information at the extent of high-performance sensors to the neural network. did.
As well as, to reveal the opportunity of practical use of the developed motion recognition device, a neuromorphic computing system that predicts vehicle paths was designed, and the developed motion recognition device was applied to it.
Consequently, it was verified that the movement of objects might be predicted more accurately by reducing energy consumption by 92.9% in comparison with existing technology.
Professor Kyung-min Kim said, “Insects use quite simple visual intelligence to acknowledge the motion of objects with surprising agility, and this study is important in that it was in a position to implement this using a memristor device that may reproduce the function of nerves.” “There may be,” he said.
He continued, “Recently, the importance of edge AI devices, comparable to mobile phones equipped with AI, is growing significantly, and this research can contribute to the implementation of an efficient vision system for motion recognition, which shall be utilized in the longer term for ▲autonomous vehicles ▲vehicle transportation systems ▲robots ▲machine vision. “It is anticipated that it could actually be applied to varied fields comparable to,” he said.
This study was co-first authored by PhD student Song Han-chan and PhD student Min-gu Lee of the Department of Materials Science and Engineering at KAIST as co-first authors, and was published online on January 29 within the international academic journal 'Advanced Materials'.
Meanwhile, this research was conducted with support from the National Research Foundation of Korea's mid-career research project, the next-generation intelligent semiconductor technology development project, the PIM artificial intelligence semiconductor core technology development project, the Nano Advanced Institute of Technology, and the KAIST Leap Research Project.
Reporter Park Soo-bin sbin08@aitimes.com