A More Stable and Reliable Artificial Neuron Array

A KAIST team has developed an advanced memristor array capable of human neuron cell replication.

KAIST announced on June 7 that its research team led by electrical and electronic engineering professor Choi Shin-hyun developed an advanced memristor array capable of human neuron cell replication.

A memristor is characterized by being capable of simultaneous information storage and operation based on input-dependent resistor states, which cannot be achieved by existing transistor chips.

The team developed a more stable and reliable artificial neuron array by using a metal oxide with a gradual oxygen concentration instead of a filament. “Existing memristor elements are less stable and hard to be given an array form for application, but our element is highly stable and can be integrated in the form of a large-capacity array based on its self-rectifying characteristics and high yield,” the team said, adding, “The element can be suitable for highly integrated and stable neuromorphic system configuration.”

Computers consume much energy in handling big data whereas the human brain can do so with little energy. In this regard, researches on neuromorphic hardware technologies are increasing so that the efficient neural signal transmission can be applied to computing. Memristor devices, which can be highly efficiently integrated, are emerging as elements for neuromorphic computing system configuration.

Existing memristors have their own limitations in terms of reliability and yield in realizing practical large-scale neural computing. The filaments in the insulators of those memristors are generated, disappear and operate at random and thus hard to control and unreliable, which means a stable neuromorphic system cannot be obtained from them.

The team solved this problem by making a variable resistance element using a gradual oxygen ion movement. In addition, it developed a unit element-based array production technique and succeeded in integrating 400 artificial neuron elements with a 100 percent yield and in the form of a crossbar array. With its improved neuromorphic system, the team succeeded in anti-microbial peptide amino acid sequence learning and realized a system for new anti-microbial peptide creation.

“The developed artificial neuron device is expected to have various applications, such as robotic nervous systems for tactile sensing and reservoir computing for time series data processing,” it explained. Details of the research have been published in Nature Communications.

Copyright © BusinessKorea. Prohibited from unauthorized reproduction and redistribution