The research team of KAIST professor Kim Kyung-min (far right). [Provided by KAIST]
The research team of KAIST professor Kim Kyung-min (far right). [Provided by KAIST]

South Korean researchers have developed a technology that harnesses the "noise" generated in semiconductors to process information the way the human brain does, rather than simply eliminating it. The advance is expected to contribute to the development of low-power, brain-inspired semiconductors capable of handling diverse signals on a single chip.

KAIST announced Sunday that a research team led by professor Kim Kyung-min of the Department of Materials Science and Engineering has developed a new neuromorphic neuron semiconductor technology that uses naturally occurring noise in semiconductors to process information in a manner similar to the human brain. The technology can select signals at a desired frequency and convert them into electrical signals resembling those produced by neurons in the brain.

In electronic devices, noise — like the static hiss of a radio — has long been treated as something to be eliminated, as it mixes irregularly with signals and interferes with accurate information reading. The human brain, by contrast, does not have neurons that respond identically to the same stimulus every time; this irregularity helps the brain respond flexibly to a wide range of stimuli and changes.

The research team used a semiconductor device called a memristor — a component whose electrical resistance changes when a voltage is applied and that retains its state even after power is cut. In simple terms, it functions as a small electronic memory chip that remembers how readily electricity flows through it.

The team observed that changing the resistance state of a memristor also alters the magnitude and fluctuation characteristics of current noise. By tuning the memristor's resistance, they found they could simultaneously control the properties of the noise it produces.

The team amplified this tuned noise to generate irregular electrical signals called "spikes" — similar to those neurons produce when transmitting information in the brain. This enabled them to build a "programmable probabilistic neuron" (PPN) that responds stochastically, like a biological nerve cell, by exploiting noise as a resource.

Just as human neurons do not always respond identically to the same stimulus, this artificial neuron varies its spike-firing probability depending on the input signal and the configured noise characteristics — turning what was once considered interference into a resource for information processing. The team also designed the system so that simply adjusting the memristor's state changes which signal speeds the artificial neuron responds to most readily.

A schematic diagram of the artificial neuron (AI-generated image). [Provided by KAIST]
A schematic diagram of the artificial neuron (AI-generated image). [Provided by KAIST]

Using these artificial neurons, the research team encoded body-movement signals in the range of a few hertz and voice signals in the kilohertz range into spike signals, then verified recognition performance. The system achieved 94.8 percent accuracy in motion recognition and 95.0 percent in speech recognition.

Because the technology can extract frequency features from different sensor signals — including motion, biosignals and voice — directly at the sensor edge, it reduces the burden of data transfer and subsequent computation. The ability to reconfigure a single piece of hardware for multiple signal types also makes it well suited for building smaller, more efficient low-power edge AI systems such as wearable devices and voice sensors.

"The significance lies in treating the noise generated by memristors not as mere error or instability, but as a resource for information processing," Kim said. "This neuron device is a brain-inspired technology that can be reconfigured on a single hardware platform to handle signals of varying speeds and frequencies, and it could serve as a signal-processing technology for low-power neuromorphic systems in the future."

He added that applying the technology to real products would require integrating the memristor with peripheral circuits — including amplifiers and comparators — onto a single semiconductor chip and scaling up to a larger number of neurons. "It will also be necessary to refine the circuit design and encoding techniques so the system operates stably under external noise and irregular inputs in real sensor environments," he said.

The findings were published Aug. 5 in Advanced Materials, an international journal in the field of materials science.


nbgkoo@heraldcorp.com