KAIST

< Members of the research team. From left: Dae Hee Kim (Ph.D. candidate), Tae Wook Ko (Ph.D. candidate), Dr. Do Hoon Kim (first author), Seo Eun Jang (Master >

In electronic devices, irregular fluctuations in signals are generally referred to as “noise.” Because noise interferes with accurate information processing, conventional semiconductor technology has mainly treated it as something to be reduced or eliminated. However, neurons in the human brain do not respond in exactly the same way every time, even to the same stimulus. Tiny internal variations in neurons change when and how often neurons are fired, and this probabilistic operation is one of the brain’s key information-processing features. Inspired by this, KAIST researchers have developed a next-generation semiconductor technology that does not remove current noise generated in memristors, but instead tunes it to a desired level and uses it to process different types of signals.

KAIST (President Choongsik Bae) announced on the August 16 that a research team led by Professor Kyung Min Kim from the Department of Materials Science and Engineering has developed a new neuromorphic neuron technology that uses noise generated in semiconductor devices for information processing, enabling selective encoding of time-series signals across different frequency bands.
※ Neuromorphic technology: A technology that processes information by mimicking the way the human brain and neurons operate.

In general, noise generated in semiconductors is regarded as an obstacle to accurate signal processing. For this reason, most electronic devices are designed to reduce or eliminate noise as much as possible. The human brain, however, works differently. Neurons, the nerve cells of the brain, do not always respond in the same way to the same stimulus because of internal probabilistic fluctuations. This irregularity actually helps the brain flexibly respond to a wide range of situations and sensory signals.

< Figure 1. Comparison of biological neurons and memristor-based probabilistic neurons >

The research team used a memristor in this study. A memristor is a semiconductor device that changes its resistance state in response to electrical stimulation and remembers that state. Until now, current noise generated in memristors has mainly been used for random number generation, which creates unpredictable numbers, or for probabilistic computing.

However, previous studies have largely focused on using the inherent randomness of memristors as it is. Technologies that can tune probabilistic response characteristics according to need had not been sufficiently realized.

< Figure 2. Memristor noise and spike-generation characteristics dependent on the resistance state >

The key insight of this study is that when the resistance state of a memristor is changed, the magnitude and behavior of its current noise also change. By presetting the resistance state of the memristor, the probability of spike generation and the response range can vary even under the same input. Using this principle, the research team implemented a “programmable probabilistic neuron (PPN)” that treats noise not simply as instability, but as an information-processing resource that can be tuned in a desired way.

This neuron can be configured to respond differently depending on how rapidly an input signal changes, in other words, its frequency. By changing only the resistance state of the memristor, the same circuit can be switched to respond sensitively to slow human activity signals in the hertz (Hz) range or fast speech signals in the kilohertz (kHz) range. Hz and kHz are units that indicate how many times a signal repeats per second, with 1 kHz equal to 1,000 Hz.

< Figure 3. Reconfigurable time-series encoding for human activity and speech signals >

In simple terms, a single artificial neuron can be reconfigured according to the speed of the signal it needs to process. When processing slowly changing signals such as human movement, it can operate in a way suited to slow variations; when processing rapidly changing signals such as speech, it can be adjusted to capture short and fast changes effectively.

The research team verified the technology using signals with different frequency ranges. The system encoded and classified human activity signals in the Hz range and speech signals in the kHz range, achieving accuracies of 94.8% in human activity recognition and 95.0% in speech recognition.

Professor Kyung Min Kim said, “The significance of this study lies in demonstrating that memristor noise can be harnessed as a tunable information-processing resource, rather than simply treated as an error or instability,” adding, “Because the same hardware can be reconfigured for signals of different speeds and frequencies, it could be used as a signal-processing technology for future low-power edge neuromorphic systems.”

This study was led by Dr. Do Hoon Kim from the Department of Materials Science and Engineering as first author, and was published in the internationally renowned materials science journal Advanced Materials on August 05.

Paper title: Noise-Tunable Memristor Enabling Programmable Probabilistic Neurons for Frequency-Selective Time-Series Signal Encoding,
DOI: https://doi.org/10.1002/adma.74529

This research was supported by the Basic Research Program in Science and Engineering and the PIM Artificial Intelligence Semiconductor Core Technology Development Program of the Ministry of Science and ICT and the National Research Foundation of Korea.