KAIST neuron uses memristor noise to switch between speech and human-activity signals

A semiconductor circuit built to reject noise is being taught to use it instead. Researchers at KAIST have developed a programmable probabilistic neuron whose behavior changes when the resistance state of a memristor changes, allowing the same hardware to process signals that move at very different speeds.

The mechanism borrows from biology. Human neurons do not fire identically every time because small internal fluctuations affect their response. In the KAIST device, current noise from the memristor creates similar variability: changing the resistance alters the noise’s amplitude and pattern, which changes the probability that the artificial neuron will produce a spike.

That gives one circuit more than one operating range. Configured one way, it can respond to slowly changing human activity signals in the hertz range. Configured another way, it can capture fast speech signals in the kilohertz range. A kilohertz, or kHz, measures a signal repeating 1,000 times per second; hertz, or Hz, measures repetitions per second.

The team tested the system on two tasks. It encoded and classified human activity signals with 94.8% accuracy and speech signals with 95.0% accuracy. Those figures come from the researchers’ demonstrations.

So what changes in practice? A future low-power edge device—one that processes data close to a sensor rather than sending everything elsewhere—could potentially reuse the same neuromorphic hardware for different kinds of input. The technology is still at the laboratory stage, but its central move is concrete: instead of spending effort eliminating every fluctuation, designers could tune some of that noise to match the signal they need to understand.