The dial that stores the state also sets the noise

One variable, two jobs

The device at the centre of the KAIST announcement is a memristor, which holds a resistance state after an electrical pulse sets it. The claim that makes it a signal-processing part is narrower than the word neuron suggests: when the resistance state is changed, the magnitude and behaviour of the current noise change with it, and presetting that state alters the probability of spike generation and the response range under the same input.^[1]

Read as a schematic, that puts one physical variable on two duties at once. The stored resistance is the memory, and the same stored resistance is the knob that sets how noisy the cell is, which is what decides the rate at which it fires. The frequency selectivity follows from there rather than from any added filtering stage: by changing only the resistance state, the group reports, the same circuit can be switched to respond to slow human activity signals in the hertz range or to speech in the kilohertz range. A quieter reading is also available, in which part of the selectivity comes from the read-out electronics around the cell rather than from the cell itself, and the announcement does not carry the measurements that would separate the two.^[1]

What the span costs

Two numbers travel with the work: 94.8 per cent accuracy on human activity recognition and 95.0 per cent on speech recognition. Both are endpoint scores on separate tasks, and the announcement puts no comparison device beside them and no energy figure under them. For a part offered as a route to low-power edge neuromorphic systems, energy is the quantity the pitch rests on, and the two accuracies leave it untouched.^[1]

The number worth waiting for is a different one. Hertz and kilohertz sit three orders of magnitude apart, and the design's whole economy depends on how many distinct resistance settings it takes to walk that distance and how tightly each has to be held. If a follow-up paper reports the count of programmed states used across the hertz-to-kilohertz span together with the drift in each state's noise level over repeated cycling, the reconfiguration claim becomes something another laboratory can price; until then the two accuracies show that each endpoint works, and say nothing about the road between them.^[1]