In March 2021, researchers at UC San Diego reported a device designed to perform a neural-network activation function directly in hardware. The work explored a way to reduce the circuitry and energy involved in that part of a network.
A controlled material transition
The prototype used a heated vanadium-dioxide layer whose resistance could change gradually. The team combined activation devices with a synaptic array and demonstrated image edge detection. The study appeared in Nature Nanotechnology.
A proof of concept
The researchers described a small experimental system, not a ready replacement for general-purpose AI hardware. More layers and more demanding tasks would require further development. The archive preserves the distinction between a promising device-level result and the engineering needed for a complete, scalable computing platform.
Research paper — Nature Nanotechnology ↗