Why making artificial neurons remember their random choices boosts computer vision

The puzzle of unpredictable brains

Your brain runs on noisy signals, yet it rarely crashes. Ordinary computer chips march to an exact clock, but biological brain cells fire with a degree of chance. Importantly, a real cell retains a trace of when it last fired, so each random pulse links to the previous one. Most synthetic circuits cannot do this. Instead, an artificial stochastic spiking neuron—a synthetic brain cell that fires with chance rather than fixed rules—usually forgets everything instantly, rolling fresh dice every time. That missing link makes it hard for synthetic chips to filter out background noise.

Shuffling ions in flat crystals

To fix this, scientists built hardware using ultra-thin sheets of a crystal called CIPS, a layered sheet containing movable copper ions. Inside this crystal, tiny copper ions slide between atomic layers when voltage pushes them. The team wired a FeFET—a transistor that stores signal weights—to a synthetic neuron made of gold contacts on CIPS. When signals leave the synapse, voltage pushes jostle the copper ions. This slowly switches the crystal's internal electric alignment, building up charge until the neuron fires a pulse.

Think of it like filling a leaky bucket that has sticky walls. Each splash of water represents an incoming voltage push. The bucket drains between splashes, but sticky water clings to the sides. If another splash arrives quickly, the water level starts higher, making the bucket spill over much sooner. This picture fits the physical build-up of charge, though the crystal does not hold liquid; it shifts microscopic electric patches called domains.

Memory in the middle ground

Using a nanoscale scanning tool called PFM, which maps tiny electric patches on surfaces, the team looked inside the crystal after firing 4-volt pulses. They found that flipped patches remained visible even after twenty minutes of rest. The neuron had not wiped clean. Because of this leftover electric alignment, earlier pulses directly guided the chance and timing of future pulses. In other words, back-to-back pulses showed strong temporal correlation, meaning past pulse timings directly shaped the next ones.

This link between pulses shifted with the input voltage. At 3.8 volts, the chance of firing was just 6%, with a mean delay of 87 milliseconds. At 4.2 volts, the neuron fired rapidly with a 90% chance and a mean delay of only 7 milliseconds. But the strongest link between back-to-back events sat right in the middle at 4.0 volts. Here, firing cost about 2.8 nanojoules per pulse.

The team put these units into an SSNN, a network of brain-like units that talk using random pulses. Tasked with reading handwritten numbers from the MNIST database, the network scored 90.6% on clean pictures at 4.0 volts. When the team scrambled the images with severe static, the memory-enabled network maintained over 87% correct scores, beating a standard memory-free design.

What we still don't know

These units remain early lab tests. The team used five-millisecond pulses, which run slow compared to modern computing chips. They still do not know how far they can shrink unit sizes and pulse lengths to cut delays and power use. It is also unclear how much long-term action depends on moving copper ions rather than trapped charges or heat.