Silicon Retina: A Neuron Learns to See Like an Eye express
A smartphone camera captures everything indiscriminately, but the eye immediately picks out what matters. The retina doesn't just pass the image to the brain: it highlights contours on the spot, suppresses noise, and adjusts to brightness. Engineers have long tried to replicate this trick in 'smart' hardware so cameras could use less energy and make decisions faster.
The new optical neuron, a tiny semiconductor device, works like a living retinal cell. It converts light into a train of electrical spikes: the brighter the light, the more frequent the spikes. The frequency ranges from zero to two thousand times per second, and the response takes microseconds — true neural speed. For comparison, the human eye reacts to a change in the image in 30–60 milliseconds, which is tens of times slower.
The device can do what the retina does: amplify brightness contrasts at object boundaries and suppress uniform backgrounds, so-called contrastive surroundings. It distinguishes colors by mimicking cone cells and adapts to light levels so it doesn't 'blind' when stepping from shadow into sunlight. When these neurons were assembled into a spiking neural network, it learned to recognize camouflaged objects better — accuracy improved by tens of percent compared to conventional architectures.
This is a step toward cameras that see like the eye: instead of streaming gigabytes of raw video, they transmit only important events. A drone with such a retina could spot a person in the bushes without wasting energy staring at a static sky.
A frog's retina doesn't transmit stationary objects to the brain at all: it only reacts to movement. The described neuron can also be tuned to ignore static backgrounds, saving even more energy.
In Peter Watts's novel 'Blindsight', vampires have a retina capable of seeing motion with inhuman clarity — roughly the problem that a spiking camera based on these new neurons solves.