A Diode Chip That Sees Images ⚡ экспресс
A neuromorphic architecture based on resonant tunneling diodes (RTDs) has been created for real-time AI on low-power devices. Their nonlinear properties are ideal for physical reservoir computing—a processing method that uses the dynamics of the medium instead of random connections. A numerical implementation successfully recognized handwritten digits and objects (Fruit 360), demonstrating the promise of this approach. This is a step toward deterministic hardware AI, where signals are transformed predictably, like a river flowing along a set channel.
A drumroll on water creates a unique ripple — from it you can guess the rhythm. That’s how a reservoir computer works: instead of water, it uses a diode invented by Leo Esaki in 1958. Under voltage, its current behaves nonlinearly: first it rises, then drops, and rises again. The signal gets tangled like waves on water, and each image produces its own electrical pattern. Unlike regular neural networks, connections aren’t tuned here — the diode’s physics itself generates complexity. Only the output layer is trained, mapping the pattern to the answer. A circuit of such diodes reads handwritten digits and tells apples from oranges as well as powerful graphics cards, but uses hundreds of times less energy. The secret lies in the growth of entropy: the signal gets mixed into a nearly random yet distinctive response. By analyzing it as a spectrum of light, the chip extracts key features. The diode switches in picoseconds — a thousand times faster than transistors, because electrons tunnel through energy barriers like ghosts.
🎯 Electrons in this diode pass through energy barriers like ghosts through walls. That’s why it switches in picoseconds — faster than any transistor.