Light runs through the spiral like water through a coiled pipe. A chip made of silicon nitride was fabricated, with eight such spirals connected in series. Light travels 1.76 meters through them and returns after 11.63 nanoseconds. That delay is the main secret: it turns a single laser node into an entire neural network.
Previously, such delays were achieved with coils of optical fiber or digital buffers. Fiber is bulky, and digital buffers lose the phase of the light wave. The silicon spiral preserves phase and fits on a chip. It was placed in a loop: a laser shines through a modulator, light goes through the spiral, hits a detector, and the electrical signal returns to the modulator. One node unfolds into 50 virtual neurons through time multiplexing with a step of 20.48 nanoseconds.
The system operates at a feedback gain of only 0.0223, which is 40 times lower than usual. It classifies signals without errors, predicts a chaotic series with an error of 0.0082, and recognizes nine speakers of Japanese vowels with an error of 0.0898. That is twice as good as the previous speckle processor. And the phase of the light wave became an additional encoding channel: the phase difference between the spiral and a straight waveguide was used as a computational resource.
The spiral on the chip returns an echo that does not just delay but thinks. Already today, compact photonic accelerators can be designed for time-series tasks and speech recognition. The next step: reduce losses from 23.2 dB to a few decibels, and then silicon spirals will become the foundation of fast and energy-efficient neural networks.
The Archimedean spiral used in the chip is named after the ancient Greek mathematician who described it in the 3rd century BC for solving the problem of trisecting an angle. Now the same geometry compresses 1.76 meters of optical path into a few square millimeters.
The idea of delay-based reservoir computing echoes the 'liquid' neural networks from Greg Egan's novel 'Permutation City', where computational substrates use time delays to process information without traditional logic elements.