Overview
- The experiment encodes digital data into the amplitude and phase of coherent light so matrix and tensor multiplications, including convolutions and attention, occur in a single pass.
- The team extends the method with multiple wavelengths, creating parallel channels that handle higher-order tensor operations.
- Computation unfolds passively during propagation, which eliminates active control and is presented as reducing power use.
- Researchers say the framework can be implemented on diverse optical platforms and plan photonic‑chip integration.
- The authors estimate integration into existing hardware platforms within three to five years, though the work remains at the experimental, peer‑reviewed stage.