A diffractive network uses time as a computing dimension

A September study in Light: Science & Applications extends diffractive neural networks from spatial patterns into both space and time. The output is spatially filtered and read as a temporal signal by a single-pixel detector.

That targets a practical bottleneck: a fast optical transformation can still be limited by the bandwidth of the camera used to read its result. The researchers demonstrate processing of spatiotemporal lidar signals and reconstruction of moving scenes.

The approach is specialized to signals and operations that suit the encoding. Its value lies in matching optical computation to the measurement task, rather than assuming every digital workload benefits from the same arrangement.