Photonic neural network achieves Gb/s multimedia processing

Muhammad Waqar Iqbal and colleagues from multiple institutions have unveiled a deep photonic neural network that pushes the boundaries of optical computing. Their system, detailed on arXiv (2605.30149), employs a digital micro-mirror device (DMD) to perform binary optical modulation at ultrafast rates, followed by optical scattering through a random medium and high-speed photodetection via a CMOS sensor. This forms a deep reservoir computing (RC) architecture where multiple layers are time-multiplexed, operating at Gigabit-per-second (Gb/s) processing speeds. The team achieved state-of-the-art results across diverse multimedia tasks—video recognition, image classification, and speech processing—by carefully tuning physical hyper-parameters like memory retention and dynamic response per layer.

This approach marks a significant leap for photonic AI, offering an all-optical pathway to real-time, high-throughput inference without the latency bottlenecks of electronic systems. The deep RC structure scales hierarchically, meaning larger networks can be built by cascading more photonic layers while maintaining Gb/s throughput. Key to the performance is the interplay between intra-layer dynamics (how each reservoir processes spatial/temporal features) and inter-layer coupling (how memory transfers between stages). By balancing these, the system extracts both fine-grained and long-range patterns efficiently. Potential applications range from autonomous driving sensor fusion to live video analytics and edge AI, where speed and energy efficiency are critical. The work demonstrates that photonic computing can rival electronic deep learning in real-world tasks while consuming far less power.