Published in iScience, researchers from Changchun Institute of Optics, Fine Mechanics and Physics at the Chinese Academy of Sciences developed an all-optical neural network using multi-channel fiber arrays to address the high energy costs and slow training times of traditional AI hardware. The study demonstrated how photonic systems could complement electronic computing for specific tasks.

The team replaced free-space light propagation in conventional optical neural networks with fiber-based transmission, creating a more compact and robust design. The system segmented input images using a 7×4 fiber array, processed features through attenuators and nonlinear couplers, and classified targets with 95% accuracy. At 1550 nm wavelength, the network achieved its highest performance, with environmental noise affecting results by less than 1%.

Key to the design was its hardware-driven approach, eliminating the need for extensive electronic training. The fiber array acted as a convolutional layer, while nonlinear couplers enabled dimensionality reduction akin to pooling in traditional neural networks. Experimental results showed distinct outputs for horizontal lines, vertical lines, and blocks, validating the system’s classification capability.

This work highlights the potential of optical computing to reduce energy consumption in AI applications. The fiber-based architecture could be integrated into edge devices or specialized processors where low latency and power efficiency are critical. Future research may explore programmable attenuators and on-chip miniaturization to expand its versatility.