An Optical Computing Chip for Executing Complex-valued Neural Network and Its On-chip Training

The optical implementation of neural networks is proposed to have advantages over electronic implementations with lower power consumption and higher computation speed. However, most optical neural networks (ONNs) utilize conventional real-valued frameworks that are designed for digital computers, forfeiting many advantages of optical computing such as efficient complex-valued operations. Complex-valued neural networks are advantageous to their real-valued counterparts by offering rich representation space, fast convergence, and strong generalizations. We propose and demonstrate an ONN that implements truly complex-valued neural networks, achieving high accuracy and strong learning capability in many benchmark tasks [1]. On the other hand, efficiently training ONNs remains a formidable challenge, due to the difficulty in obtaining gradient information from a physical device. We propose an efficient on-chip training protocol for ONNs and demonstrate it by several practical tasks [2]. The protocol is gradient-free and physical agnostic, and is applicable for various types of chip structures, especially those that cannot be analytically decomposed and characterized. The protocol is robust to experimental perturbations like imperfect phase detection and photodetection noise. Our results present a promising avenue towards deep complex networks with smaller chip size, stronger performance, and flexible reconfiguration to realistic applications (e.g., facial recognition, natural language processing, and autonomous vehicles).