Easily Scalable Photonic Tensor Core Based on Tunable Units with Single Internal Phase Shifters

Abstract

Photonic neural networks (PNNs) show tremendous potential for artificial intelligence applications due to their higher computational rates than their traditional electronic counterpart. However, the scale‐up of PNN relies on the number of cascaded computing units, which is limited by the accumulated transmission attenuation. Here, a topology of PNN with Mach–Zehnder interferometers based on a single‐tuned phase shifter that implements arbitrary nonnegative or real‐valued matrices for vector‐matrix multiplication is proposed. Compared with the universal matrix mesh, the new configuration exhibits two orders of magnitude lower optical path loss and a twofold reduction in the number of the tunable phase shifter. An 8 × 8 reconfigurable chip is designed and fabricated, and it is experimentally verified that the 2 × 4 nonnegative‐valued matrix and the 2 × 2 real‐valued matrix are implemented in the proposed topology. Higher than 85% inference accuracies are obtained in the Modified National Institute of Standards and Technology handwritten digit recognition tasks with these matrices in the PNNs. Therefore, with much lower optical path loss and comparable computing accuracy, the proposed PNN configuration can be easily scaled up to tackle higher dimensional matrix multiplication, which is highly desired in tasks like voice and image recognition.

Multimode diffractive optical neural network

“…Potential multimode in this space is not utilized as much as possible. To realize the coupling between each single mode, a multilayer directional coupler array 4,5,18,19 is necessary; however, this coupling connection between the modes can be completed by a metaline. The previous DONN is designed based on the DAM.…”

“…16 However, with the growth of data dimension or processing depth, the overheads in footprint and the number of devices of the MZI network or MRR array increase significantly and require complex correction. 7,[17][18][19][20][21] Conversely, on-chip diffractive optical neural networks (DONN) exhibit remarkable integration capabilities. 11,[22][23][24] In our preceding studies, the length of the silicon (Si) etching slots in the DONNs is optimally designed to modulate the phase of the optical field carrying information, allowing classification, regression, and convolution computations to be actualized.…”