Highly integrated all-optical nonlinear deep neural network for multi-thread processing
Jialong Zhang, Bo Wu, Shiji Zhang, Junwei Cheng, Yilun Wang, Hailong Zhou, Jianji Dong, Xinliang Zhang
Abstract
Optical neural networks have emerged as feasible alternatives to their electronic counterparts, offering significant benefits such as low power consumption, low latency, and high parallelism. However, the realization of ultra-compact nonlinear deep neural networks and multi-thread processing remain crucial challenges for optical computing. We present a monolithically integrated all-optical nonlinear diffractive deep neural network (AON-D2NN) chip for the first time. The all-optical nonlinear activation function is implemented using germanium microstructures, which provide low loss and are compatible with the standard silicon photonics fabrication process. Assisted by the germanium activation function, the classification accuracy is improved by 9.1% for four-classification tasks. In addition, the chip’s reconfigurability enables multi-task learning in situ via an innovative cross-training algorithm, yielding two task-specific inference results with accuracies of 95% and 96%, respectively. Furthermore, leveraging the wavelength-dependent response of the chip, the multi-thread nonlinear optical neural network is implemented for the first time, capable of handling two different tasks in parallel. The proposed AON-D2NN contains three hidden layers with a footprint of only 0.73 mm2. It can achieve ultra-low latency (172 ps), paving the path for realizing high-performance optical neural networks.