Multi-Wavelength Diffractive Photonic Neural Network for Multi-Task Learning

Photonic neural networks use photons instead of electrons to perform artificial intelligence (AI) tasks, with the advantage of high-speed, low-power information processing. However, existing architectures are designed for a single task, which cannot reuse multiple tasks in parallel in a single system because the competition among different tasks will degrade the model performance. In this paper, a novel optical multi-task learning system is proposed by designing a multi-wavelength diffractive photonic neural network (DPNN) using a joint optimization method. By encoding the input of multiple tasks into multi-wavelength channels, the system can significantly reduce the competition to execute multi-tasks in parallel with high precision. We design the two-task and four-task DPNNs with two and four spectral channels respectively, for classifying different inputs from the EMNIST, KMNIST, FMNIST, and MNIST databases. Numerical evaluations show that for multitask learning, multi-wavelength DPNNs achieve significantly higher classification accuracies than single-wavelength DPNNs under the same network size. Moreover, as the network size increases, the classification accuracy of multi-wavelength DPNNs is comparable to that of individually training multiple single-wavelength DPNNs to perform multiple tasks separately. Our work provides a proven technical solution for developing high-throughput neuromorphic photonic computing and more general artificial intelligence systems to perform multiple tasks in parallel.