QDRelu: An Activation Function Based On Quantum-Dot Spin-VCSELs

Activation function has become an indispensable part of artificial neural networks (ANNs). Nowadays, the realization of activation function mainly depends on electronic instrument, which need a lot of computing resources and power consumption. In this paper, we research quantum dot spin-polarized vertical-cavity surface-emitting laser and investigate the relationship between the total output intensity and normalized pump intensity through numerical simulation. Then QDRelu function based on the relationship is presented to realize the behavior of Relu function. At the same time, CIFAR10-based and MNIST-based recognition tasks are carried out to verify the feasibility of QDRelu function with respect to traditional Relu function, these tasks are implemented in several typical artificial neural network models. The results indicate that QDRelu function not only takes advantage of the merits of photonic technology such as high speed, low power consumption and high parallelism but also ensures accuracy of whole artificial neural network.