上海科技大学知识管理系统

In this paper, a scheme for realizing low-threshold dual-polarization electro-optic nonlinear activation functions is proposed, designed, and experimentally demonstrated. This scheme relies on a structure composed of an optical power splitter and a thermo-optic modulator based on Mach-Zehnder interferometer, which are optimized by using direct binary search and particle swarm optimization algorithms. When a fixed proportion of the input optical signal is transformed into the electrical signal and the remaining input optical signal is modulated by using the thermo-optic effect, the rectified linear unit activation function with a low-threshold power of 0.2 mW for transverse-electric polarization and the Gaussian activation function with a low-threshold power of 0.1 mW for transverse-magnetic polarization can be measured separately. If the above-mentioned two nonlinear activation functions are introduced into the convolutional neural network to perform the modified National Institute of Standards and Technology handwritten digit classification task, validation accuracies of 97.3% and 96.85% will be achieved.