A Realization of Activation Function Based on Inverse Design

The realization of nonlinear activation functions by optical means, as an important component of optical neural networks, has been a hot spot. In this work, we use inverse design tools to design a compact patterned nonlinear activation function generator in silicon-carbide-on-insulator and our nonlinear activation function achieves an outstanding performance in MNIST task (98.86%) and relatively low energy loss (40.21%).