Reconfigurable Nonlinear Activation Functions by Add-Drop Resonator Mach–Zehnder Interferometers

Abstract

We propose and experimentally demonstrate a reconfigurable nonlinear activation function (NAF) unit based on add-drop resonator Mach-Zehnder interferometers (ADRMZIs) for photonic neural networks. By incorporating a tunable Mach-Zehnder interferometer (MZI) as the coupler within the ADRMZI, we successfully realized four distinct nonlinear activation functions: sigmoid, leaky ReLU, sinusoidal, and Gaussian functions, utilizing chips fabricated on a silicon photonic platform. We also assessed the performance of the NAF in an artificial neural network (ANN) trained for the MNIST handwritten digit classification task. Numerical calculations based on the measured leaky ReLU function demonstrated a prediction accuracy exceeding 97.49%, significantly surpassing the 92.41% accuracy achieved using linear operations alone. Our proposed ADRMZI-based NAF unit provides a versatile and reconfigurable solution for implementing diverse NAFs, thereby enhancing the computational capacity and flexibility of photonic neural networks for machine learning tasks.