Multi-Functional Low-Threshold Nonlinear Activations for Photonic Neural Networks

In recent years, with the advancement of optical computing technology, optoelectronic activation functions have emerged as a critical component for high-performance optical computing systems. Here, we propose an all-analogue opto-electronic nonlinear activation unit. Through the continuous adjustment of the bias voltage and the reference light intensity, it is feasible to implement multiple activation functions, including Sigmoid, Exponential Linear Unit (ELU), leaky-ELU, Inverse ELU, log-log, and other atypical functions. Embedding these nonlinear activation functions into a convolutional neural network for Modified National Institute of Standards and Technology (MNIST) hand-written digit-classification and Fashion-MNIST tasks, we can achieve 99.21% and 89.47% high inference accuracy respectively. This continuous programmable unit features a simple structure, incurs low costs, and is highly amenable to integration. It can operate at an extremely low optical power of -24 dBm, making it suitable in low-light conditions. The device is well-suited for being deployed as a nonlinear unit in the photonic neural network at the hardware level, effectively establishing the technical groundwork for the future implementation of large-scale photonic neural networks.