Multilayer nonlinear diffraction neural networks with programmable and fast ReLU activation function
Yu Ning, Qian Ma, Qiang Xiao, Xinxin Gao, Qian Wen Wu, Ze Gu, Rui Li, Long Chen, Jian Wei You, Tie Jun Cui
Abstract
Optical diffractive neural networks are emerging for improving speed and energy efficiency in machine learning. However, the challenges of nonlinear activation functions (e.g., latency issues, high power consumption, and cascading complexity) impede their performance and practical deployment. Here, we propose a programmable multilayer full-space nonlinear neural network operating in the microwave frequency band. Its nonlinear layers are constructed using programmable metasurfaces integrated with RF components, implementing a ReLU-like activation function. The nonlinear architecture achieves a nanosecond-scale delay (17.7 ns), representing orders of magnitude improvement in speed over photoelectric conversion-based nonlinearities. Moreover, the nonlinearity is characterized by exceedingly low thresholds and reconfigurable nonlinear activation functions. The system demonstrates remarkable classification capability in image classification and real-time human posture recognition tasks. Characterized by low latency, high speed, low power consumption, and flexible nonlinear activation, this architecture holds great promise for applications in security screening, medical rehabilitation, human-computer interaction, and numerous other fields. To address the challenge of high-speed nonlinear activation in diffractive neural networks, the authors introduce a programmable, low-power, and fast nonlinear diffractive network that exhibits an activation latency of 17.7 ns, enabling tasks like real-time human posture recognition.