可变光学环境下衍射神经网络的仿真模型研究

Research on Simulation Model of Diffractive Neural Network in Variable Optical Environments

摘要: 光学衍射神经网络(optical diffractive neural network, ODNN)已被证明可以很好地完成机器学习的推理任务。然而目前大部分ODNN的研究主要集中在大气环境下，其他空间环境下的ODNN性能讨论甚少。本文基于瑞利–索末菲衍射理论，研究分析了不同光学环境下ODNN的训练及测试效果。在532 nm的可见光工作波段下，光学衍射神经网络分别在空气和水环境中训练，在其能准确执行推理能力的情况下，将网络模型在仿真阶段置于新的空间环境测试，研究网络性能的变化趋势，分析ODNN对的测试环境适应性。本文仿真结果表明，光学衍射神经网络对测试环境的变化响应敏感，网络的推理能力随着测试环境与训练条件的偏差增大而下降，直至消失。以此为切入点，为光学衍射神经网络增加一个额外的自由度，有望在ODNN的多任务等方面提供新的研究方向。

Abstract: The optical diffractive neural network (ODNN) has been demonstrated to perform well in machine learning inference tasks. However, most of the current research on ODNN is focused on atmospheric environments, with limited discussion on its performance in other spatial environments. This paper investigates and analyzes the training and testing effects of ODNN in different optical environments based on the Rayleigh-Sommerfeld diffraction theory. ODNN models are trained in both air and water environments at the visible light wavelength of 532 nm. After demonstrating their accurate inference capability, the network models are tested in new spatial environments during simulation to study the trends in network performance and analyze the adaptability of ODNN to testing environments. The simulation results indicate that ODNN is sensitive to changes in the testing environment, and the inference capability of the network decreases as the deviation between the testing environment and training conditions increases, eventually disappearing. This observation provides a starting point to introduce an additional degree of freedom to ODNN, which may offer new research directions in areas such as multi-tasking.