基于光学可调非线性函数的目标分类

Target Recognition Based on Optically Tunable Nonlinear Functions

摘要: 非线性函数在神经网络中扮演着重要的角色，通过对神经元引入非线性因素，从而使得神经网络得以逼近任何非线性函数，但随着网络本身规模的增加也导致训练时间和算力成本过高。尽管目前已经提出基于电子计算的优化方法，却仍存在可拓展性差、计算自由度单一等问题。因此，文章提出了一种利用大量可编程光学神经元的光子计算作为解决方法，以高鲁棒性、高精确度方法进行Tanh和Sigmoid非线性计算。实验结果表明，所提出的基于衍射所构建的非线性计算模块在神经网络的应用方面具有相对较高的泛化能力。对于5 × 5数据输入维度，Sigmoid函数实现了9.66 × 10⁻⁴的误差损失，Tanh函数实现了1.02 × 10⁻³的损失。在MNIST手写数字分类任务上，通过实验验证了架构在面对传输中高斯噪声等现实因素干扰时具有相当的稳定性和鲁棒性。

Abstract: Nonlinear functions play a crucial role in neural networks by introducing nonlinearity into neurons, enabling the network to approximate arbitrary nonlinear mappings. However, as the scale of neural networks continues to grow, the associated training time and computational costs have become increasingly prohibitive. Although several optimization strategies based on electronic computing have been proposed, they still suffer from limited scalability and restricted computational flexibility. To address these challenges, this paper proposes a photonic computing approach based on a large number of programmable optical neurons to perform nonlinear computations with high robustness and precision. Specifically, nonlinear functions such as Tanh and Sigmoid are implemented using a diffractive optical structure. Experimental results demonstrate that the proposed diffraction-based nonlinear computing module exhibits strong generalization capabilities when applied to neural network tasks. For a 5 × 5 input dimension, the Sigmoid function achieved a loss of 9.66 × 10⁻⁴, while the Tanh function yielded a loss of 1.02 × 10⁻³. Furthermore, evaluation on the MNIST handwritten digit classification task shows that the proposed architecture maintains considerable stability and robustness in the presence of practical disturbances such as Gaussian noise during transmission.