One of the pressing issues for optical neural networks (ONNs) is the performance degradation introduced by parameter uncertainties in practical optical components. Hereby, we propose a novel two-step ex situ training scheme to configure phase shifts in a Mach–Zehnder-interferometer-based feedforward ONN, where a stochastic gradient descent algorithm followed by a genetic algorithm considering four types of practical imprecisions is employed. By doing so, the learning process features fast convergence and high computational efficiency, and the trained ONN is robust to varying degrees and types of imprecisions. We investigate the effectiveness of our scheme by using practical machine learning tasks including Iris and MNIST classifications, showing more than 23% accuracy improvement after training and accuracy (90.8% in an imprecise ONN with three hidden layers and 224 tunable thermal-optic phase shifters) comparable to the ideal one (92.0%).

中文翻译：

---

使用遗传算法的广义鲁棒训练方案，用于具有不精确分量的光学神经网络

光学神经网络 (ONN) 面临的紧迫问题之一是实际光学元件中参数不确定性导致的性能下降。因此，我们提出了一种新颖的两步非原位训练方案来配置基于马赫-曾德干涉仪的前馈 ONN 中的相移，其中采用了随机梯度下降算法和考虑四种实际不精确度的遗传算法。通过这样做，学习过程具有快速收敛和高计算效率的特点，并且训练后的 ONN 对不同程度和类型的不精确性具有鲁棒性。我们通过使用包括Iris在内的实际机器学习任务来调查我们方案的有效性和 MNIST 分类，训练后的准确率提高了 23% 以上，准确率（在具有三个隐藏层和 224 个可调谐热光移相器的不精确 ONN 中为 90.8%）与理想的（92.0%）相当。