Convolutional neural networks suffer performance degradation when deployed on analog optical computing systems due to analog noise. To address this issue, a hybrid training approach is used, where forward propagation is performed by an optical convolution processor and back-propagation by a digital electronic computer. Networks trained with this method achieve inference accuracies on MNIST, Fashion-MNIST, and CIFAR-10 comparable to fully digital training, effectively mitigating the impact of analog computation noise.