Abstract

This study presents the development of FatNet, an optical fully convolutional neural network designed to transform traditional in silico architectures into a format optimized for high-resolution optical processing. By leveraging a free-space  system, the network utilizes high-resolution feature maps and kernels without compromising frame rates. Unlike standard classifiers, FatNet integrates feature extraction and classification into a single fully convolutional framework, maximizing the parallelism of optical systems and reducing electronic-optical conversion overhead. Evaluation using the CIFAR100 dataset demonstrates that FatNet performs 8.2 times fewer convolution operations than ResNet-18 with only a marginal 6% reduction in accuracy. These results indicate that the optical implementation of FatNet achieves significantly faster inference than ResNet-18. This research provides a promising foundation for the advancement of deep learning models specifically tailored for the burgeoning era of optical computing through high-resolution kernel training.