A manufacture-friendly design framework for optical neural networks

We introduced a Python-based design framework for manufacture-friendly optical ANN. It provides the cross-level interoperability between the photonic circuit chip layout and the neural networks infrastructure to enable the optical ANN with better tolerance to the device-by-device or chip-by-chip deviation. It allows a wide range of abstract on levels to describe the behavior of optical ANN: from the lowest-level functionality of manipulating the properties and arrangement of individual phase shifters on a photonic circuit chip, to the highest-level features of designing optical ANN via PyTorch-like development-library as well as its optimization with the well-established machine learning algorithms such as back-propagation. On all the levels, the physical design of the photonic circuit chip can be integrated and synchronized with the construction of the neural networks accounting the influences of the fabrication-deviation with the assistance of IPKISS, a Python-based tool for photonic circuit design. As a demonstration, we use our framework to design the LeNet-5 networks, which can be executed on the photonic circuit chip with non-uniformed grating coupling efficiency. Our LeNet-5 networks achieves the precision around 97.5% for MNIST task.