High accuracy inference by an optical neural network implementation

In recent years, sensing and imaging have significantly progressed due to AI algorithms such as Neural Network (NN). The main issues of applying NNs to information processing are the limited processing speed and high energy consumption of electronic processors. Optical Neural Network (ONN), which utilizes diffraction and propagation of light for processing, is an intriguing implementation of an ultra-fast and low-energy-consuming NN. However, previous studies of ONN are mainly on simulations due to the experimental difficulty of processing more than hundreds of input data. In hardware implementations, the performance or the classification accuracy of ONNs can be reduced by the noise and the displacements. Therefore, not only must the ONN achieve high theoretical accuracy, but it must also be robust to these experimental errors. In this study, the classification of 1,000 MNIST input data (100 data for each of 10 classes) was realized experimentally as well as theoretically, taking advantage of our novel setup with a variable spatial light modulator (SLM). With our experimental configuration, we investigated the classification accuracy with several loss functions for the ONN training. The inference accuracy of the MNIST classification task was up to 97% in the simulation and ~95% in the experiment by softmax-cross-entropy (SCE) loss function. Also, the classification accuracy of 98% for a Surface crack classification and 93% for a Pollen classification was achieved experimentally. These results show that SCE can realize high-accuracy classification in the ONN implementation. Our results revealed the high application capability of the optical neural network for practical sensing tasks.