Joint loss function design in diffractive optical neural network classifiers for high power efficiency

The high speed, wide bandwidth, and parallel processing capabilities of a diffractive optical neural network (DONN) stimulate its applications in computer vision for image recognition and information processing tasks. This paper presents a joint loss function (J-SCE) that combines classification performance and diffractive power efficiency, thereby improving the power efficiency of the DONN classifier from 0.92% to 12.89% while maintaining a classification accuracy of 95.36%. The J-SCE function improves the overall power efficiency of the system by directing energy more effectively toward the target area. Furthermore, the J-SCE function enhances the system's robustness to noise and overall stability. This work significantly contributes to the application of DONN classifiers in practical image recognition and other information processing scenarios.