A mixed mutation strategy genetic algorithm for the effective training and design of optical neural networks

Abstract

The optical neural networks (ONNs) have received lots of attention due to their superior speed and energy efficiency in large-scale linear matrix operations compared to electrical neural networks. Undoubtedly, the training and designing weights and other hyperparameters of ONNs to enhance performance is a pivotal concern. We present a mixed mutation strategy genetic algorithm (M2SGA) to deal with the problem of inefficient training. By combining the single-point mutation (SM), the uniform mutation (UM), and the Gaussian mutation (GM) operators, our method can achieve faster convergence speed and better robustness for the parameter training of ONNs than the standard genetic algorithm (SGA). Furthermore, an N-layer feedforward optical neural network based on Mach-Zehnder interferometers (MZIs) and electro-optic modulators (EOMs) is constructed to verify the feasibility of our scheme. And, four datasets are used to further confirm the effect of our proposed scheme on the classification tasks. Additionally, the impact of alternative operators on the performance of the algorithm is evaluated. Experimental results demonstrate that M2SGA exhibits higher accuracy and lower mean squared error compared with SGA. Consequently, M2SGA holds tremendous potential for applications in object detection, image processing and other related fields thanks to its great effect in datasets classification.