Weight Initialization for Complex Number Convolutional Neural Networks Using Genetic Algorithms

Abstract

Weight initialization using Glorot method is one of the effective methods for convolutional neural networks (CNN). In addition, previous studies have also used this method for complex number CNNs. However, there is no research to confirm that Glorot weight initialization method has the best effect for complex number CNN. This study focuses on optimizing complex number CNNs based on how to initialize connection weights of neurons. This paper proposes to use genetic algorithm (GA) to find optimal parameters for Rayleigh distribution to initialize weights before training complex number CNN. Experiments show that GA has found better parameters when using the Glorot method as in previous studies.