GACNN: Training Deep Convolutional Neural Networks with Genetic Algorithm (1909.13354v1)

Abstract: Convolutional Neural Networks (CNNs) have gained a significant attraction in the recent years due to their increasing real-world applications. Their performance is highly dependent to the network structure and the selected optimization method for tuning the network parameters. In this paper, we propose novel yet efficient methods for training convolutional neural networks. The most of current state of the art learning method for CNNs are based on Gradient decent. In contrary to the traditional CNN training methods, we propose to optimize the CNNs using methods based on Genetic Algorithms (GAs). These methods are carried out using three individual GA schemes, Steady-State, Generational, and Elitism. We present new genetic operators for crossover, mutation and also an innovative encoding paradigm of CNNs to chromosomes aiming to reduce the resulting chromosome's size by a large factor. We compare the effectiveness and scalability of our encoding with the traditional encoding. Furthermore, the performance of individual GA schemes used for training the networks were compared with each other in means of convergence rate and overall accuracy. Finally, our new encoding alongside the superior GA-based training scheme is compared to Backpropagation training with Adam optimization.