Hyperparameters Optimization of Convolutional Neural Networks using Evolutionary Algorithms

This paper presents an approach for tuning hyperparameters in Convolutional Neural Networks (CNNs) by adopting evolutionary algorithms, i.e., Genetic Algorithms (GAs). CNNs include abounding hyperparameters that must be adjusted cautiously to accomplish the highest classification accuracy in case of image classification tasks or other related tasks. GAs are used to effectively go across the enormous search space of CNNs hyperparameters and choose the finest CNN architecture that operates well on a given task. CNN architecture contains a sequence of several convolutional layers followed by some fully-connected layers, as the number of layers increases, the number of corresponding hyperparameters will exponentially increase, examples of hyperparameters include but not limited to: number of convolutional filters in each convolutional layer, the number of nodes in the fully-connected layers, and the type of activation function. The proposed tuning framework was tested and evaluated using the SVHN dataset for digit classification of printed digits cropped from pictures of house number plates. The proposed approach can generate a CNN architecture with a validation accuracy of 92.31%.