A Novel Approach for Optimization of Convolution Neural Network with Particle Swarm Optimization and Genetic Algorithm for Face Recognition

Abstract

Convolutional neural networks are contemporary deep learning models that are employed for many various applications. In general, the filter size, number of filters, number of convolutional layers, number of fully connected layers, activation function and learning rate are some of the hyperparameters that significantly determine how well a CNN performs.. Generally, these hyperparameters are selected manually and varied for each CNN model depending on the application and dataset. During optimization, CNN could get stuck in local minima. To overcome this, metaheuristic algorithms are used for optimization. In this work, the CNN structure is first constructed with randomly chosen hyperparameters and these parameters are optimized using Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) algorithm. A CNN with optimized hyperparameters is used for face recognition. CNNs optimized with these algorithms use RMSprop optimizer instead of stochastic gradient descent. This RMSprop optimizer helps the CNN reach global minimum quickly. It has been observed that optimizing with GA and PSO improves the performance of CNNs. It also reduces the time it takes for the CNN to reach the global minimum.