Hybrid particle swarm training for convolution neural network (CNN)

Abstract

Convolutional Neural Networks(CNN) are one of the most used neural networks in the present time. Its applications are extremely varied. Most recently they have been proving helpful with deep learning, as well. Since it is growing in more convoluted domains, its training complexity is also increasing. To tackle this problem, many hybrid algorithms have been implemented. In this paper, Particle Swarm Optimization (PSO) is used to reduce the overall complexity of the algorithm. The hybrid of PSO used with CNN decreases the required number of epochs for training and the dependency on GPU system. The algorithm so designed is capable of achieving 3–4% increase in accuracy with lesser number of epochs. The advantage of which is decreased hardware requirements for training of CNNs. The hybrid training algorithm is also capable of overcoming the local minima problem of the regular backpropagation training methodology.