Swarm Intelligence Based Ensemble Learning of Deep Neural Networks

Ensemble learning is a powerful tool in machine learning and it is very significant to utilize the mechanism of ensemble learning to improve the performance of deep learning for recognition and classification tasks. In this paper, we propose a general ensemble learning framework of deep neural networks based on swarm intelligence algorithms for solving the combination coefficients to the outputs of these component deep neural networks. We formulate the weights assigning problem for the deep neural networks as an optimization problem whose objective function is highly complicated and use swarm intelligence algorithms to solve it. We evaluate this ensemble learning framework on two real-world datasets, Market-1501 for person re-identification and CIFAR for image classification, and show that it outperforms a single deep neural network remarkably.