Evolving convolutional autoencoders using multi-objective Particle Swarm Optimization

Convolution autoencoders have shown that they have useful applications in multiple areas related to image classification and generation. The architecture of deep neural networks has a huge impact on the performance of the network. In recent years, metaheuristics have become quite popular for the optimization of the architectures of deep neural networks. In this paper, we propose a multi-objective Particle Swarm Optimization (PSO) for designing flexible convolution autoencoders by evolving the arrangement of convolution and pooling layers along with the number of parameters. A novel encoding strategy for PSO particles is introduced, which allows flexible positioning of pooling layers. To enhance the exploration of the search space, velocity and the position update methods for variable-length particles in PSO are also modified. The proposed method is evaluated on different datasets for image classification. The experimental results show that the resultant architectures are more generalized, optimized, and accurate than their competitors.