Multi-objective Evolution for Deep Neural Network Architecture Search

Petra Vidnerová, Roman Neruda

Abstract: In this paper, we propose a multi-objective evolutionary algorithm for automatic deep neural architecture search. The algorithm optimizes the performance of the model together with the number of network parameters. This allows exploring architectures that are both successful and compact. We test the proposed solution on several image classification data sets including MNIST, fashionMNIST and CIFAR-10, and we consider deep architectures including convolutional and fully connected networks. The effects of using two different versions of multi-objective selections are also examined in the paper. Our approach outperforms both the considered baseline architectures and the standard genetic algorithm used in our previous work.