Multiobjective Evolution for Convolutional Neural Network Architecture Search

Abstrakt

The choice of an architecture is crucial for the performance of the neural network, and thus automatic methods for architecture search have been proposed to provide a data-dependent solution to this problem. In this paper, we deal with an automatic neural architecture search for convolutional neural networks. We propose a novel approach for architecture selection based on multi-objective evolutionary optimisation. Our algorithm optimises not only the performance of the network, but it controls also the size of the network, in terms of the number of network parameters. The proposed algorithm is evaluated on experiments, including MNIST and fashionMNIST classification problems. Our approach outperforms both the considered baseline architectures and the standard genetic algorithm.