Designing Deep Neural Network Topologies with Population-Based Metaheuristics


Over last years the deep neural network has become one of the most popular classification methods with performance comparable and in some cases even superior to humans in the wide range of applications. However, there are still some major challenges regarding the deep neural networks. One of the biggest, with the huge impact on the classification performance, is the design of such deep neural network. In this paper, we propose a population-based metaheuristics approach for designing a deep neural network topology in a straightforward automatic manner, which performance we compare against the conventional classifiers across three different datasets. With the usage of our proposed method, unlike the conventional classifiers, we were able to achieve high classification performance with no major performance drops throughout all tested datasets.

Proceedings of the Central European Conference on Information and IntelligentSystems