A Generic Framework for Evolution of Deep Neural Networks Using Genetic Algorithms

Abstract

This paper presents an approach for the evolution of the deep neural networks (DNN) using genetic algorithms. The deep artificial neural networks are used generally for classification tasks. Depending upon the problem at hand, the designers decide on how many layers, how many number of nodes in each layers, what activation functions to be used at layers, etc. The term genetic algorithms is taken from the biological world and used in the evolution process. Here, we utilized a basic genetic algorithm concept to automatically build the optimum deep neural network suiting for the given classification task. The evolved networks are evaluated on the accuracy of classification needed, and genetic algorithm will find the best-suited DNN architecture automatically.