Evolving node transfer functions in deep neural networks for pattern recognition

Abstract

Theoretical results suggest that in order to learn complicated functions that can represent high-level features in the computer vision field, one may need to use deep architectures. The popular choice among scientists and engineers for modeling deep architectures are feed-forward Deep Artificial Neural Networks. One of the latest research areas in this field is the evolution of Artificial Neural Networks: NeuroEvolution. This paper explores the effect of evolving a Node Transfer Function and its parameters, along with the evolution of connection weights and an architecture in Deep Neural Networks for Pattern Recognition problems. The results strongly indicate the importance of evolving Node Transfer Functions for shortening the time of training Deep Artificial Neural Networks using NeuroEvolution.