Evolving Deep Neural Networks

Abstract

Human nervous system has evolved for over 600 million years and can accomplish a wide variety of tasks e ortlessly - telling whether a visual scene contains animals or buildings feels trivial to us, for example. For Arti cial Neural Networks (ANNs) to carry out activities like these, requires careful design of networks by experts over years of di cult research, and typically address one speci c task, such as to nd what's in a photograph, or to help diagnose a disease. Preferably for any given task, one would want an automated technique to generate the right architecture. One approach to generate these architectures is through the use of evolutionary techniques. In this work, we test three methods i.e. CGP technique, ECGP technique, and a new crossover technique of CGP, for generating CNN architectures and report the results. To our knowledge, this is the rst attempt on using either ECGP or the new crossover technique of CGP for evolving CNN architectures. This study is still in progress.