Probabilistic grammar-based deep neuroevolution

Abstract

Designing deep neural networks by human engineers can be challenging because there are various choices of deep neural network structures. We developed a deep neuroevolution system to automatically design deep neural network structures using deep neuroevolution. Our approach defines a set of structures using a probabilistic grammar and searches for best network structures using Probabilistic Model Building Genetic Programming. Our approach takes advantage of the probabilistic dependencies found among the structures of networks. The system was applied to tackle the problem of the physiological signal classification of abnormal heart rhythm. In the classification problem, our discovered model is more accurate than AlexNet. Our discovered model uses about 2% of the total amount of parameters of AlexNet.