Plenary -Evolutionary Computation for Automated Design of Deep Neural Networks

Bing Xue; Victoria University, Wellington, NZ; ABSTRACT: Deep neural networks (DNNs) have achieved great success in a wide range of challenging tasks in recent years, such as image classification and natural language processing. The deep architectures in DNNs play a crucial role in their performance, via learning meaningful features directly from the raw data without explicit feature engineering. However, many powerful DNN architectures are manually designed, which requires rich expertise and experience in both the DNN and the target problem domains, but they are often not available to interested users in reality. Neural architecture search (NAS) can address this issue by automatically designing DNN architectures, where evolutionary computation (EC) based methods (i.e. evolutionary NAS, ENAS) have recently gained much attention and success. In this talk, I will mainly take image classification as the application area to discuss how EC methods, such as genetic algorithms and particle swarm optimisation, can be used to achieve NAS, and the key components in designing an ENAS algorithm, such as individual representation, fitness function and search mechanism. State-of-the-art ENAS algorithms will be reviewed and discussed in terms of how they achieve improved performance in terms of the classification accuracy, model complexity, computational efficiency, etc . Finally, the talk will also cover current challenging issues and open research questions in this area.