Deep neural networks (DNNs) have shown remarkable performance in solving various real-world problems. In principle, the achievements of DNNs are mainly contributed by their deep architectures, which can learn meaningful representations at different levels. The learned representations can greatly enhance the performance of the subsequent machine learning algorithms. However, designing an optimal deep architecture for a particular problem requires rich domain knowledge of both the investigated problem domain and for DNNs, which is not necessarily held by every end-user.
Neural Architecture Search (NAS), as an emerging technique to automate the design of DNN architectures without or with little domain expertise, is a potentially promising technique to solve the aforementioned critical issue. However, NAS is a complex optimization problem challenged with discrete, bi-level, and computation expensive characteristics. Currently, reinforcement learning, gradient-based algorithms, and evolutionary computation (EC) are the three main techniques for NAS. However, the reinforcement learning-based NAS algorithms have typically a high computational complexity. Although gradient-based NAS can perform efficiently, it often relaxes the original NAS problem into a continuous optimization problem without any convincing justification. Furthermore, neither of them can fully automate the NAS process in practice.
EC-based NAS (ENAS) algorithms can solve the NAS problems in their natural form. ENAS algorithms can also automate the design of DNN architectures. In the literature, EC has been successfully applied to both architecture search and weight optimization of shallow and medium-scale neural networks for over 30 years. Unfortunately, these works cannot be properly scaled to DNNs, which typically have a large number of architectural parameters and weights. Although a number of ENAS algorithms have been published in top-tier journals and conferences in recent years, there is still much room to make improvements in addressing NAS problems.
This special issue is particularly focused on novel theories, algorithms and applications on ENAS. The topics of interest include but are not limited to:
Manuscripts should be prepared according to the “Information for Authors” section at https://cis.ieee.org/publications/t-evolutionary-computation/tevc-information-for-authors and submissions should be made through the journal submission website at https://ieee.atyponrex.com/journal/TEVC-IEEE by selecting the Manuscript Type “ENAS” and clearly adding “ENAS Special Issue Paper” to the comments to the Editor-in-Chief. Submission of a manuscript implies that it is the authors' original unpublished work and is not being submitted for possible publication elsewhere..