A Survey on Evolutionary Construction of Deep Neural Networks

Automated construction of deep neural networks (DNNs) has become a research hot spot nowadays because DNN’s performance is heavily influenced by its architecture and parameters, which are highly task-dependent, but it notoriously difficult to find the most appropriate DNN in terms parameters best solve given task. In this work, we provide an insight into automated process formulating multilevel multiobjective large-scale optimization problem with constraints, where nonconvex, nondifferentiable, black-box nature make evolutionary algorithms (EAs) stand out as promising solver. Then, give systematical review existing techniques from different aspects analyze pros cons using EA-based methods each aspect. This work aims help researchers better understand why, where, how utilize EAs for meanwhile, EA task so that they may focus more on EA-favored scenarios devise effective techniques.