Abstract: In recent years, both industry and academia have made significant advances in deep learning (DL). However, configuring the hyperparameters of deep models typically requires significant computational overhead and expert knowledge. To overcome these aforementioned challenges, evolutionary computation (EC), as an efficient heuristic search, has demonstrated significant advantages in the automated configuration of DL models, i.e., evolutionary DL (EDL). We describe EDL from the perspective of automated machine learning. Particularly, we first depict the concept of EDL from EC and DL perspectives and regard EDL as an optimization problem. Consequently, we systematically introduce data preparation, model generation, and model deployment from the DL lifecycle. In addition, we analyze and discuss the solution representation and search paradigms. Finally, we provide applications, open issues, and potential research directions related to EDL. This study reviews the advancements in EDL and offers insightful guidelines for its development.