.png){kind=icon}
.png){kind=icon}
Return
A survey: evolutionary deep learning
.png){kind=icon}
PRE
AI
DOI：10.1007/s00500-023-08316-4.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deep learning (DL) has made remarkable progress on various real-world tasks, but its construction pipeline strongly relies on human scientists. Furthermore, evolutionary computing (EC), as an optimization tool based on the biological evolution mechanism, has good performance on complex optimization problems. It provides a new way to construct DL models and has generated many sparks in the DL field, especially in automatic machine learning (AutoML). Although many reviews have been conducted on AutoML, in recent years, few comprehensive works have studied on the application of EC in DL, which is called evolutionary deep learning (EDL). After a thorough investigation, we think that EDL can be divided into four parts: (1) learning rule optimization, (2) hyperparameter optimization, (3) neural architecture search, and (4) other EDL-related works. In this work, we introduce the classic optimization methods and the challenges of EDL with respect to these four parts, review the related work, and then present the future research prospects. This work clearly and comprehensively reviews the concept and research content of EDL, which can help readers quickly find the intersection between EC and DL and seek their inspiration.
Keywords:
Evolutionary computing
Biological evolution
Deep learning
Automatic machine learning