Deep learning (DL), as a subfield of artificial intelligence (AI) and machine learning (ML), is the fastest-growing trend in data analysis and is regarded as a breakthrough. Over the past few years, there has been an ongoing shift toward using DL methods in different applications, mainly due to the increasing data accessibility and computational processing power. DL models characterized by neural networks are learning methods with multiple levels of representation that learn the semantic and discriminative features in a sequential bottom-to-up manner from the data. They are composed of several levels of non-linear modules that each modify the representation at a lower level into a higher or slightly more abstract level. As such, very complex functions can be learned without depending on human-crafted features.
DL has been used in several research fields, such as speech recognition, stereo vision, medical image recognition, remote sensing, time-series analysis, biomedicine, agriculture, and geosciences. One of the limiting factors of using DL models is that they require significant amounts of training samples compared to conventional ML methods To date, several DL architectures have been introduced, of which the stacked autoencoder, convolutional neural network, generative adversarial network, deep belief network, and recurrent neural network have become mainstream. DL techniques have had significant successes in several fields, which have been widely accepted as challenges in recent decades. Moreover, by growing big data and their applications in practical productions and developed time-efficient networks or public online free or commercial cloud computing platforms, such as Google, Amazon, Microsoft, and IBM, much more attention will be paid to develop new DL networks for the practical projects.