Abstract

 Deep Learning has had a huge impact on a wide variety of computational tasks and it has become one of the major pillars of modern Artificial Intelligence. This paper presents a comprehensive and consistent organizational structure on recent developments in the area of deep learning, including algorithm enhancements and proof-of-concept implementations. The survey was conducted through a methodological literature review of scientific databases such as ACM Digital Library, Scopus, Web of Science, IEEE Xplore, and arXiv using a keyword menu designed to identify the most common core applications, architectures, and cutting-edge research fronts. We synthesized the development of leading popular deep learning libraries linking to models that are well known such as CNN, RNN, to the present day of more advanced and sophisticated models such as transformers, graph neural networks, and diffusion-based generating models. The use of various applications in different domains has been reviewed and their performance, limitations, and interdisciplinary impacts have been detailed. There are many different persistent challenges persistent (such as lack of data, explainability, too much computational power required to run AI models, ethics, and adversarial robustness) that are pointing to key research gaps in the field that still exist. At the conclusion of this study there will be multiple examples of future work in this field (explainable AI, resource-efficient modelling, learning from multiple types of data, collaboration between humans & AI, and creating general-purpose intelligent systems). By reviewing all current literature thoroughly, this study aims to serve as a comprehensive, accurate, up-to-date reference for researchers, practitioners and policy-makers who either study the current state or predict future states of deep learning.

Statistics