[Paper Review] Deep Learning -- A first Meta-Survey of selected Reviews across Scientific Disciplines, their Commonalities, Challenges and Research Impact

Jan Egger, Antonio Pepe+4 more|arXiv (Cornell University)|Nov 16, 2020

COVID-19 diagnosis using AI115 references34 citations

TL;DR

This meta-survey synthesizes 61 deep learning review papers across computer vision, natural language processing, medical informatics, and interdisciplinary domains to map the field's evolution, common architectures, challenges, and research impact. It reveals rapid growth in publications—over 11,000 PubMed results for 'deep learning' by Q3 2020, with 90% from the prior three years—and identifies key trends, limitations, and future directions across subfields, offering a high-level, categorized overview of deep learning's scientific and practical influence.

ABSTRACT

Deep learning belongs to the field of artificial intelligence, where machines perform tasks that typically require some kind of human intelligence. Similar to the basic structure of a brain, a deep learning algorithm consists of an artificial neural network, which resembles the biological brain structure. Mimicking the learning process of humans with their senses, deep learning networks are fed with (sensory) data, like texts, images, videos or sounds. These networks outperform the state-of-the-art methods in different tasks and, because of this, the whole field saw an exponential growth during the last years. This growth resulted in way over 10,000 publications per year in the last years. For example, the search engine PubMed alone, which covers only a sub-set of all publications in the medical field, provides already over 11,000 results in Q3 2020 for the search term 'deep learning', and around 90% of these results are from the last three years. Consequently, a complete overview over the field of deep learning is already impossible to obtain and, in the near future, it will potentially become difficult to obtain an overview over a subfield. However, there are several review articles about deep learning, which are focused on specific scientific fields or applications, for example deep learning advances in computer vision or in specific tasks like object detection. With these surveys as a foundation, the aim of this contribution is to provide a first high-level, categorized meta-survey of selected reviews on deep learning across different scientific disciplines. The categories (computer vision, language processing, medical informatics and additional works) have been chosen according to the underlying data sources (image, language, medical, mixed). In addition, we review the common architectures, methods, pros, cons, evaluations, challenges and future directions for every sub-category.

Motivation & Objective

• To provide a high-level, categorized meta-survey of deep learning review literature across scientific disciplines due to the field's exponential growth and information overload.
• To analyze the research impact of selected deep learning reviews through citation and reference counts, reflecting their influence in specific domains.
• To identify common deep learning architectures, methods, pros, cons, evaluation practices, and challenges across subfields such as computer vision, NLP, and medical informatics.
• To outline future research directions and critical challenges—such as interpretability, generalization, and ethical risks—based on synthesized findings from the reviewed literature.
• To offer researchers a consolidated, structured reference point for navigating the rapidly expanding deep learning research landscape.

Proposed method

• Conducted a multi-database search (IEEE Xplore, Scopus, DBLP, PubMed, Web of Science, Google Scholar) using keywords 'Deep Learning' and 'Review' or 'Survey'.
• Screened titles and abstracts to exclude non-review works, resulting in 61 selected review/survey publications for analysis.
• Categorized reviews into four main domains: computer vision, natural language processing, medical informatics, and additional interdisciplinary works based on data modality.
• Extracted and analyzed key information from each review, including referenced works, citations, common architectures, methods, pros, cons, evaluation metrics, and challenges.
• Synthesized findings into structured tables and summaries to enable high-level comparison across subfields and to assess research impact.
• Included selected preprints with high citation counts to capture emerging, influential works not yet peer-reviewed.

Experimental results

Research questions

• RQ1 What are the dominant deep learning architectures and methods used across major scientific domains such as computer vision, NLP, and medical informatics?
• RQ2 How has the research impact of deep learning reviews been quantified through citation and reference counts, and what does this reveal about the field's growth and influence?
• RQ3 What are the most frequently cited challenges and limitations in deep learning applications across different domains, and how do they vary by data modality?
• RQ4 How do evaluation practices and performance benchmarks differ across subfields, and what trends emerge in model performance and generalization?
• RQ5 What are the key future research directions and ethical concerns highlighted in the reviewed literature, particularly regarding interpretability, robustness, and real-world deployment?

Key findings

• Over 11,000 PubMed results for 'deep learning' were recorded by Q3 2020, with approximately 90% published in the last three years, indicating explosive growth in the field.
• The meta-survey identified 61 high-impact review papers across four main categories, with medical informatics and computer vision showing particularly strong citation and reference counts, reflecting their research influence.
• Convolutional Neural Networks (CNNs) and Transformers emerged as dominant architectures in computer vision and NLP, respectively, with Vision Transformers gaining traction in image analysis.
• Common challenges across domains include model interpretability (the 'black box' problem), data scarcity, domain shift, and robustness to distributional shifts, especially in medical applications.
• Despite high performance in controlled settings, deep learning models often fail in real-world scenarios due to unforeseen edge cases, as seen in self-driving car accidents and biased image classification.
• The field shows a strong trend toward multimodal and hybrid models, especially in medical imaging, where fusion of imaging, clinical, and genomic data is increasingly explored for improved diagnostic accuracy.