Deep learning is too large to review as a whole: OpenAlex lists more than 340,000 works with the phrase in the title, about 68,000 of them from 2025. Read LeCun, Bengio and Hinton (2015) for the concepts, Schmidhuber (2015) for the history, and then go straight to a survey for your application area, such as Litjens et al. (2017) for medical imaging. Many architecture papers (on convolutional networks, residual networks and transformers) do not have deep learning in the title, so they are missing here by construction.
How the literature is organised
Deep learning is a set of methods, and its literature is organised by architecture, task and application domain, not by competing theories. Two overview papers from 2015 serve as the common reference. LeCun, Bengio and Hinton explain representation learning, backpropagation, convolutional and recurrent networks in nine pages. Schmidhuber gives a historical survey that, in its own words, summarises relevant work, much of it from the previous millennium.
After that the most-cited papers divide into tools, methods and landmark applications. Tools: the PyTorch paper (Paszke et al., 2019). Methods: data augmentation (Shorten and Khoshgoftaar, 2019), physics-informed neural networks (Raissi, Perdikaris and Karniadakis, 2019), object detection reviewed by Zhao et al. (2019). Applications: Gulshan et al. (2016) validated a deep learning algorithm for detecting diabetic retinopathy in retinal photographs against grading by ophthalmologists; Litjens et al. (2017) surveyed medical image analysis; Mathis et al. (2018) introduced DeepLabCut for tracking animal movement; Reichstein et al. (2019) set an agenda for Earth system science.
Main debates
Why do very large networks generalise at all? Zhang et al. (2021), "Understanding deep learning (still) requires rethinking generalization", show through systematic experiments that the traditional explanations fail to account for it, and the paper is the most-cited entry to that question on this page. A second issue with a review of its own is uncertainty (Abdar et al., 2021, on its quantification). Opacity has a literature of its own, listed on the explainable AI page.
Where recent work is heading
General reviews by Alzubaidi et al. (2021) and Sarker (2021) give current taxonomies. Specialised reviews cover attention mechanisms (Niu, Zhong and Yu, 2021), ensembles (Ganaie et al., 2022), anomaly detection (Pang et al., 2022) and time-series forecasting (Lim and Zohren, 2021). The most-cited recent applications are in biomedicine: nnU-Net for segmentation (Isensee et al., 2021) and ProteinMPNN for protein sequence design (Dauparas et al., 2022).
Most-cited foundational papers
Published before 2021 and ranked by how often later work cites them. Read the abstract of each and the full text of the three or four closest to your question. Citation count measures attention, not quality, so treat this as a map of what the field has argued about rather than a ranking of what is true.
Zhong-Qiu Zhao and 3 others (2019). IEEE Transactions on Neural Networks and Learning Systems.
Cited by 5,500doi:10.1109/tnnls.2018.2876865
Most-cited papers since 2021
Primary studies and conceptual papers from 2021 onwards. A paper published in 2024 has had a few years to accumulate citations where the works in the section above have had decades, so compare these counts with each other rather than with the ones above.
Huiqiang Xie and 3 others (2021). IEEE Transactions on Signal Processing.
Cited by 1,570Open accessdoi:10.1109/tsp.2021.3071210
Recent reviews and meta-analyses
The fastest way into a literature. A good review gives you the structure of the field, a reference list to mine and, in its limitations section, the gaps other researchers have already spotted.
Bryan Lim, Stefan Zohren (2021). Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences.
Cited by 1,785Open accessdoi:10.1098/rsta.2020.0209
How big the literature is, and where it is published
OpenAlex indexes 343,352 works whose title matches this topic. The chart shows how many were published each year from 2000 to 2025; the current year is left out because it is incomplete.
Show the numbers as a table
Year
Works
2000
7
2001
10
2002
14
2003
16
2004
14
2005
30
2006
27
2007
22
2008
28
2009
45
2010
57
2011
89
2012
110
2013
179
2014
398
2015
832
2016
1,853
2017
4,521
2018
9,454
2019
15,942
2020
22,718
2021
29,226
2022
34,899
2023
43,789
2024
53,338
2025
67,955
Journals behind the most-cited work
Counted across the 170 most-cited works on the topic, not across everything published. Browsing recent issues of the first two or three is a reliable way to find current work that has not yet been cited much.
IEEE Access7 papers
Nature Medicine6 papers
Journal Of Big Data5 papers
Nature5 papers
ACM Computing Surveys4 papers
Nature Methods4 papers
Cell3 papers
Computers and Electronics in Agriculture3 papers
Search strings to copy
Written for databases that accept Boolean operators (Scopus, Web of Science, ERIC, PubMed, EBSCO). Quotation marks keep a phrase together, an asterisk stands in for the end of a word, and OR groups go in brackets. British and American spellings are written out with OR rather than covered by a single-character wildcard, because those wildcards differ between databases: PubMed’s help page documents the asterisk only, and asks for at least four characters before it. Limit the search to title and abstract first; widen it only if you get too little. Our guide to starting a literature review covers how to record what you searched.
Narrow by task
"deep learning" AND ("image segmentation" OR "object detection" OR "time series" OR "anomaly detection" OR "text classification")
Narrow by domain
("deep learning" OR "convolutional neural network*" OR "neural network*") AND (medical OR agriculture OR "remote sensing" OR finance OR climate)
Practical issues
"deep learning" AND ("data augmentation" OR "uncertainty quantification" OR "small data" OR "transfer learning" OR (generalisation OR generalization))
Explore deep learning in Wonders →Opens Wonders on this question. Once you are signed in it becomes a project with suggested sub-topics and keywords you can edit before searching. The trial is 14 days.
Sub-topics to narrow into
A thesis-sized question usually sits inside one of these, combined with a population or a setting.
Foundations and history
LeCun, Bengio and Hinton (2015) and Schmidhuber (2015).
Every paper above has a DOI, a part of the reference that is easy to leave out. Here is one of them, “Deep learning” (2015), in the two styles students ask about most:
APA 7th edition
LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444. https://doi.org/10.1038/nature14539
MLA 9th edition
LeCun, Yann, et al. “Deep Learning.” Nature, vol. 521, no. 7553, 2015, pp. 436–44, https://doi.org/10.1038/nature14539.
Check the details against the article itself before you submit: databases, including the one behind this page, sometimes carry the online-first year rather than the volume year. Full rules and more examples are in our guides to APA, MLA, Chicago, Harvard, Vancouver and ABNT, with the rest in the citation guides. You can also format a reference from its DOI with our free citation tools.
Frequently asked questions
What is the most cited deep learning paper?
Among works with the phrase in the title it is LeCun, Bengio and Hinton (2015) in Nature, with about 85,000 citations in OpenAlex. Some architecture papers without the phrase in their titles are cited even more, which is a limitation of any title-based list.
How do I write a literature review on such a large topic?
Do not review deep learning; review deep learning for one task in one domain, over a stated period. Start from a recent survey in that niche, check who cites it, and use a structured protocol so that your inclusion decisions are defensible. Our literature review guides explain how to document a search.
Why is a software paper (PyTorch) among the most cited?
Paszke et al. (2019) describes the PyTorch library, and researchers cite it when they use it. As with database papers in biology, the citation count reflects use as a tool. Cite it if you use PyTorch, not as evidence for a claim.
How this page was made
The lists come from OpenAlex, an open index of scholarly works whose data are published under a CC0 licence, queried on September 21, 2026 for works whose title matches "deep learning". Only works with a DOI are listed. Each one was checked against the publisher’s own record at Crossref or DataCite (title, year, first author, journal, volume and pages), and in three cases, where the publisher deposited no byline, against PubMed; anything OpenAlex or Crossref flags as retracted was left out, and an editor took out results that matched the words but not the subject. Citation counts are OpenAlex’s on that date and are usually lower than Google Scholar’s, which counts more kinds of document. Ranking by citations tells you what a field has relied on, not what is correct; several heavily cited papers on any topic are cited because later work disputes them. Books without a DOI are missing, which matters in fields where the founding text is a book.
Try Wonders for free