deep learning#Deep neural networks

Most modern deep learning models are based on multi-layered neural networks such as convolutional neural networks and transformers, although they can also include propositional formulas or latent variables organized layer-wise in deep generative models such as the nodes in deep belief networks and deep Boltzmann machines.{{cite journal|last=Bengio|first=Yoshua|year=2009|title=Learning Deep Architectures for AI|url=http://sanghv.com/download/soft/machine%20learning,%20artificial%20intelligence,%20mathematics%20ebooks/ML/learning%20deep%20architectures%20for%20AI%20%282009%29.pdf|journal=Foundations and Trends in Machine Learning|volume=2|issue=1|pages=1–127|doi=10.1561/2200000006|citeseerx=10.1.1.701.9550|s2cid=207178999 |access-date=3 September 2015|archive-url=https://web.archive.org/web/20160304084250/http://sanghv.com/download/soft/machine%20learning,%20artificial%20intelligence,%20mathematics%20ebooks/ML/learning%20deep%20architectures%20for%20AI%20(2009).pdf|archive-date=4 March 2016|url-status=dead}}

Fundamentally, deep learning refers to a class of machine learning algorithms in which a hierarchy of layers is used to transform input data into a progressively more abstract and composite representation. For example, in an image recognition model, the raw input may be an image (represented as a tensor of pixels). The first representational layer may attempt to identify basic shapes such as lines and circles, the second layer may compose and encode arrangements of edges, the third layer may encode a nose and eyes, and the fourth layer may recognize that the image contains a face.

Importantly, a deep learning process can learn which features to optimally place at which level on its own. Prior to deep learning, machine learning techniques often involved hand-crafted feature engineering to transform the data into a more suitable representation for a classification algorithm to operate on. In the deep learning approach, features are not hand-crafted and the model discovers useful feature representations from the data automatically. This does not eliminate the need for hand-tuning; for example, varying numbers of layers and layer sizes can provide different degrees of abstraction.{{cite journal|last2=Courville|first2=A.|last3=Vincent|first3=P.|year=2013|title=Representation Learning: A Review and New Perspectives|journal=IEEE Transactions on Pattern Analysis and Machine Intelligence|volume=35|issue=8|pages=1798–1828|arxiv=1206.5538|doi=10.1109/tpami.2013.50|pmid=23787338|last1=Bengio|first1=Y.|s2cid=393948}}

The word "deep" in "deep learning" refers to the number of layers through which the data is transformed. More precisely, deep learning systems have a substantial credit assignment path (CAP) depth. The CAP is the chain of transformations from input to output. CAPs describe potentially causal connections between input and output. For a feedforward neural network, the depth of the CAPs is that of the network and is the number of hidden layers plus one (as the output layer is also parameterized). For recurrent neural networks, in which a signal may propagate through a layer more than once, the CAP depth is potentially unlimited. No universally agreed-upon threshold of depth divides shallow learning from deep learning, but most researchers agree that deep learning involves CAP depth higher than two. CAP of depth two has been shown to be a universal approximator in the sense that it can emulate any function.{{Cite book|url=https://books.google.com/books?id=9CqQDwAAQBAJ&pg=PA15|title=Human Behavior and Another Kind in Consciousness: Emerging Research and Opportunities: Emerging Research and Opportunities|last=Shigeki|first=Sugiyama|date=12 April 2019|publisher=IGI Global|isbn=978-1-5225-8218-2|language=en}} Beyond that, more layers do not add to the function approximator ability of the network. Deep models (CAP > two) are able to extract better features than shallow models and hence, extra layers help in learning the features effectively.

Deep learning architectures can be constructed with a greedy layer-by-layer method.{{cite conference | first1=Yoshua | last1=Bengio | first2=Pascal | last2=Lamblin | first3=Dan | last3=Popovici | first4=Hugo | last4=Larochelle | title=Greedy layer-wise training of deep networks | year=2007 | url=http://papers.nips.cc/paper/3048-greedy-layer-wise-training-of-deep-networks.pdf | conference=Advances in neural information processing systems | pages=153–160 | access-date=2019-10-06 | archive-date=2019-10-20 | archive-url=https://web.archive.org/web/20191020195638/http://papers.nips.cc/paper/3048-greedy-layer-wise-training-of-deep-networks.pdf | url-status=live }} Deep learning helps to disentangle these abstractions and pick out which features improve performance.

Deep learning algorithms can be applied to unsupervised learning tasks. This is an important benefit because unlabeled data is more abundant than the labeled data. Examples of deep structures that can be trained in an unsupervised manner are deep belief networks.{{cite journal | last1 = Hinton | first1 = G.E. | year = 2009| title = Deep belief networks | journal = Scholarpedia | volume = 4 | issue = 5| page = 5947 | doi=10.4249/scholarpedia.5947| bibcode = 2009SchpJ...4.5947H| doi-access = free }}

The term Deep Learning was introduced to the machine learning community by Rina Dechter in 1986,Rina Dechter (1986). Learning while searching in constraint-satisfaction problems. University of California, Computer Science Department, Cognitive Systems Laboratory.[https://www.researchgate.net/publication/221605378_Learning_While_Searching_in_Constraint-Satisfaction-Problems Online] {{Webarchive|url=https://web.archive.org/web/20160419054654/https://www.researchgate.net/publication/221605378_Learning_While_Searching_in_Constraint-Satisfaction-Problems|date=2016-04-19}} and to artificial neural networks by Igor Aizenberg and colleagues in 2000, in the context of Boolean threshold neurons.{{cite book |author1=Aizenberg, I.N. |url=https://link.springer.com/book/10.1007/978-1-4757-3115-6 |title=Multi-Valued and Universal Binary Neurons |author2=Aizenberg, N.N. |author3=Vandewalle, J. |date=2000 |publisher=Science & Business Media |isbn=978-0-7923-7824-2 |doi=10.1007/978-1-4757-3115-6 |access-date=27 December 2023}}Co-evolving recurrent neurons learn deep memory POMDPs. Proc. GECCO, Washington, D. C., pp. 1795–1802, ACM Press, New York, NY, USA, 2005. Although the history of its appearance is apparently more complicated.{{Cite journal |last=Fradkov |first=Alexander L. |date=2020-01-01 |title=Early History of Machine Learning |journal=IFAC-PapersOnLine |series=21st IFAC World Congress |volume=53 |issue=2 |pages=1385–1390 |doi=10.1016/j.ifacol.2020.12.1888 |issn=2405-8963 |s2cid=235081987 |doi-access=free}}

Deep neural networks are generally interpreted in terms of the universal approximation theorem{{cite journal | last1 = Cybenko | year = 1989 | title = Approximations by superpositions of sigmoidal functions | url = http://deeplearning.cs.cmu.edu/pdfs/Cybenko.pdf | journal = Mathematics of Control, Signals, and Systems | volume = 2 | issue = 4 | pages = 303–314 | doi = 10.1007/bf02551274 | bibcode = 1989MCSS....2..303C | s2cid = 3958369 | url-status = dead | archive-url = https://web.archive.org/web/20151010204407/http://deeplearning.cs.cmu.edu/pdfs/Cybenko.pdf | archive-date = 10 October 2015 }}{{cite journal | last1 = Hornik | first1 = Kurt | year = 1991 | title = Approximation Capabilities of Multilayer Feedforward Networks | journal = Neural Networks | volume = 4 | issue = 2| pages = 251–257 | doi=10.1016/0893-6080(91)90009-t| s2cid = 7343126 }}{{cite book|first=Simon S. |last=Haykin|title=Neural Networks: A Comprehensive Foundation|url={{google books |plainurl=y |id=bX4pAQAAMAAJ}}|year=1999|publisher=Prentice Hall|isbn=978-0-13-273350-2}}{{cite book|first=Mohamad H. |last=Hassoun|title=Fundamentals of Artificial Neural Networks|url={{google books |plainurl=y |id=Otk32Y3QkxQC|page=48}}|year=1995|publisher=MIT Press|isbn=978-0-262-08239-6|page=48}}Lu, Z., Pu, H., Wang, F., Hu, Z., & Wang, L. (2017). [http://papers.nips.cc/paper/7203-the-expressive-power-of-neural-networks-a-view-from-the-width The Expressive Power of Neural Networks: A View from the Width] {{Webarchive|url=https://web.archive.org/web/20190213005539/http://papers.nips.cc/paper/7203-the-expressive-power-of-neural-networks-a-view-from-the-width |date=2019-02-13 }}. Neural Information Processing Systems, 6231-6239. or probabilistic inference.{{cite journal |last1=Orhan |first1=A. E. |last2=Ma |first2=W. J. |date=2017 |title=Efficient probabilistic inference in generic neural networks trained with non-probabilistic feedback |journal=Nature Communications |volume=8 |issue=1 |pages=138 | pmid=28743932 | doi=10.1038/s41467-017-00181-8|pmc=5527101 |bibcode=2017NatCo...8..138O | doi-access=free}}{{cite journal |last1=Deng |first1=L. |last2=Yu |first2=D. |year=2014 |title=Deep Learning: Methods and Applications |url=http://research.microsoft.com/pubs/209355/DeepLearning-NowPublishing-Vol7-SIG-039.pdf |url-status=live |journal=Foundations and Trends in Signal Processing |volume=7 |issue=3–4 |pages=1–199 |doi=10.1561/2000000039 |archive-url=https://web.archive.org/web/20160314152112/http://research.microsoft.com/pubs/209355/DeepLearning-NowPublishing-Vol7-SIG-039.pdf |archive-date=2016-03-14 |access-date=2014-10-18}}{{cite journal|last=Schmidhuber|first=J.|s2cid=11715509|year=2015|title=Deep Learning in Neural Networks: An Overview|journal=Neural Networks|volume=61|pages=85–117|arxiv=1404.7828|doi=10.1016/j.neunet.2014.09.003|pmid=25462637}}{{cite book|first=Kevin P. |last=Murphy|title=Machine Learning: A Probabilistic Perspective|url={{google books |plainurl=y |id=NZP6AQAAQBAJ}}|date=24 August 2012|publisher=MIT Press|isbn=978-0-262-01802-9}}

The classic universal approximation theorem concerns the capacity of feedforward neural networks with a single hidden layer of finite size to approximate continuous functions. In 1989, the first proof was published by George Cybenko for sigmoid activation functions and was generalised to feed-forward multi-layer architectures in 1991 by Kurt Hornik. Recent work also showed that universal approximation also holds for non-bounded activation functions such as Kunihiko Fukushima's rectified linear unit.{{cite journal |last1=Fukushima |first1=K. |date=1969 |title=Visual feature extraction by a multilayered network of analog threshold elements |journal=IEEE Transactions on Systems Science and Cybernetics |volume=5 |issue=4 |pages=322–333 |doi=10.1109/TSSC.1969.300225}}{{cite journal | last1 = Sonoda | first1 = Sho | last2=Murata | first2=Noboru | s2cid = 12149203 | year = 2017 | title = Neural network with unbounded activation functions is universal approximator | journal = Applied and Computational Harmonic Analysis | volume = 43 | issue = 2 | pages = 233–268 | doi = 10.1016/j.acha.2015.12.005| arxiv = 1505.03654 }}

The universal approximation theorem for deep neural networks concerns the capacity of networks with bounded width but the depth is allowed to grow. Lu et al. proved that if the width of a deep neural network with ReLU activation is strictly larger than the input dimension, then the network can approximate any Lebesgue integrable function; if the width is smaller or equal to the input dimension, then a deep neural network is not a universal approximator.

The probabilistic interpretation derives from the field of machine learning. It features inference, as well as the optimization concepts of training and testing, related to fitting and generalization, respectively. More specifically, the probabilistic interpretation considers the activation nonlinearity as a cumulative distribution function. The probabilistic interpretation led to the introduction of dropout as regularizer in neural networks. The probabilistic interpretation was introduced by researchers including Hopfield, Widrow and Narendra and popularized in surveys such as the one by Bishop.{{cite book|title=Pattern Recognition and Machine Learning|author=Bishop, Christopher M.|year=2006|publisher=Springer|url=http://users.isr.ist.utl.pt/~wurmd/Livros/school/Bishop%20-%20Pattern%20Recognition%20And%20Machine%20Learning%20-%20Springer%20%202006.pdf|isbn=978-0-387-31073-2|access-date=2017-08-06|archive-date=2017-01-11|archive-url=https://web.archive.org/web/20170111005101/http://users.isr.ist.utl.pt/~wurmd/Livros/school/Bishop%20-%20Pattern%20Recognition%20And%20Machine%20Learning%20-%20Springer%20%202006.pdf|url-status=live}}

There are two types of artificial neural network (ANN): feedforward neural network (FNN) or multilayer perceptron (MLP) and recurrent neural networks (RNN). RNNs have cycles in their connectivity structure, FNNs don't. In the 1920s, Wilhelm Lenz and Ernst Ising created the Ising model{{Cite web|url=https://www.hs-augsburg.de/~harsch/anglica/Chronology/20thC/Ising/isi_fm00.html|title=bibliotheca Augustana|website=www.hs-augsburg.de}}{{cite journal |doi=10.1103/RevModPhys.39.883|title=History of the Lenz-Ising Model|year=1967|last1=Brush|first1=Stephen G.|journal=Reviews of Modern Physics|volume=39|issue=4|pages=883–893|bibcode=1967RvMP...39..883B}} which is essentially a non-learning RNN architecture consisting of neuron-like threshold elements. In 1972, Shun'ichi Amari made this architecture adaptive.{{cite journal |last1=Amari |first1=Shun-Ichi |title=Learning patterns and pattern sequences by self-organizing nets of threshold elements|journal= IEEE Transactions |date=1972 |volume=C |issue=21 |pages=1197–1206 }} His learning RNN was republished by John Hopfield in 1982.{{cite journal |last1=Hopfield |first1=J. J. |title=Neural networks and physical systems with emergent collective computational abilities |journal= Proceedings of the National Academy of Sciences|date=1982 |volume=79 |issue=8 |pages=2554–2558 |doi=10.1073/pnas.79.8.2554 |pmid=6953413 |pmc=346238 |bibcode=1982PNAS...79.2554H |doi-access=free }} Other early recurrent neural networks were published by Kaoru Nakano in 1971.{{cite book |last1=Nakano |first1=Kaoru |chapter=Learning Process in a Model of Associative Memory |title=Pattern Recognition and Machine Learning |date=1971|pages=172–186 |doi=10.1007/978-1-4615-7566-5_15 |isbn=978-1-4615-7568-9 }}{{cite journal |last1=Nakano |first1=Kaoru |title=Associatron-A Model of Associative Memory|journal= IEEE Transactions on Systems, Man, and Cybernetics |date=1972 |volume=SMC-2 |issue=3|pages=380–388 |doi=10.1109/TSMC.1972.4309133 }} Already in 1948, Alan Turing produced work on "Intelligent Machinery" that was not published in his lifetime,{{cite book |last1=Turing |first1=Alan |chapter=Intelligent Machinery |editor-last=Ince |editor-first=D.C. |title=Collected Works of AM Turing: Mechanical Intelligence |volume=1 |page=107 |publisher=Elsevier Science Publishers |year=1992 |orig-date=1948 |isbn=0-444-88058-5}} containing "ideas related to artificial evolution and learning RNNs".

Frank Rosenblatt (1958){{Cite journal |last=Rosenblatt |first=F. |date=1958 |title=The perceptron: A probabilistic model for information storage and organization in the brain. |url=https://doi.apa.org/doi/10.1037/h0042519 |journal=Psychological Review |language=en |volume=65 |issue=6 |pages=386–408 |doi=10.1037/h0042519 |pmid=13602029 |issn=1939-1471}} proposed the perceptron, an MLP with 3 layers: an input layer, a hidden layer with randomized weights that did not learn, and an output layer. He later published a 1962 book that also introduced variants and computer experiments, including a version with four-layer perceptrons "with adaptive preterminal networks" where the last two layers have learned weights (here he credits H. D. Block and B. W. Knight).{{cite book |last=Rosenblatt |first=Frank |author-link=Frank Rosenblatt |title=Principles of Neurodynamics |publisher=Spartan, New York |year=1962}}{{rp|section 16}} The book cites an earlier network by R. D. Joseph (1960){{cite book |last=Joseph |first=R. D. |title=Contributions to Perceptron Theory, Cornell Aeronautical Laboratory Report No. VG-11 96--G-7, Buffalo |year=1960}} "functionally equivalent to a variation of" this four-layer system (the book mentions Joseph over 30 times). Should Joseph therefore be considered the originator of proper adaptive multilayer perceptrons with learning hidden units? Unfortunately, the learning algorithm was not a functional one, and fell into oblivion.

The first working deep learning algorithm was the Group method of data handling, a method to train arbitrarily deep neural networks, published by Alexey Ivakhnenko and Lapa in 1965. They regarded it as a form of polynomial regression,{{cite book|first1=A. G. |last1=Ivakhnenko |first2=V. G. |last2=Lapa |title=Cybernetics and Forecasting Techniques|url={{google books |plainurl=y |id=rGFgAAAAMAAJ}}|year=1967|publisher=American Elsevier Publishing Co.|isbn=978-0-444-00020-0}} or a generalization of Rosenblatt's perceptron.{{Cite journal |last=Ivakhnenko |first=A.G. |date=March 1970 |title=Heuristic self-organization in problems of engineering cybernetics |url=https://linkinghub.elsevier.com/retrieve/pii/0005109870900920 |journal=Automatica |language=en |volume=6 |issue=2 |pages=207–219 |doi=10.1016/0005-1098(70)90092-0}} A 1971 paper described a deep network with eight layers trained by this method,{{Cite journal|last=Ivakhnenko|first=Alexey|date=1971|title=Polynomial theory of complex systems|url=http://gmdh.net/articles/history/polynomial.pdf|journal=IEEE Transactions on Systems, Man, and Cybernetics|pages=364–378|doi=10.1109/TSMC.1971.4308320|volume=SMC-1|issue=4|access-date=2019-11-05|archive-date=2017-08-29|archive-url=https://web.archive.org/web/20170829230621/http://www.gmdh.net/articles/history/polynomial.pdf|url-status=live}} which is based on layer by layer training through regression analysis. Superfluous hidden units are pruned using a separate validation set. Since the activation functions of the nodes are Kolmogorov-Gabor polynomials, these were also the first deep networks with multiplicative units or "gates".

The first deep learning multilayer perceptron trained by stochastic gradient descent{{Cite journal | last1 = Robbins | first1 = H. | author-link = Herbert Robbins| last2 = Monro | first2 = S. | doi = 10.1214/aoms/1177729586 | title = A Stochastic Approximation Method | journal = The Annals of Mathematical Statistics | volume = 22 | issue = 3 | pages = 400 | year = 1951 | doi-access = free }} was published in 1967 by Shun'ichi Amari.{{cite journal |last1=Amari |first1=Shun'ichi |author-link=Shun'ichi Amari|title=A theory of adaptive pattern classifier|journal= IEEE Transactions |date=1967 |volume=EC |issue=16 |pages=279–307}} In computer experiments conducted by Amari's student Saito, a five layer MLP with two modifiable layers learned internal representations to classify non-linearily separable pattern classes.{{cite arXiv |eprint=2212.11279 |class=cs.NE |first=Jürgen |last=Schmidhuber |author-link=Jürgen Schmidhuber |title=Annotated History of Modern AI and Deep Learning |date=2022}} Subsequent developments in hardware and hyperparameter tunings have made end-to-end stochastic gradient descent the currently dominant training technique.

In 1969, Kunihiko Fukushima introduced the ReLU (rectified linear unit) activation function. The rectifier has become the most popular activation function for deep learning.{{cite arXiv |eprint=1710.05941 |class=cs.NE |first1=Prajit |last1=Ramachandran |first2=Zoph |last2=Barret |title=Searching for Activation Functions |date=October 16, 2017 |last3=Quoc |first3=V. Le}}

Deep learning architectures for convolutional neural networks (CNNs) with convolutional layers and downsampling layers began with the Neocognitron introduced by Kunihiko Fukushima in 1979, though not trained by backpropagation.{{cite journal |last1=Fukushima |first1=K. |year=1979 |title=Neural network model for a mechanism of pattern recognition unaffected by shift in position—Neocognitron |journal=Trans. IECE (In Japanese)|volume= J62-A |issue=10 |pages=658–665 |doi=10.1007/bf00344251 |pmid=7370364 |s2cid=206775608}}{{cite journal |last1=Fukushima |first1=K. |year=1980 |title=Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position |journal=Biol. Cybern. |volume=36 |issue=4 |pages=193–202 |doi=10.1007/bf00344251 |pmid=7370364 |s2cid=206775608}}

Backpropagation is an efficient application of the chain rule derived by Gottfried Wilhelm Leibniz in 1673{{Cite book |last=Leibniz |first=Gottfried Wilhelm Freiherr von |url=https://books.google.com/books?id=bOIGAAAAYAAJ&q=leibniz+altered+manuscripts&pg=PA90 |title=The Early Mathematical Manuscripts of Leibniz: Translated from the Latin Texts Published by Carl Immanuel Gerhardt with Critical and Historical Notes (Leibniz published the chain rule in a 1676 memoir) |date=1920 |publisher=Open court publishing Company |isbn=9780598818461 |language=en}} to networks of differentiable nodes. The terminology "back-propagating errors" was actually introduced in 1962 by Rosenblatt, but he did not know how to implement this, although Henry J. Kelley had a continuous precursor of backpropagation in 1960 in the context of control theory.{{cite journal |last1=Kelley |first1=Henry J. |author-link=Henry J. Kelley |year=1960 |title=Gradient theory of optimal flight paths |journal=ARS Journal |volume=30 |issue=10 |pages=947–954 |doi=10.2514/8.5282}} The modern form of backpropagation was first published in Seppo Linnainmaa's master thesis (1970).{{cite thesis |first=Seppo |last=Linnainmaa |author-link=Seppo Linnainmaa |year=1970 |type=Masters |title=The representation of the cumulative rounding error of an algorithm as a Taylor expansion of the local rounding errors |language=fi |publisher=University of Helsinki |page=6–7}}{{cite journal |last1=Linnainmaa |first1=Seppo |author-link=Seppo Linnainmaa |year=1976 |title=Taylor expansion of the accumulated rounding error |journal=BIT Numerical Mathematics |volume=16 |issue=2 |pages=146–160 |doi=10.1007/bf01931367 |s2cid=122357351}} G.M. Ostrovski et al. republished it in 1971.Ostrovski, G.M., Volin,Y.M., and Boris, W.W. (1971). On the computation of derivatives. Wiss. Z. Tech. Hochschule for Chemistry, 13:382–384. Paul Werbos applied backpropagation to neural networks in 1982{{cite book |last=Werbos |first=Paul |author-link=Paul Werbos |title=System modeling and optimization |publisher=Springer |year=1982 |pages=762–770 |chapter=Applications of advances in nonlinear sensitivity analysis |access-date=2 July 2017 |chapter-url=http://werbos.com/Neural/SensitivityIFIPSeptember1981.pdf |archive-url=https://web.archive.org/web/20160414055503/http://werbos.com/Neural/SensitivityIFIPSeptember1981.pdf |archive-date=14 April 2016 |url-status=live}} (his 1974 PhD thesis, reprinted in a 1994 book,{{cite book |last=Werbos |first=Paul J. |title=The Roots of Backpropagation : From Ordered Derivatives to Neural Networks and Political Forecasting |location=New York |publisher=John Wiley & Sons |year=1994 |isbn=0-471-59897-6 }} did not yet describe the algorithm{{cite web | last = Schmidhuber | first = Juergen | title = Who Invented Backpropagation? | author-link=Juergen Schmidhuber| publisher = IDSIA, Switzerland | url = https://people.idsia.ch/~juergen/who-invented-backpropagation.html | date = 25 Oct 2014 | access-date = 14 Sep 2024 | archive-url = http://web.archive.org/web/20240730110408/https://people.idsia.ch/~juergen/who-invented-backpropagation.html | archive-date = 30 July 2024 | quote = }}). In 1986, David E. Rumelhart et al. popularised backpropagation but did not cite the original work.{{Cite journal |last1=Rumelhart |first1=David E. |last2=Hinton |first2=Geoffrey E. |last3=Williams |first3=Ronald J. |date=October 1986 |title=Learning representations by back-propagating errors |url=https://www.nature.com/articles/323533a0 |journal=Nature |language=en |volume=323 |issue=6088 |pages=533–536 |doi=10.1038/323533a0 |bibcode=1986Natur.323..533R |issn=1476-4687}}Rumelhart, David E., Geoffrey E. Hinton, and R. J. Williams. "[https://apps.dtic.mil/dtic/tr/fulltext/u2/a164453.pdf Learning Internal Representations by Error Propagation] {{Webarchive|url=https://web.archive.org/web/20221013070443/https://apps.dtic.mil/dtic/tr/fulltext/u2/a164453.pdf |date=2022-10-13 }}". David E. Rumelhart, James L. McClelland, and the PDP research group. (editors), Parallel distributed processing: Explorations in the microstructure of cognition, Volume 1: Foundation. MIT Press, 1986.

The time delay neural network (TDNN) was introduced in 1987 by Alex Waibel to apply CNN to phoneme recognition. It used convolutions, weight sharing, and backpropagation.{{cite conference |title=Phoneme Recognition Using Time-Delay Neural Networks |last1=Waibel |first1=Alex |date=December 1987 |location=Tokyo, Japan |conference=Meeting of the Institute of Electrical, Information and Communication Engineers (IEICE) | url=https://isl.anthropomatik.kit.edu/pdf/Waibel1987a.pdf}}Alexander Waibel et al., [http://www.inf.ufrgs.br/~engel/data/media/file/cmp121/waibel89_TDNN.pdf Phoneme Recognition Using Time-Delay Neural Networks] IEEE Transactions on Acoustics, Speech, and Signal Processing, Volume 37, No. 3, pp. 328. – 339 March 1989. In 1988, Wei Zhang applied a backpropagation-trained CNN to alphabet recognition.{{cite journal |last=Zhang |first=Wei |date=1988 |title=Shift-invariant pattern recognition neural network and its optical architecture |url=https://drive.google.com/file/d/1nN_5odSG_QVae54EsQN_qSz-0ZsX6wA0/view?usp=sharing |journal=Proceedings of Annual Conference of the Japan Society of Applied Physics}}

In 1989, Yann LeCun et al. created a CNN called LeNet for recognizing handwritten ZIP codes on mail. Training required 3 days.LeCun et al., "Backpropagation Applied to Handwritten Zip Code Recognition", Neural Computation, 1, pp. 541–551, 1989. In 1990, Wei Zhang implemented a CNN on optical computing hardware.{{cite journal |last=Zhang |first=Wei |date=1990 |title=Parallel distributed processing model with local space-invariant interconnections and its optical architecture |url=https://drive.google.com/file/d/0B65v6Wo67Tk5ODRzZmhSR29VeDg/view?usp=sharing |journal=Applied Optics |volume=29 |issue=32 |pages=4790–7 |bibcode=1990ApOpt..29.4790Z |doi=10.1364/AO.29.004790 |pmid=20577468}} In 1991, a CNN was applied to medical image object segmentation{{cite journal |last=Zhang |first=Wei |date=1991 |title=Image processing of human corneal endothelium based on a learning network |url=https://drive.google.com/file/d/0B65v6Wo67Tk5cm5DTlNGd0NPUmM/view?usp=sharing |journal=Applied Optics |volume=30 |issue=29 |pages=4211–7 |bibcode=1991ApOpt..30.4211Z |doi=10.1364/AO.30.004211 |pmid=20706526}} and breast cancer detection in mammograms.{{cite journal |last=Zhang |first=Wei |date=1994 |title=Computerized detection of clustered microcalcifications in digital mammograms using a shift-invariant artificial neural network |url=https://drive.google.com/file/d/0B65v6Wo67Tk5Ml9qeW5nQ3poVTQ/view?usp=sharing |journal=Medical Physics |volume=21 |issue=4 |pages=517–24 |bibcode=1994MedPh..21..517Z |doi=10.1118/1.597177 |pmid=8058017}} LeNet-5 (1998), a 7-level CNN by Yann LeCun et al., that classifies digits, was applied by several banks to recognize hand-written numbers on checks digitized in 32x32 pixel images.{{cite journal |last=LeCun |first=Yann |author2=Léon Bottou |author3=Yoshua Bengio |author4=Patrick Haffner |year=1998 |title=Gradient-based learning applied to document recognition |url=http://yann.lecun.com/exdb/publis/pdf/lecun-01a.pdf |journal=Proceedings of the IEEE |volume=86 |issue=11 |pages=2278–2324 |citeseerx=10.1.1.32.9552 |doi=10.1109/5.726791 |s2cid=14542261 |access-date=October 7, 2016}}

Recurrent neural networks (RNN) were further developed in the 1980s. Recurrence is used for sequence processing, and when a recurrent network is unrolled, it mathematically resembles a deep feedforward layer. Consequently, they have similar properties and issues, and their developments had mutual influences. In RNN, two early influential works were the Jordan network (1986){{Cite journal |last=Jordan |first=Michael I. |date=1986 |title=Attractor dynamics and parallelism in a connectionist sequential machine |url=https://escholarship.org/uc/item/1fg2j76h |journal=Proceedings of the Annual Meeting of the Cognitive Science Society |language=en |volume=8 }} and the Elman network (1990),{{Cite journal |last=Elman |first=Jeffrey L. |date=March 1990 |title=Finding Structure in Time |url=https://onlinelibrary.wiley.com/doi/10.1207/s15516709cog1402_1 |journal=Cognitive Science |language=en |volume=14 |issue=2 |pages=179–211 |doi=10.1207/s15516709cog1402_1 |issn=0364-0213}} which applied RNN to study problems in cognitive psychology.

In the 1980s, backpropagation did not work well for deep learning with long credit assignment paths. To overcome this problem, in 1991, Jürgen Schmidhuber proposed a hierarchy of RNNs pre-trained one level at a time by self-supervised learning where each RNN tries to predict its own next input, which is the next unexpected input of the RNN below.{{cite journal |last1=Schmidhuber |first1=Jürgen |date= April 1991 |title=Neural Sequence Chunkers | author-link=Jürgen Schmidhuber |url=https://people.idsia.ch/~juergen/FKI-148-91ocr.pdf|journal= TR FKI-148, TU Munich}}{{cite journal |last1=Schmidhuber |first1=Jürgen |year=1992 |title=Learning complex, extended sequences using the principle of history compression (based on TR FKI-148, 1991) |url=https://sferics.idsia.ch/pub/juergen/chunker.pdf|journal=Neural Computation |volume=4 |issue=2 |pages=234–242 |doi=10.1162/neco.1992.4.2.234 |s2cid=18271205 }} This "neural history compressor" uses predictive coding to learn internal representations at multiple self-organizing time scales. This can substantially facilitate downstream deep learning. The RNN hierarchy can be collapsed into a single RNN, by distilling a higher level chunker network into a lower level automatizer network. In 1993, a neural history compressor solved a "Very Deep Learning" task that required more than 1000 subsequent layers in an RNN unfolded in time.{{Cite book |last=Schmidhuber |first=Jürgen |url=https://sferics.idsia.ch/pub/juergen/habilitation.pdf |archive-url=https://web.archive.org/web/20220516075450/https://sferics.idsia.ch/pub/juergen/habilitation.pdf |url-status=dead |archive-date=May 16, 2022 |title=Habilitation thesis: System modeling and optimization |year=1993}} Page 150 ff demonstrates credit assignment across the equivalent of 1,200 layers in an unfolded RNN. The "P" in ChatGPT refers to such pre-training.

Sepp Hochreiter's diploma thesis (1991)S. Hochreiter., "[http://people.idsia.ch/~juergen/SeppHochreiter1991ThesisAdvisorSchmidhuber.pdf Untersuchungen zu dynamischen neuronalen Netzen]". {{Webarchive|url=https://web.archive.org/web/20150306075401/http://people.idsia.ch/~juergen/SeppHochreiter1991ThesisAdvisorSchmidhuber.pdf |date=2015-03-06 }}. Diploma thesis. Institut f. Informatik, Technische Univ. Munich. Advisor: J. Schmidhuber, 1991. implemented the neural history compressor, and identified and analyzed the vanishing gradient problem.{{cite book|chapter-url={{google books |plainurl=y |id=NWOcMVA64aAC}}|title=A Field Guide to Dynamical Recurrent Networks|last=Hochreiter|first=S.|display-authors=etal|date=15 January 2001|publisher=John Wiley & Sons|isbn=978-0-7803-5369-5|chapter=Gradient flow in recurrent nets: the difficulty of learning long-term dependencies|editor-last2=Kremer|editor-first2=Stefan C.|editor-first1=John F.|editor-last1=Kolen}} Hochreiter proposed recurrent residual connections to solve the vanishing gradient problem. This led to the long short-term memory (LSTM), published in 1995.{{Cite Q|Q98967430}} LSTM can learn "very deep learning" tasks with long credit assignment paths that require memories of events that happened thousands of discrete time steps before. That LSTM was not yet the modern architecture, which required a "forget gate", introduced in 1999,{{Cite book |doi = 10.1049/cp:19991218|chapter = Learning to forget: Continual prediction with LSTM|title = 9th International Conference on Artificial Neural Networks: ICANN '99|volume = 1999|pages = 850–855|year = 1999|last1 = Gers|first1 = Felix| last2 = Schmidhuber|first2 = Jürgen| last3 = Cummins|first3 = Fred| isbn = 0-85296-721-7}} which became the standard RNN architecture.

In 1991, Jürgen Schmidhuber also published adversarial neural networks that contest with each other in the form of a zero-sum game, where one network's gain is the other network's loss.{{cite conference| title = A possibility for implementing curiosity and boredom in model-building neural controllers | last1 = Schmidhuber | first1 = Jürgen | author-link = Jürgen Schmidhuber | date = 1991 | publisher = MIT Press/Bradford Books| book-title = Proc. SAB'1991| pages = 222–227}}{{cite journal|last1=Schmidhuber|first1=Jürgen|author-link=Jürgen Schmidhuber|year=2010|title=Formal Theory of Creativity, Fun, and Intrinsic Motivation (1990-2010)|journal= IEEE Transactions on Autonomous Mental Development|volume=2|issue=3|pages=230–247|doi=10.1109/TAMD.2010.2056368 |s2cid=234198 }} The first network is a generative model that models a probability distribution over output patterns. The second network learns by gradient descent to predict the reactions of the environment to these patterns. This was called "artificial curiosity". In 2014, this principle was used in generative adversarial networks (GANs).{{Cite journal|last=Schmidhuber|first=Jürgen| author-link = Jürgen Schmidhuber |date=2020|title=Generative Adversarial Networks are Special Cases of Artificial Curiosity (1990) and also Closely Related to Predictability Minimization (1991)|journal=Neural Networks |language=en|volume=127|pages=58–66|doi=10.1016/j.neunet.2020.04.008 |pmid=32334341 |arxiv=1906.04493 |s2cid=216056336 }}

During 1985–1995, inspired by statistical mechanics, several architectures and methods were developed by Terry Sejnowski, Peter Dayan, Geoffrey Hinton, etc., including the Boltzmann machine,{{Cite journal |last1=Ackley |first1=David H. |last2=Hinton |first2=Geoffrey E. |last3=Sejnowski |first3=Terrence J. |date=1985-01-01 |title=A learning algorithm for boltzmann machines |url=https://www.sciencedirect.com/science/article/pii/S0364021385800124 |journal=Cognitive Science |volume=9 |issue=1 |pages=147–169 |doi=10.1016/S0364-0213(85)80012-4 |issn=0364-0213}} restricted Boltzmann machine,{{cite book |last=Smolensky |first=Paul |title=Parallel Distributed Processing: Explorations in the Microstructure of Cognition, Volume 1: Foundations |title-link=Connectionism |publisher=MIT Press |year=1986 |isbn=0-262-68053-X |editor1-last=Rumelhart |editor1-first=David E. |pages=[https://archive.org/details/paralleldistribu00rume/page/194 194–281] |chapter=Chapter 6: Information Processing in Dynamical Systems: Foundations of Harmony Theory |editor2-last=McLelland |editor2-first=James L. |chapter-url=https://stanford.edu/~jlmcc/papers/PDP/Volume%201/Chap6_PDP86.pdf}} Helmholtz machine,{{Cite journal |last1=Peter |first1=Dayan |last2=Hinton |first2=Geoffrey E. |last3=Neal |first3=Radford M. |last4=Zemel |first4=Richard S. |date=1995 |title=The Helmholtz machine. |journal=Neural Computation |volume=7 |issue=5 |pages=889–904 |doi=10.1162/neco.1995.7.5.889 |pmid=7584891 |s2cid=1890561 |hdl-access=free |author-link1=Peter Dayan |author-link2=Geoffrey Hinton |author-link3=Radford M. Neal |author-link4=Richard Zemel |hdl=21.11116/0000-0002-D6D3-E}} {{closed access}} and the wake-sleep algorithm.{{Cite journal |last1=Hinton |first1=Geoffrey E. |author-link=Geoffrey Hinton |last2=Dayan |first2=Peter |last3=Frey |first3=Brendan J. |last4=Neal |first4=Radford |date=1995-05-26 |title=The wake-sleep algorithm for unsupervised neural networks |journal=Science |volume=268 |issue=5214 |pages=1158–1161 |bibcode=1995Sci...268.1158H |doi=10.1126/science.7761831 |pmid=7761831 |s2cid=871473 |author-link2=Peter Dayan |author-link3=Brendan Frey}} These were designed for unsupervised learning of deep generative models. However, those were more computationally expensive compared to backpropagation. Boltzmann machine learning algorithm, published in 1985, was briefly popular before being eclipsed by the backpropagation algorithm in 1986. (p. 112 {{Cite book |last=Sejnowski |first=Terrence J. |title=The Deep Learning Revolution |date=2018 |publisher=The MIT Press |isbn=978-0-262-03803-4 |location=Cambridge, Massachusetts}}). A 1988 network became state of the art in protein structure prediction, an early application of deep learning to bioinformatics.{{Cite journal |last1=Qian |first1=Ning |last2=Sejnowski |first2=Terrence J. |date=1988-08-20 |title=Predicting the secondary structure of globular proteins using neural network models |url=https://dx.doi.org/10.1016/0022-2836%2888%2990564-5 |journal=Journal of Molecular Biology |volume=202 |issue=4 |pages=865–884 |doi=10.1016/0022-2836(88)90564-5 |pmid=3172241 |issn=0022-2836}}

Both shallow and deep learning (e.g., recurrent nets) of ANNs for speech recognition have been explored for many years.{{Cite journal|last1=Morgan|first1=Nelson|last2=Bourlard |first2=Hervé |last3=Renals |first3=Steve |last4=Cohen |first4=Michael|last5=Franco |first5=Horacio |date=1 August 1993 |title=Hybrid neural network/hidden markov model systems for continuous speech recognition |journal=International Journal of Pattern Recognition and Artificial Intelligence|volume=07|issue=4|pages=899–916|doi=10.1142/s0218001493000455|issn=0218-0014}}{{Cite journal|last=Robinson|first=T.|author-link=Tony Robinson (speech recognition)|date=1992|title=A real-time recurrent error propagation network word recognition system|url=http://dl.acm.org/citation.cfm?id=1895720|journal=ICASSP|pages=617–620|isbn=9780780305328|series=Icassp'92|access-date=2017-06-12|archive-date=2021-05-09|archive-url=https://web.archive.org/web/20210509123135/https://dl.acm.org/doi/10.5555/1895550.1895720|url-status=live}}{{Cite journal|last1=Waibel|first1=A.|last2=Hanazawa|first2=T.|last3=Hinton|first3=G.|last4=Shikano|first4=K.|last5=Lang|first5=K. J.|date=March 1989|title=Phoneme recognition using time-delay neural networks|journal=IEEE Transactions on Acoustics, Speech, and Signal Processing|volume=37|issue=3|pages=328–339|doi=10.1109/29.21701|issn=0096-3518|hdl=10338.dmlcz/135496|s2cid=9563026 |url=http://dml.cz/bitstream/handle/10338.dmlcz/135496/Kybernetika_38-2002-6_2.pdf|access-date=2019-09-24|archive-date=2021-04-27|archive-url=https://web.archive.org/web/20210427001446/https://dml.cz/bitstream/handle/10338.dmlcz/135496/Kybernetika_38-2002-6_2.pdf|url-status=live}} These methods never outperformed non-uniform internal-handcrafting Gaussian mixture model/Hidden Markov model (GMM-HMM) technology based on generative models of speech trained discriminatively.{{cite journal | last1 = Baker | first1 = J. | last2 = Deng | first2 = Li | last3 = Glass | first3 = Jim | last4 = Khudanpur | first4 = S. | last5 = Lee | first5 = C.-H. | last6 = Morgan | first6 = N. | last7 = O'Shaughnessy | first7 = D. | year = 2009 | title = Research Developments and Directions in Speech Recognition and Understanding, Part 1 | journal = IEEE Signal Processing Magazine | volume = 26 | issue = 3| pages = 75–80 | doi=10.1109/msp.2009.932166| bibcode = 2009ISPM...26...75B | hdl = 1721.1/51891 | s2cid = 357467 }} Key difficulties have been analyzed, including gradient diminishing and weak temporal correlation structure in neural predictive models.{{Cite web|url=https://www.researchgate.net/publication/41229141|title=Artificial Neural Networks and their Application to Speech/Sequence Recognition|last=Bengio|first=Y.|date=1991|publisher=McGill University Ph.D. thesis|access-date=2017-06-12|archive-date=2021-05-09|archive-url=https://web.archive.org/web/20210509123131/https://www.researchgate.net/publication/41229141_Artificial_neural_networks_and_their_application_to_sequence_recognition|url-status=live}}{{cite journal | last1 = Deng | first1 = L. | last2 = Hassanein | first2 = K. | last3 = Elmasry | first3 = M. | year = 1994 | title = Analysis of correlation structure for a neural predictive model with applications to speech recognition | journal = Neural Networks | volume = 7 | issue = 2| pages = 331–339 | doi=10.1016/0893-6080(94)90027-2}} Additional difficulties were the lack of training data and limited computing power.

Most speech recognition researchers moved away from neural nets to pursue generative modeling. An exception was at SRI International in the late 1990s. Funded by the US government's NSA and DARPA, SRI researched in speech and speaker recognition. The speaker recognition team led by Larry Heck reported significant success with deep neural networks in speech processing in the 1998 NIST Speaker Recognition benchmark.{{cite journal | last1 = Doddington | first1 = G. | last2 = Przybocki | first2 = M. | last3 = Martin | first3 = A. | last4 = Reynolds | first4 = D. | year = 2000 | title = The NIST speaker recognition evaluation ± Overview, methodology, systems, results, perspective | journal = Speech Communication | volume = 31 | issue = 2| pages = 225–254 | doi=10.1016/S0167-6393(99)00080-1}}{{cite journal |last1=Heck |first1=L. |last2=Konig |first2=Y. |last3=Sonmez |first3=M. |last4=Weintraub |first4=M. |year=2000 |title=Robustness to Telephone Handset Distortion in Speaker Recognition by Discriminative Feature Design |journal=Speech Communication |volume=31 |issue=2 |pages=181–192 |doi=10.1016/s0167-6393(99)00077-1}} It was deployed in the Nuance Verifier, representing the first major industrial application of deep learning.L.P Heck and R. Teunen. "Secure and Convenient Transactions with Nuance Verifier". Nuance Users Conference, April 1998.

The principle of elevating "raw" features over hand-crafted optimization was first explored successfully in the architecture of deep autoencoder on the "raw" spectrogram or linear filter-bank features in the late 1990s, showing its superiority over the Mel-Cepstral features that contain stages of fixed transformation from spectrograms. The raw features of speech, waveforms, later produced excellent larger-scale results.{{Cite web|url=https://www.researchgate.net/publication/266030526|title=Acoustic Modeling with Deep Neural Networks Using Raw Time Signal for LVCSR (PDF Download Available)|website=ResearchGate|access-date=14 June 2017|archive-date=9 May 2021|archive-url=https://web.archive.org/web/20210509123218/https://www.researchgate.net/publication/266030526_Acoustic_Modeling_with_Deep_Neural_Networks_Using_Raw_Time_Signal_for_LVCSR|url-status=live}}

Neural networks entered a lull, and simpler models that use task-specific handcrafted features such as Gabor filters and support vector machines (SVMs) became the preferred choices in the 1990s and 2000s, because of artificial neural networks' computational cost and a lack of understanding of how the brain wires its biological networks.{{Citation needed|date=August 2024}}

In 2003, LSTM became competitive with traditional speech recognizers on certain tasks.{{Cite web |last1=Graves |first1=Alex |last2=Eck |first2=Douglas |last3=Beringer |first3=Nicole |last4=Schmidhuber |first4=Jürgen |date=2003 |title=Biologically Plausible Speech Recognition with LSTM Neural Nets |url=ftp://ftp.idsia.ch/pub/juergen/bioadit2004.pdf |url-status=live |archive-url=https://web.archive.org/web/20210509123139/ftp://ftp.idsia.ch/pub/juergen/bioadit2004.pdf |archive-date=2021-05-09 |access-date=2016-04-09 |website=1st Intl. Workshop on Biologically Inspired Approaches to Advanced Information Technology, Bio-ADIT 2004, Lausanne, Switzerland |pages=175–184}} In 2006, Alex Graves, Santiago Fernández, Faustino Gomez, and Schmidhuber combined it with connectionist temporal classification (CTC){{Cite journal |last1=Graves |first1=Alex |author1-link=Alex Graves (computer scientist) |last2=Fernández |first2=Santiago |last3=Gomez |first3=Faustino |last4=Schmidhuber |first4=Jürgen |author4-link=Jürgen Schmidhuber |date=2006 |title=Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks |journal=Proceedings of the International Conference on Machine Learning, ICML 2006 |pages=369–376 |citeseerx=10.1.1.75.6306}} in stacks of LSTMs.Santiago Fernandez, Alex Graves, and Jürgen Schmidhuber (2007). [https://mediatum.ub.tum.de/doc/1289941/file.pdf An application of recurrent neural networks to discriminative keyword spotting] {{Webarchive|url=https://web.archive.org/web/20181118164457/https://mediatum.ub.tum.de/doc/1289941/file.pdf|date=2018-11-18}}. Proceedings of ICANN (2), pp. 220–229. In 2009, it became the first RNN to win a pattern recognition contest, in connected handwriting recognition.Graves, Alex; and Schmidhuber, Jürgen; Offline Handwriting Recognition with Multidimensional Recurrent Neural Networks, in Bengio, Yoshua; Schuurmans, Dale; Lafferty, John; Williams, Chris K. I.; and Culotta, Aron (eds.), Advances in Neural Information Processing Systems 22 (NIPS'22), December 7th–10th, 2009, Vancouver, BC, Neural Information Processing Systems (NIPS) Foundation, 2009, pp. 545–552

In 2006, publications by Geoff Hinton, Ruslan Salakhutdinov, Osindero and Teh{{Cite journal |last=Hinton |first=Geoffrey E. |date=1 October 2007 |title=Learning multiple layers of representation |url=http://www.cell.com/trends/cognitive-sciences/abstract/S1364-6613(07)00217-3 |url-status=live |journal=Trends in Cognitive Sciences |volume=11 |issue=10 |pages=428–434 |doi=10.1016/j.tics.2007.09.004 |issn=1364-6613 |pmid=17921042 |s2cid=15066318 |archive-url=https://web.archive.org/web/20131011071435/http://www.cell.com/trends/cognitive-sciences/abstract/S1364-6613(07)00217-3 |archive-date=11 October 2013 |access-date=12 June 2017}}{{Cite journal |last1=Hinton |first1=G. E. |author-link1=Geoff Hinton |last2=Osindero |first2=S. |last3=Teh |first3=Y. W. |year=2006 |title=A Fast Learning Algorithm for Deep Belief Nets |url=http://www.cs.toronto.edu/~hinton/absps/fastnc.pdf |url-status=live |journal=Neural Computation |volume=18 |issue=7 |pages=1527–1554 |doi=10.1162/neco.2006.18.7.1527 |pmid=16764513 |s2cid=2309950 |archive-url=https://web.archive.org/web/20151223164129/http://www.cs.toronto.edu/~hinton/absps/fastnc.pdf |archive-date=2015-12-23 |access-date=2011-07-20}} deep belief networks were developed for generative modeling. They are trained by training one restricted Boltzmann machine, then freezing it and training another one on top of the first one, and so on, then optionally fine-tuned using supervised backpropagation.G. E. Hinton., "[http://www.csri.utoronto.ca/~hinton/absps/ticsdraft.pdf Learning multiple layers of representation]". {{Webarchive|url=https://web.archive.org/web/20180522112408/http://www.csri.utoronto.ca/~hinton/absps/ticsdraft.pdf|date=2018-05-22}}. Trends in Cognitive Sciences, 11, pp. 428–434, 2007. They could model high-dimensional probability distributions, such as the distribution of MNIST images, but convergence was slow.{{Cite journal |last=Hinton |first=Geoffrey E. |date=October 2007 |title=Learning multiple layers of representation |url=https://linkinghub.elsevier.com/retrieve/pii/S1364661307002173 |journal=Trends in Cognitive Sciences |language=en |volume=11 |issue=10 |pages=428–434 |doi=10.1016/j.tics.2007.09.004|pmid=17921042 }}{{Cite journal |last1=Hinton |first1=Geoffrey E. |last2=Osindero |first2=Simon |last3=Teh |first3=Yee-Whye |date=July 2006 |title=A Fast Learning Algorithm for Deep Belief Nets |url=https://direct.mit.edu/neco/article/18/7/1527-1554/7065 |journal=Neural Computation |language=en |volume=18 |issue=7 |pages=1527–1554 |doi=10.1162/neco.2006.18.7.1527 |pmid=16764513 |issn=0899-7667}}{{Cite journal |last=Hinton |first=Geoffrey E. |date=2009-05-31 |title=Deep belief networks |journal=Scholarpedia |language=en |volume=4 |issue=5 |pages=5947 |doi=10.4249/scholarpedia.5947 |doi-access=free |bibcode=2009SchpJ...4.5947H |issn=1941-6016}}

The impact of deep learning in industry began in the early 2000s, when CNNs already processed an estimated 10% to 20% of all the checks written in the US, according to Yann LeCun.Yann LeCun (2016). Slides on Deep Learning [https://indico.cern.ch/event/510372/ Online] {{Webarchive|url=https://web.archive.org/web/20160423021403/https://indico.cern.ch/event/510372/|date=2016-04-23}} Industrial applications of deep learning to large-scale speech recognition started around 2010.

The 2009 NIPS Workshop on Deep Learning for Speech Recognition was motivated by the limitations of deep generative models of speech, and the possibility that given more capable hardware and large-scale data sets that deep neural nets might become practical. It was believed that pre-training DNNs using generative models of deep belief nets (DBN) would overcome the main difficulties of neural nets. However, it was discovered that replacing pre-training with large amounts of training data for straightforward backpropagation when using DNNs with large, context-dependent output layers produced error rates dramatically lower than then-state-of-the-art Gaussian mixture model (GMM)/Hidden Markov Model (HMM) and also than more-advanced generative model-based systems.{{cite journal | last1 = Hinton | first1 = G. | last2 = Deng | first2 = L. | last3 = Yu | first3 = D. | last4 = Dahl | first4 = G. | last5 = Mohamed | first5 = A. | last6 = Jaitly | first6 = N. | last7 = Senior | first7 = A. | last8 = Vanhoucke | first8 = V. | last9 = Nguyen | first9 = P. | last10 = Sainath | first10 = T. | author10-link = Tara Sainath | last11 = Kingsbury | first11 = B. | year = 2012 | title = Deep Neural Networks for Acoustic Modeling in Speech Recognition: The Shared Views of Four Research Groups| journal = IEEE Signal Processing Magazine | volume = 29 | issue = 6| pages = 82–97 | doi=10.1109/msp.2012.2205597 | bibcode = 2012ISPM...29...82H| s2cid = 206485943 }} The nature of the recognition errors produced by the two types of systems was characteristically different, offering technical insights into how to integrate deep learning into the existing highly efficient, run-time speech decoding system deployed by all major speech recognition systems.{{cite book|last2=Deng|first2=L.|date=2014|title=Automatic Speech Recognition: A Deep Learning Approach (Publisher: Springer)|url={{google books |plainurl=y |id=rUBTBQAAQBAJ}}|isbn=978-1-4471-5779-3|last1=Yu|first1=D.|publisher=Springer }}{{cite web|title=Deng receives prestigious IEEE Technical Achievement Award - Microsoft Research|url=https://www.microsoft.com/en-us/research/blog/deng-receives-prestigious-ieee-technical-achievement-award/|website=Microsoft Research|date=3 December 2015|access-date=16 March 2018|archive-date=16 March 2018|archive-url=https://web.archive.org/web/20180316084821/https://www.microsoft.com/en-us/research/blog/deng-receives-prestigious-ieee-technical-achievement-award/|url-status=live}} Analysis around 2009–2010, contrasting the GMM (and other generative speech models) vs. DNN models, stimulated early industrial investment in deep learning for speech recognition. That analysis was done with comparable performance (less than 1.5% in error rate) between discriminative DNNs and generative models.{{cite web| last1=Deng| first1=L.| last2=Hinton| first2=G.| last3=Kingsbury| first3=B.| date=May 2013| title=New types of deep neural network learning for speech recognition and related applications: An overview (ICASSP)| url=https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/ICASSP-2013-DengHintonKingsbury-revised.pdf| publisher=Microsoft| access-date=27 December 2023| archive-date=2017-09-26| archive-url=https://web.archive.org/web/20170926190920/https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/ICASSP-2013-DengHintonKingsbury-revised.pdf| url-status=live}}{{Cite web|url=https://www.superlectures.com/interspeech2014/downloadFile?id=6&type=slides&filename=achievements-and-challenges-of-deep-learning-from-speech-analysis-and-recognition-to-language-and-multimodal-processing|title=Keynote talk: 'Achievements and Challenges of Deep Learning - From Speech Analysis and Recognition To Language and Multimodal Processing'|last=Li|first=Deng|date=September 2014|website=Interspeech|access-date=2017-06-12|archive-date=2017-09-26|archive-url=https://web.archive.org/web/20170926190732/https://www.superlectures.com/interspeech2014/downloadFile?id=6&type=slides&filename=achievements-and-challenges-of-deep-learning-from-speech-analysis-and-recognition-to-language-and-multimodal-processing|url-status=live}}

In 2010, researchers extended deep learning from TIMIT to large vocabulary speech recognition, by adopting large output layers of the DNN based on context-dependent HMM states constructed by decision trees.{{cite journal|last1=Yu|first1=D.|last2=Deng|first2=L.|date=2010|title=Roles of Pre-Training and Fine-Tuning in Context-Dependent DBN-HMMs for Real-World Speech Recognition|url=https://www.microsoft.com/en-us/research/publication/roles-of-pre-training-and-fine-tuning-in-context-dependent-dbn-hmms-for-real-world-speech-recognition/|journal=NIPS Workshop on Deep Learning and Unsupervised Feature Learning|access-date=2017-06-14|archive-date=2017-10-12|archive-url=https://web.archive.org/web/20171012095148/https://www.microsoft.com/en-us/research/publication/roles-of-pre-training-and-fine-tuning-in-context-dependent-dbn-hmms-for-real-world-speech-recognition/|url-status=live}}{{Cite book|last1=Seide|first1=F.|last2=Li|first2=G.|last3=Yu|first3=D.|title=Interspeech 2011 |chapter=Conversational speech transcription using context-dependent deep neural networks |date=2011 |chapter-url=https://www.microsoft.com/en-us/research/publication/conversational-speech-transcription-using-context-dependent-deep-neural-networks |pages=437–440|doi=10.21437/Interspeech.2011-169|s2cid=398770 |access-date=2017-06-14|archive-date=2017-10-12|archive-url=https://web.archive.org/web/20171012095522/https://www.microsoft.com/en-us/research/publication/conversational-speech-transcription-using-context-dependent-deep-neural-networks/|url-status=live}}{{Cite journal|last1=Deng|first1=Li|last2=Li|first2=Jinyu|last3=Huang|first3=Jui-Ting|last4=Yao|first4=Kaisheng|last5=Yu|first5=Dong|last6=Seide|first6=Frank|last7=Seltzer|first7=Mike|last8=Zweig|first8=Geoff|last9=He|first9=Xiaodong|date=1 May 2013|title=Recent Advances in Deep Learning for Speech Research at Microsoft|url=https://www.microsoft.com/en-us/research/publication/recent-advances-in-deep-learning-for-speech-research-at-microsoft/|journal=Microsoft Research|access-date=14 June 2017|archive-date=12 October 2017|archive-url=https://web.archive.org/web/20171012044053/https://www.microsoft.com/en-us/research/publication/recent-advances-in-deep-learning-for-speech-research-at-microsoft/|url-status=live}}

File:AI-ML-DL.svg

The deep learning revolution started around CNN- and GPU-based computer vision.

Although CNNs trained by backpropagation had been around for decades and GPU implementations of NNs for years,{{cite journal |last1=Oh |first1=K.-S. |last2=Jung |first2=K. |year=2004 |title=GPU implementation of neural networks |journal=Pattern Recognition |volume=37 |issue=6 |pages=1311–1314 |bibcode=2004PatRe..37.1311O |doi=10.1016/j.patcog.2004.01.013}} including CNNs,{{Citation |last1=Chellapilla |first1=Kumar |title=High performance convolutional neural networks for document processing |date=2006 |url=https://hal.inria.fr/inria-00112631/document |access-date=2021-02-14 |archive-url=https://web.archive.org/web/20200518193413/https://hal.inria.fr/inria-00112631/document |archive-date=2020-05-18 |url-status=live |last2=Puri |first2=Sidd |last3=Simard |first3=Patrice}} faster implementations of CNNs on GPUs were needed to progress on computer vision. Later, as deep learning becomes widespread, specialized hardware and algorithm optimizations were developed specifically for deep learning.{{cite arXiv |eprint=1703.09039 |class=cs.CV |first1=Vivienne |last1=Sze |first2=Yu-Hsin |last2=Chen |author1-link=Vivienne Sze |title=Efficient Processing of Deep Neural Networks: A Tutorial and Survey |last3=Yang |first3=Tien-Ju |last4=Emer |first4=Joel |year=2017}}

A key advance for the deep learning revolution was hardware advances, especially GPU. Some early work dated back to 2004. In 2009, Raina, Madhavan, and Andrew Ng reported a 100M deep belief network trained on 30 Nvidia GeForce GTX 280 GPUs, an early demonstration of GPU-based deep learning. They reported up to 70 times faster training.{{Cite book |last1=Raina |first1=Rajat |last2=Madhavan |first2=Anand |last3=Ng |first3=Andrew Y. |chapter=Large-scale deep unsupervised learning using graphics processors |date=2009-06-14 |title=Proceedings of the 26th Annual International Conference on Machine Learning |chapter-url=https://doi.org/10.1145/1553374.1553486 |series=ICML '09 |location=New York, NY, USA |publisher=Association for Computing Machinery |pages=873–880 |doi=10.1145/1553374.1553486 |isbn=978-1-60558-516-1}}

In 2011, a CNN named DanNet{{Cite journal |last1=Cireşan |first1=Dan Claudiu |last2=Meier |first2=Ueli |last3=Gambardella |first3=Luca Maria |last4=Schmidhuber |first4=Jürgen |date=21 September 2010 |title=Deep, Big, Simple Neural Nets for Handwritten Digit Recognition |journal=Neural Computation |volume=22 |issue=12 |pages=3207–3220 |arxiv=1003.0358 |doi=10.1162/neco_a_00052 |issn=0899-7667 |pmid=20858131 |s2cid=1918673}}{{Cite journal|last1=Ciresan|first1=D. C.|last2=Meier|first2=U.|last3=Masci|first3=J.|last4=Gambardella|first4=L.M.|last5=Schmidhuber|first5=J.|date=2011|title=Flexible, High Performance Convolutional Neural Networks for Image Classification|url=http://ijcai.org/papers11/Papers/IJCAI11-210.pdf|journal=International Joint Conference on Artificial Intelligence|doi=10.5591/978-1-57735-516-8/ijcai11-210|access-date=2017-06-13|archive-date=2014-09-29|archive-url=https://web.archive.org/web/20140929094040/http://ijcai.org/papers11/Papers/IJCAI11-210.pdf|url-status=live}} by Dan Ciresan, Ueli Meier, Jonathan Masci, Luca Maria Gambardella, and Jürgen Schmidhuber achieved for the first time superhuman performance in a visual pattern recognition contest, outperforming traditional methods by a factor of 3. It then won more contests.{{Cite book|url=http://papers.nips.cc/paper/4741-deep-neural-networks-segment-neuronal-membranes-in-electron-microscopy-images.pdf|title=Advances in Neural Information Processing Systems 25|last1=Ciresan|first1=Dan|last2=Giusti|first2=Alessandro|last3=Gambardella|first3=Luca M.|last4=Schmidhuber|first4=Jürgen|date=2012|publisher=Curran Associates, Inc.|editor-last=Pereira|editor-first=F.|pages=2843–2851|editor-last2=Burges|editor-first2=C. J. C.|editor-last3=Bottou|editor-first3=L.|editor-last4=Weinberger|editor-first4=K. Q.|access-date=2017-06-13|archive-date=2017-08-09|archive-url=https://web.archive.org/web/20170809081713/http://papers.nips.cc/paper/4741-deep-neural-networks-segment-neuronal-membranes-in-electron-microscopy-images.pdf|url-status=live}}{{Cite book |last1=Ciresan |first1=D. |title=Medical Image Computing and Computer-Assisted Intervention – MICCAI 2013 |last2=Giusti |first2=A. |last3=Gambardella |first3=L.M. |last4=Schmidhuber |first4=J. |date=2013 |isbn=978-3-642-38708-1 |series=Lecture Notes in Computer Science |volume=7908 |pages=411–418 |chapter=Mitosis Detection in Breast Cancer Histology Images with Deep Neural Networks |doi=10.1007/978-3-642-40763-5_51 |pmid=24579167 |issue=Pt 2}} They also showed how max-pooling CNNs on GPU improved performance significantly.

In 2012, Andrew Ng and Jeff Dean created an FNN that learned to recognize higher-level concepts, such as cats, only from watching unlabeled images taken from YouTube videos.{{cite arXiv |eprint=1112.6209 |class=cs.LG |first1=Andrew |last1=Ng |first2=Jeff |last2=Dean |title=Building High-level Features Using Large Scale Unsupervised Learning |year=2012}}

In October 2012, AlexNet by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton won the large-scale ImageNet competition by a significant margin over shallow machine learning methods. Further incremental improvements included the VGG-16 network by Karen Simonyan and Andrew Zisserman{{cite arXiv|last1=Simonyan |first1=Karen |last2=Andrew |first2=Zisserman |title=Very Deep Convolution Networks for Large Scale Image Recognition |year=2014 |class=cs.CV |eprint=1409.1556}} and Google's Inceptionv3.{{Cite journal |last=Szegedy |first=Christian |date=2015 |title=Going deeper with convolutions |url=https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/43022.pdf |journal=Cvpr2015|arxiv=1409.4842 }}

The success in image classification was then extended to the more challenging task of generating descriptions (captions) for images, often as a combination of CNNs and LSTMs.{{cite arXiv |eprint=1411.4555 |class=cs.CV |first1=Oriol |last1=Vinyals |first2=Alexander |last2=Toshev |title=Show and Tell: A Neural Image Caption Generator |last3=Bengio |first3=Samy |last4=Erhan |first4=Dumitru |year=2014}}.{{cite arXiv |eprint=1411.4952 |class=cs.CV |first1=Hao |last1=Fang |first2=Saurabh |last2=Gupta |title=From Captions to Visual Concepts and Back |last3=Iandola |first3=Forrest |last4=Srivastava |first4=Rupesh |last5=Deng |first5=Li |last6=Dollár |first6=Piotr |last7=Gao |first7=Jianfeng |last8=He |first8=Xiaodong |last9=Mitchell |first9=Margaret |last10=Platt |first10=John C |last11=Lawrence Zitnick |first11=C |last12=Zweig |first12=Geoffrey |year=2014}}.{{cite arXiv |eprint=1411.2539 |class=cs.LG |first1=Ryan |last1=Kiros |first2=Ruslan |last2=Salakhutdinov |title=Unifying Visual-Semantic Embeddings with Multimodal Neural Language Models |last3=Zemel |first3=Richard S |year=2014}}.

In 2014, the state of the art was training “very deep neural network” with 20 to 30 layers.{{Citation |last1=Simonyan |first1=Karen |title=Very Deep Convolutional Networks for Large-Scale Image Recognition |date=2015-04-10 |arxiv=1409.1556 |last2=Zisserman |first2=Andrew}} Stacking too many layers led to a steep reduction in training accuracy,{{cite arXiv |eprint=1502.01852 |class=cs.CV |first1=Kaiming |last1=He |first2=Xiangyu |last2=Zhang |title=Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification |last3=Ren |first3=Shaoqing |last4=Sun |first4=Jian |year=2016}} known as the "degradation" problem.{{Cite conference |last1=He |first1=Kaiming |last2=Zhang |first2=Xiangyu |last3=Ren |first3=Shaoqing |last4=Sun |first4=Jian |date=10 Dec 2015 |title=Deep Residual Learning for Image Recognition |arxiv=1512.03385}} In 2015, two techniques were developed to train very deep networks: the Highway Network was published in May 2015, and the residual neural network (ResNet){{Cite conference |last1=He |first1=Kaiming |last2=Zhang |first2=Xiangyu |last3=Ren |first3=Shaoqing |last4=Sun |first4=Jian |date=2016 |title=Deep Residual Learning for Image Recognition |url=https://ieeexplore.ieee.org/document/7780459 |location=Las Vegas, NV, USA |publisher=IEEE |pages=770–778 |arxiv=1512.03385 |doi=10.1109/CVPR.2016.90 |isbn=978-1-4673-8851-1 |journal=2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)}} in Dec 2015. ResNet behaves like an open-gated Highway Net.

Around the same time, deep learning started impacting the field of art. Early examples included Google DeepDream (2015), and neural style transfer (2015),{{cite arXiv |eprint=1508.06576 |class=cs.CV |first1=Leon A. |last1=Gatys |first2=Alexander S. |last2=Ecker |title=A Neural Algorithm of Artistic Style |date=26 August 2015 |first3=Matthias |last3=Bethge}} both of which were based on pretrained image classification neural networks, such as VGG-19.

Generative adversarial network (GAN) by (Ian Goodfellow et al., 2014){{cite conference |last1=Goodfellow |first1=Ian |last2=Pouget-Abadie |first2=Jean |last3=Mirza |first3=Mehdi |last4=Xu |first4=Bing |last5=Warde-Farley |first5=David |last6=Ozair |first6=Sherjil |last7=Courville |first7=Aaron |last8=Bengio |first8=Yoshua |year=2014 |title=Generative Adversarial Networks |url=https://papers.nips.cc/paper/5423-generative-adversarial-nets.pdf |conference=Proceedings of the International Conference on Neural Information Processing Systems (NIPS 2014) |pages=2672–2680 |archive-url=https://web.archive.org/web/20191122034612/http://papers.nips.cc/paper/5423-generative-adversarial-nets.pdf |archive-date=22 November 2019 |access-date=20 August 2019 |url-status=live}} (based on Jürgen Schmidhuber's principle of artificial curiosity)

became state of the art in generative modeling during 2014-2018 period. Excellent image quality is achieved by Nvidia's StyleGAN (2018){{Cite web |date=December 14, 2018 |title=GAN 2.0: NVIDIA's Hyperrealistic Face Generator |url=https://syncedreview.com/2018/12/14/gan-2-0-nvidias-hyperrealistic-face-generator/ |access-date=October 3, 2019 |website=SyncedReview.com}} based on the Progressive GAN by Tero Karras et al.{{cite arXiv |eprint=1710.10196 |class=cs.NE |first1=T. |last1=Karras |first2=T. |last2=Aila |title=Progressive Growing of GANs for Improved Quality, Stability, and Variation |date=26 February 2018 |last3=Laine |first3=S. |last4=Lehtinen |first4=J.}} Here the GAN generator is grown from small to large scale in a pyramidal fashion. Image generation by GAN reached popular success, and provoked discussions concerning deepfakes.{{Cite web |title=Prepare, Don't Panic: Synthetic Media and Deepfakes |url=https://lab.witness.org/projects/synthetic-media-and-deep-fakes/ |url-status=live |archive-url=https://web.archive.org/web/20201202231744/https://lab.witness.org/projects/synthetic-media-and-deep-fakes/ |archive-date=2 December 2020 |access-date=25 November 2020 |publisher=witness.org}} Diffusion models (2015){{Cite journal |last1=Sohl-Dickstein |first1=Jascha |last2=Weiss |first2=Eric |last3=Maheswaranathan |first3=Niru |last4=Ganguli |first4=Surya |date=2015-06-01 |title=Deep Unsupervised Learning using Nonequilibrium Thermodynamics |url=http://proceedings.mlr.press/v37/sohl-dickstein15.pdf |journal=Proceedings of the 32nd International Conference on Machine Learning |language=en |publisher=PMLR |volume=37 |pages=2256–2265|arxiv=1503.03585 }} eclipsed GANs in generative modeling since then, with systems such as DALL·E 2 (2022) and Stable Diffusion (2022).

In 2015, Google's speech recognition improved by 49% by an LSTM-based model, which they made available through Google Voice Search on smartphone.Google Research Blog. The neural networks behind Google Voice transcription. August 11, 2015. By Françoise Beaufays http://googleresearch.blogspot.co.at/2015/08/the-neural-networks-behind-google-voice.html{{Cite web |last1=Sak |first1=Haşim |last2=Senior |first2=Andrew |last3=Rao |first3=Kanishka |last4=Beaufays |first4=Françoise |last5=Schalkwyk |first5=Johan |date=September 2015 |title=Google voice search: faster and more accurate |url=http://googleresearch.blogspot.ch/2015/09/google-voice-search-faster-and-more.html |url-status=live |archive-url=https://web.archive.org/web/20160309191532/http://googleresearch.blogspot.ch/2015/09/google-voice-search-faster-and-more.html |archive-date=2016-03-09 |access-date=2016-04-09}}

Deep learning is part of state-of-the-art systems in various disciplines, particularly computer vision and automatic speech recognition (ASR). Results on commonly used evaluation sets such as TIMIT (ASR) and MNIST (image classification), as well as a range of large-vocabulary speech recognition tasks have steadily improved.{{Cite book |last1=Singh |first1=Premjeet |title=2021 International Conference on Computer Communication and Informatics (ICCCI) |last2=Saha |first2=Goutam |last3=Sahidullah |first3=Md |year=2021 |isbn=978-1-7281-5875-4 |pages=1–4 |chapter=Non-linear frequency warping using constant-Q transformation for speech emotion recognition |doi=10.1109/ICCCI50826.2021.9402569 |arxiv=2102.04029 |s2cid=231846518}} Convolutional neural networks were superseded for ASR by LSTM.{{Cite web |last1=Sak |first1=Hasim |last2=Senior |first2=Andrew |last3=Beaufays |first3=Francoise |date=2014 |title=Long Short-Term Memory recurrent neural network architectures for large scale acoustic modeling |url=https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/43905.pdf |url-status=dead |archive-url=https://web.archive.org/web/20180424203806/https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/43905.pdf |archive-date=24 April 2018}}{{cite arXiv |eprint=1410.4281 |class=cs.CL |first1=Xiangang |last1=Li |first2=Xihong |last2=Wu |title=Constructing Long Short-Term Memory based Deep Recurrent Neural Networks for Large Vocabulary Speech Recognition |year=2014}}{{Cite web |last1=Zen |first1=Heiga |last2=Sak |first2=Hasim |date=2015 |title=Unidirectional Long Short-Term Memory Recurrent Neural Network with Recurrent Output Layer for Low-Latency Speech Synthesis |url=https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/43266.pdf |url-status=live |archive-url=https://web.archive.org/web/20210509123113/https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/43266.pdf |archive-date=2021-05-09 |access-date=2017-06-13 |website=Google.com |publisher=ICASSP |pages=4470–4474}} but are more successful in computer vision.

Yoshua Bengio, Geoffrey Hinton and Yann LeCun were awarded the 2018 Turing Award for "conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing".{{Cite web |title=2018 ACM A.M. Turing Award Laureates |url=https://awards.acm.org/about/2018-turing |access-date=2024-08-07 |website=awards.acm.org |language=en}}

{{Main|Artificial neural network}}

{{multiple image

| direction = horizontal

| total_width = 400

| footer =

| image1 = Simplified neural network training example.svg

| alt1 =

| caption1 = Simplified example of training a neural network in object detection: The network is trained by multiple images that are known to depict starfish and sea urchins, which are correlated with "nodes" that represent visual features. The starfish match with a ringed texture and a star outline, whereas most sea urchins match with a striped texture and oval shape. However, the instance of a ring textured sea urchin creates a weakly weighted association between them.

| image2 = Simplified neural network example.svg

| alt2 =

| caption2 = Subsequent run of the network on an input image (left):{{cite book|author=Ferrie, C., & Kaiser, S.|year=2019|title=Neural Networks for Babies|publisher=Sourcebooks|isbn=978-1492671206}} The network correctly detects the starfish. However, the weakly weighted association between ringed texture and sea urchin also confers a weak signal to the latter from one of two intermediate nodes. In addition, a shell that was not included in the training gives a weak signal for the oval shape, also resulting in a weak signal for the sea urchin output. These weak signals may result in a false positive result for sea urchin.
In reality, textures and outlines would not be represented by single nodes, but rather by associated weight patterns of multiple nodes.}}

Artificial neural networks (ANNs) or connectionist systems are computing systems inspired by the biological neural networks that constitute animal brains. Such systems learn (progressively improve their ability) to do tasks by considering examples, generally without task-specific programming. For example, in image recognition, they might learn to identify images that contain cats by analyzing example images that have been manually labeled as "cat" or "no cat" and using the analytic results to identify cats in other images. They have found most use in applications difficult to express with a traditional computer algorithm using rule-based programming.

An ANN is based on a collection of connected units called artificial neurons, (analogous to biological neurons in a biological brain). Each connection (synapse) between neurons can transmit a signal to another neuron. The receiving (postsynaptic) neuron can process the signal(s) and then signal downstream neurons connected to it. Neurons may have state, generally represented by real numbers, typically between 0 and 1. Neurons and synapses may also have a weight that varies as learning proceeds, which can increase or decrease the strength of the signal that it sends downstream.

Typically, neurons are organized in layers. Different layers may perform different kinds of transformations on their inputs. Signals travel from the first (input), to the last (output) layer, possibly after traversing the layers multiple times.

The original goal of the neural network approach was to solve problems in the same way that a human brain would. Over time, attention focused on matching specific mental abilities, leading to deviations from biology such as backpropagation, or passing information in the reverse direction and adjusting the network to reflect that information.

Neural networks have been used on a variety of tasks, including computer vision, speech recognition, machine translation, social network filtering, playing board and video games and medical diagnosis.

As of 2017, neural networks typically have a few thousand to a few million units and millions of connections. Despite this number being several order of magnitude less than the number of neurons on a human brain, these networks can perform many tasks at a level beyond that of humans (e.g., recognizing faces, or playing "Go"{{Cite journal|last1=Silver|first1=David|last2=Huang|first2=Aja|last3=Maddison|first3=Chris J.|last4=Guez|first4=Arthur|last5=Sifre|first5=Laurent|last6=Driessche|first6=George van den|last7=Schrittwieser|first7=Julian|last8=Antonoglou|first8=Ioannis|last9=Panneershelvam|first9=Veda|s2cid=515925|date=January 2016|title=Mastering the game of Go with deep neural networks and tree search|journal=Nature|volume=529|issue=7587|pages=484–489|doi=10.1038/nature16961|issn=1476-4687|pmid=26819042|bibcode=2016Natur.529..484S}}).

A deep neural network (DNN) is an artificial neural network with multiple layers between the input and output layers. There are different types of neural networks but they always consist of the same components: neurons, synapses, weights, biases, and functions.{{citation |title=A Guide to Deep Learning and Neural Networks |url=https://serokell.io/blog/deep-learning-and-neural-network-guide#components-of-neural-networks |access-date=2020-11-16 |archive-date=2020-11-02 |archive-url=https://web.archive.org/web/20201102151103/https://serokell.io/blog/deep-learning-and-neural-network-guide#components-of-neural-networks |url-status=live }} These components as a whole function in a way that mimics functions of the human brain, and can be trained like any other ML algorithm.{{Citation needed|date=November 2020}}

For example, a DNN that is trained to recognize dog breeds will go over the given image and calculate the probability that the dog in the image is a certain breed. The user can review the results and select which probabilities the network should display (above a certain threshold, etc.) and return the proposed label. Each mathematical manipulation as such is considered a layer, {{cite journal | last1 = Kumar | first1 = Nishant | last2 = Raubal | first2 = Martin | title = Applications of deep learning in congestion detection, prediction and alleviation: A survey | journal = Transportation Research Part C: Emerging Technologies | volume = 133 | pages = 103432 | year = 2021 | doi = 10.1016/j.trc.2021.103432 | hdl = 10230/42143 | s2cid = 240420107 | doi-access = free | arxiv = 2102.09759 | bibcode = 2021TRPC..13303432K | hdl-access = free }} and complex DNN have many layers, hence the name "deep" networks.

DNNs can model complex non-linear relationships. DNN architectures generate compositional models where the object is expressed as a layered composition of primitives.{{Cite journal|last1=Szegedy|first1=Christian|last2=Toshev|first2=Alexander|last3=Erhan|first3=Dumitru|date=2013|title=Deep neural networks for object detection|url=https://papers.nips.cc/paper/5207-deep-neural-networks-for-object-detection|journal=Advances in Neural Information Processing Systems|pages=2553–2561|access-date=2017-06-13|archive-date=2017-06-29|archive-url=https://web.archive.org/web/20170629172111/http://papers.nips.cc/paper/5207-deep-neural-networks-for-object-detection|url-status=live}} The extra layers enable composition of features from lower layers, potentially modeling complex data with fewer units than a similarly performing shallow network. For instance, it was proved that sparse multivariate polynomials are exponentially easier to approximate with DNNs than with shallow networks.{{cite conference|last1=Rolnick|first1=David|last2=Tegmark|first2=Max|date=2018|title=The power of deeper networks for expressing natural functions|url=https://openreview.net/pdf?id=SyProzZAW|conference=ICLR 2018|book-title=International Conference on Learning Representations|access-date=2021-01-05|archive-date=2021-01-07|archive-url=https://web.archive.org/web/20210107183647/https://openreview.net/pdf?id=SyProzZAW|url-status=live}}

Deep architectures include many variants of a few basic approaches. Each architecture has found success in specific domains. It is not always possible to compare the performance of multiple architectures, unless they have been evaluated on the same data sets.

DNNs are typically feedforward networks in which data flows from the input layer to the output layer without looping back. At first, the DNN creates a map of virtual neurons and assigns random numerical values, or "weights", to connections between them. The weights and inputs are multiplied and return an output between 0 and 1. If the network did not accurately recognize a particular pattern, an algorithm would adjust the weights.{{Cite news|last=Hof|first=Robert D.|title=Is Artificial Intelligence Finally Coming into Its Own?|work=MIT Technology Review|url=https://www.technologyreview.com/s/513696/deep-learning/|access-date=10 July 2018|archive-url=https://web.archive.org/web/20190331092832/https://www.technologyreview.com/s/513696/deep-learning/|archive-date=31 March 2019}} That way the algorithm can make certain parameters more influential, until it determines the correct mathematical manipulation to fully process the data.

Recurrent neural networks, in which data can flow in any direction, are used for applications such as language modeling.{{cite journal|last1=Gers|first1=Felix A.|last2=Schmidhuber|first2=Jürgen|year=2001|title=LSTM Recurrent Networks Learn Simple Context Free and Context Sensitive Languages|url=http://elartu.tntu.edu.ua/handle/lib/30719|journal=IEEE Transactions on Neural Networks|volume=12|issue=6|pages=1333–1340|doi=10.1109/72.963769|pmid=18249962|s2cid=10192330 |access-date=2020-02-25|archive-date=2020-01-26|archive-url=https://web.archive.org/web/20200126045722/http://elartu.tntu.edu.ua/handle/lib/30719|url-status=live}}{{cite arXiv |eprint=1602.02410|last1=Jozefowicz|first1=Rafal|title=Exploring the Limits of Language Modeling|last2=Vinyals|first2=Oriol|last3=Schuster|first3=Mike|last4=Shazeer|first4=Noam|last5=Wu|first5=Yonghui|class=cs.CL|year=2016}}{{cite arXiv |eprint=1512.00103|last1=Gillick|first1=Dan|title=Multilingual Language Processing from Bytes|last2=Brunk|first2=Cliff|last3=Vinyals|first3=Oriol|last4=Subramanya|first4=Amarnag|class=cs.CL|year=2015}}{{Cite journal|last=Mikolov|first=T.|display-authors=etal|date=2010|title=Recurrent neural network based language model|url=http://www.fit.vutbr.cz/research/groups/speech/servite/2010/rnnlm_mikolov.pdf|journal=Interspeech|pages=1045–1048|doi=10.21437/Interspeech.2010-343|s2cid=17048224 |access-date=2017-06-13|archive-date=2017-05-16|archive-url=https://web.archive.org/web/20170516181940/http://www.fit.vutbr.cz/research/groups/speech/servite/2010/rnnlm_mikolov.pdf|url-status=live}} Long short-term memory is particularly effective for this use.{{Cite journal |last1=Hochreiter |first1=Sepp |last2=Schmidhuber |first2=Jürgen |date=1 November 1997 |title=Long Short-Term Memory |journal=Neural Computation |volume=9 |issue=8 |pages=1735–1780 |doi=10.1162/neco.1997.9.8.1735 |issn=0899-7667 |pmid=9377276 |s2cid=1915014}}{{Cite web|url=https://www.researchgate.net/publication/220320057|title=Learning Precise Timing with LSTM Recurrent Networks (PDF Download Available)|website=ResearchGate|access-date=13 June 2017|archive-date=9 May 2021|archive-url=https://web.archive.org/web/20210509123147/https://www.researchgate.net/publication/220320057_Learning_Precise_Timing_with_LSTM_Recurrent_Networks|url-status=live}}

Convolutional neural networks (CNNs) are used in computer vision.{{cite journal |last1=LeCun |first1=Y. |display-authors=etal |year= 1998|title=Gradient-based learning applied to document recognition |journal=Proceedings of the IEEE |volume=86 |issue=11 |pages=2278–2324 |doi=10.1109/5.726791|s2cid=14542261 |url=http://elartu.tntu.edu.ua/handle/lib/38369 }} CNNs also have been applied to acoustic modeling for automatic speech recognition (ASR).{{Cite book |doi=10.1109/icassp.2013.6639347|isbn=978-1-4799-0356-6|chapter=Deep convolutional neural networks for LVCSR|title=2013 IEEE International Conference on Acoustics, Speech and Signal Processing|pages=8614–8618|year=2013|last1=Sainath|first1=Tara N.|author1-link= Tara Sainath |last2=Mohamed|first2=Abdel-Rahman|last3=Kingsbury|first3=Brian|last4=Ramabhadran|first4=Bhuvana|author4-link=Bhuvana Ramabhadran|s2cid=13816461}}

As with ANNs, many issues can arise with naively trained DNNs. Two common issues are overfitting and computation time.

DNNs are prone to overfitting because of the added layers of abstraction, which allow them to model rare dependencies in the training data. Regularization methods such as Ivakhnenko's unit pruning or weight decay ($\backslash ell\_2$-regularization) or sparsity ($\backslash ell\_1$-regularization) can be applied during training to combat overfitting.{{Cite book |doi=10.1109/icassp.2013.6639349|isbn=978-1-4799-0356-6|arxiv=1212.0901|citeseerx=10.1.1.752.9151|chapter=Advances in optimizing recurrent networks|title=2013 IEEE International Conference on Acoustics, Speech and Signal Processing|pages=8624–8628|year=2013|last1=Bengio|first1=Yoshua|last2=Boulanger-Lewandowski|first2=Nicolas|last3=Pascanu|first3=Razvan|s2cid=12485056}} Alternatively dropout regularization randomly omits units from the hidden layers during training. This helps to exclude rare dependencies.{{Cite journal|last=Dahl|first=G.|display-authors=etal|date=2013|title=Improving DNNs for LVCSR using rectified linear units and dropout|url=http://www.cs.toronto.edu/~gdahl/papers/reluDropoutBN_icassp2013.pdf|journal=ICASSP|access-date=2017-06-13|archive-date=2017-08-12|archive-url=https://web.archive.org/web/20170812140509/http://www.cs.toronto.edu/~gdahl/papers/reluDropoutBN_icassp2013.pdf|url-status=live}} Another interesting recent development is research into models of just enough complexity through an estimation of the intrinsic complexity of the task being modelled. This approach has been successfully applied for multivariate time series prediction tasks such as traffic prediction.{{cite journal| last1 = Kumar | first1 = Nishant | last2 = Martin | first2 = Henry | last3 = Raubal | first3 = Martin | title = Enhancing Deep Learning-Based City-Wide Traffic Prediction Pipelines Through Complexity Analysis | journal = Data Science for Transportation | volume = 6 | pages = Article 24| year = 2024

| issue = 3 | doi = 10.1007/s42421-024-00109-x| doi-access = free| hdl = 20.500.11850/695425 | hdl-access = free }} Finally, data can be augmented via methods such as cropping and rotating such that smaller training sets can be increased in size to reduce the chances of overfitting.{{Cite web|url=https://www.coursera.org/learn/convolutional-neural-networks/lecture/AYzbX/data-augmentation|title=Data Augmentation - deeplearning.ai {{!}} Coursera|website=Coursera|access-date=30 November 2017|archive-date=1 December 2017|archive-url=https://web.archive.org/web/20171201032606/https://www.coursera.org/learn/convolutional-neural-networks/lecture/AYzbX/data-augmentation|url-status=live}}

DNNs must consider many training parameters, such as the size (number of layers and number of units per layer), the learning rate, and initial weights. Sweeping through the parameter space for optimal parameters may not be feasible due to the cost in time and computational resources. Various tricks, such as batching (computing the gradient on several training examples at once rather than individual examples){{Cite journal|last=Hinton|first=G. E.|date=2010|title=A Practical Guide to Training Restricted Boltzmann Machines|url=https://www.researchgate.net/publication/221166159|journal=Tech. Rep. UTML TR 2010-003|access-date=2017-06-13|archive-date=2021-05-09|archive-url=https://web.archive.org/web/20210509123211/https://www.researchgate.net/publication/221166159_A_brief_introduction_to_Weightless_Neural_Systems|url-status=live}} speed up computation. Large processing capabilities of many-core architectures (such as GPUs or the Intel Xeon Phi) have produced significant speedups in training, because of the suitability of such processing architectures for the matrix and vector computations.{{cite book|last1=You|first1=Yang|title=Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis on - SC '17|pages=1–12|last2=Buluç|first2=Aydın|last3=Demmel|first3=James|chapter=Scaling deep learning on GPU and knights landing clusters|chapter-url=https://dl.acm.org/citation.cfm?doid=3126908.3126912|publisher=SC '17, ACM|date=November 2017|access-date=5 March 2018|doi=10.1145/3126908.3126912|isbn=9781450351140|s2cid=8869270|url=http://www.escholarship.org/uc/item/6ch40821|archive-date=29 July 2020|archive-url=https://web.archive.org/web/20200729133850/https://escholarship.org/uc/item/6ch40821|url-status=live}}{{cite journal|last1=Viebke|first1=André|last2=Memeti|first2=Suejb|last3=Pllana|first3=Sabri|last4=Abraham|first4=Ajith|s2cid=14135321|title=CHAOS: a parallelization scheme for training convolutional neural networks on Intel Xeon Phi|journal=The Journal of Supercomputing|volume=75|pages=197–227|doi=10.1007/s11227-017-1994-x|arxiv=1702.07908|bibcode=2017arXiv170207908V|year=2019}}

Alternatively, engineers may look for other types of neural networks with more straightforward and convergent training algorithms. CMAC (cerebellar model articulation controller) is one such kind of neural network. It doesn't require learning rates or randomized initial weights. The training process can be guaranteed to converge in one step with a new batch of data, and the computational complexity of the training algorithm is linear with respect to the number of neurons involved.Ting Qin, et al. "A learning algorithm of CMAC based on RLS". Neural Processing Letters 19.1 (2004): 49-61.Ting Qin, et al. "[http://www-control.eng.cam.ac.uk/Homepage/papers/cued_control_997.pdf Continuous CMAC-QRLS and its systolic array]". {{Webarchive|url=https://web.archive.org/web/20181118122850/http://www-control.eng.cam.ac.uk/Homepage/papers/cued_control_997.pdf |date=2018-11-18 }}. Neural Processing Letters 22.1 (2005): 1-16.

Since the 2010s, advances in both machine learning algorithms and computer hardware have led to more efficient methods for training deep neural networks that contain many layers of non-linear hidden units and a very large output layer.{{cite web|last1=Research|first1=AI|title=Deep Neural Networks for Acoustic Modeling in Speech Recognition|url=http://airesearch.com/ai-research-papers/deep-neural-networks-for-acoustic-modeling-in-speech-recognition/|website=airesearch.com|access-date=23 October 2015|date=23 October 2015|archive-date=1 February 2016|archive-url=https://web.archive.org/web/20160201033801/http://airesearch.com/ai-research-papers/deep-neural-networks-for-acoustic-modeling-in-speech-recognition/|url-status=live}} By 2019, graphics processing units (GPUs), often with AI-specific enhancements, had displaced CPUs as the dominant method for training large-scale commercial cloud AI .{{cite news |title=GPUs Continue to Dominate the AI Accelerator Market for Now |url=https://www.informationweek.com/big-data/ai-machine-learning/gpus-continue-to-dominate-the-ai-accelerator-market-for-now/a/d-id/1336475 |access-date=11 June 2020 |work=InformationWeek |date=December 2019 |language=en |archive-date=10 June 2020 |archive-url=https://web.archive.org/web/20200610094310/https://www.informationweek.com/big-data/ai-machine-learning/gpus-continue-to-dominate-the-ai-accelerator-market-for-now/a/d-id/1336475 |url-status=live }} OpenAI estimated the hardware computation used in the largest deep learning projects from AlexNet (2012) to AlphaZero (2017) and found a 300,000-fold increase in the amount of computation required, with a doubling-time trendline of 3.4 months.{{cite news |last1=Ray |first1=Tiernan |title=AI is changing the entire nature of computation |url=https://www.zdnet.com/article/ai-is-changing-the-entire-nature-of-compute/ |access-date=11 June 2020 |work=ZDNet |date=2019 |language=en |archive-date=25 May 2020 |archive-url=https://web.archive.org/web/20200525144635/https://www.zdnet.com/article/ai-is-changing-the-entire-nature-of-compute/ |url-status=live }}{{cite web |title=AI and Compute |url=https://openai.com/blog/ai-and-compute/ |website=OpenAI |access-date=11 June 2020 |language=en |date=16 May 2018 |archive-date=17 June 2020 |archive-url=https://web.archive.org/web/20200617200602/https://openai.com/blog/ai-and-compute/ |url-status=live }}

Special electronic circuits called deep learning processors were designed to speed up deep learning algorithms. Deep learning processors include neural processing units (NPUs) in Huawei cellphones{{Cite web|url=https://consumer.huawei.com/en/press/news/2017/ifa2017-kirin970/|title=HUAWEI Reveals the Future of Mobile AI at IFA 2017 | HUAWEI Latest News | HUAWEI Global|website=consumer.huawei.com}} and cloud computing servers such as tensor processing units (TPU) in the Google Cloud Platform.{{Cite journal|last1=P|first1=JouppiNorman|last2=YoungCliff|last3=PatilNishant|last4=PattersonDavid|last5=AgrawalGaurav|last6=BajwaRaminder|last7=BatesSarah|last8=BhatiaSuresh|last9=BodenNan|last10=BorchersAl|last11=BoyleRick|date=2017-06-24|title=In-Datacenter Performance Analysis of a Tensor Processing Unit|journal=ACM SIGARCH Computer Architecture News|volume=45|issue=2|pages=1–12|language=EN|doi=10.1145/3140659.3080246|doi-access=free|arxiv=1704.04760}} Cerebras Systems has also built a dedicated system to handle large deep learning models, the CS-2, based on the largest processor in the industry, the second-generation Wafer Scale Engine (WSE-2).{{Cite web |last=Woodie |first=Alex |date=2021-11-01 |title=Cerebras Hits the Accelerator for Deep Learning Workloads |url=https://www.datanami.com/2021/11/01/cerebras-hits-the-accelerator-for-deep-learning-workloads/ |access-date=2022-08-03 |website=Datanami}}{{Cite web |date=2021-04-20 |title=Cerebras launches new AI supercomputing processor with 2.6 trillion transistors |url=https://venturebeat.com/2021/04/20/cerebras-systems-launches-new-ai-supercomputing-processor-with-2-6-trillion-transistors/ |access-date=2022-08-03 |website=VentureBeat |language=en-US}}

Atomically thin semiconductors are considered promising for energy-efficient deep learning hardware where the same basic device structure is used for both logic operations and data storage.

In 2020, Marega et al. published experiments with a large-area active channel material for developing logic-in-memory devices and circuits based on floating-gate field-effect transistors (FGFETs).{{cite journal|title=Logic-in-memory based on an atomically thin semiconductor|year=2020|doi=10.1038/s41586-020-2861-0|last1=Marega|first1=Guilherme Migliato|last2=Zhao|first2=Yanfei|last3=Avsar|first3=Ahmet|last4=Wang|first4=Zhenyu|last5=Tripati|first5=Mukesh|last6=Radenovic|first6=Aleksandra|last7=Kis|first7=Anras|journal=Nature|volume=587|issue=2|pages=72–77|pmid=33149289|pmc=7116757|bibcode=2020Natur.587...72M }}

In 2021, J. Feldmann et al. proposed an integrated photonic hardware accelerator for parallel convolutional processing.{{cite journal |title=Parallel convolutional processing using an integrated photonic tensor |year=2021 |doi=10.1038/s41586-020-03070-1 |last1=Feldmann |first1=J. |last2=Youngblood|first2=N. |last3=Karpov |first3=M. | last4=Gehring |first4=H. | display-authors=3 | journal=Nature |volume=589 |issue=2 |pages=52–58|pmid=33408373 |arxiv=2002.00281 |s2cid=211010976 }} The authors identify two key advantages of integrated photonics over its electronic counterparts: (1) massively parallel data transfer through wavelength division multiplexing in conjunction with frequency combs, and (2) extremely high data modulation speeds. Their system can execute trillions of multiply-accumulate operations per second, indicating the potential of integrated photonics in data-heavy AI applications.

{{Main|Speech recognition}}

Large-scale automatic speech recognition is the first and most convincing successful case of deep learning. LSTM RNNs can learn "Very Deep Learning" tasks that involve multi-second intervals containing speech events separated by thousands of discrete time steps, where one time step corresponds to about 10 ms. LSTM with forget gates is competitive with traditional speech recognizers on certain tasks.

The initial success in speech recognition was based on small-scale recognition tasks based on TIMIT. The data set contains 630 speakers from eight major dialects of American English, where each speaker reads 10 sentences.{{cite book| title=TIMIT Acoustic-Phonetic Continuous Speech Corpus| author1=Garofolo, J.S.| author2=Lamel, L.F.| author3=Fisher, W.M.| author4=Fiscus, J.G.| author5=Pallett, D.S.| author6=Dahlgren, N.L.| author7=Zue, V.| url=https://catalog.ldc.upenn.edu/LDC93S1| publisher=Linguistic Data Consortium| date=1993| access-date=27 December 2023| doi=10.35111/17gk-bn40| isbn=1-58563-019-5}} Its small size lets many configurations be tried. More importantly, the TIMIT task concerns phone-sequence recognition, which, unlike word-sequence recognition, allows weak phone bigram language models. This lets the strength of the acoustic modeling aspects of speech recognition be more easily analyzed. The error rates listed below, including these early results and measured as percent phone error rates (PER), have been summarized since 1991.

The debut of DNNs for speaker recognition in the late 1990s and speech recognition around 2009-2011 and of LSTM around 2003–2007, accelerated progress in eight major areas:

• Scale-up/out and accelerated DNN training and decoding
• Sequence discriminative training
• Feature processing by deep models with solid understanding of the underlying mechanisms
• Adaptation of DNNs and related deep models
• Multi-task and transfer learning by DNNs and related deep models
• CNNs and how to design them to best exploit domain knowledge of speech
• RNN and its rich LSTM variants
• Other types of deep models including tensor-based models and integrated deep generative/discriminative models.

All major commercial speech recognition systems (e.g., Microsoft Cortana, Xbox, Skype Translator, Amazon Alexa, Google Now, Apple Siri, Baidu and iFlyTek voice search, and a range of Nuance speech products, etc.) are based on deep learning.{{Cite magazine|url=https://www.wired.com/2014/12/skype-used-ai-build-amazing-new-language-translator/|title=How Skype Used AI to Build Its Amazing New Language Translator {{!}} WIRED|magazine=Wired|access-date=14 June 2017|date=17 December 2014|last1=McMillan|first1=Robert|archive-date=8 June 2017|archive-url=https://web.archive.org/web/20170608062106/https://www.wired.com/2014/12/skype-used-ai-build-amazing-new-language-translator/|url-status=live}}{{cite arXiv |eprint=1412.5567|last1=Hannun|first1=Awni|title=Deep Speech: Scaling up end-to-end speech recognition|last2=Case|first2=Carl|last3=Casper|first3=Jared|last4=Catanzaro|first4=Bryan|last5=Diamos|first5=Greg|last6=Elsen|first6=Erich|last7=Prenger|first7=Ryan|last8=Satheesh|first8=Sanjeev|last9=Sengupta|first9=Shubho|last10=Coates|first10=Adam|last11=Ng|first11=Andrew Y|class=cs.CL|year=2014}}

{{Main|Computer vision}}

File:Deep Learning in mussel farming.webm in mussel aquaculture.]]

A common evaluation set for image classification is the MNIST database data set. MNIST is composed of handwritten digits and includes 60,000 training examples and 10,000 test examples. As with TIMIT, its small size lets users test multiple configurations. A comprehensive list of results on this set is available.{{cite web|url=http://yann.lecun.com/exdb/mnist/.|title=MNIST handwritten digit database, Yann LeCun, Corinna Cortes and Chris Burges|website=yann.lecun.com|access-date=2014-01-28|archive-date=2014-01-13|archive-url=https://web.archive.org/web/20140113175237/http://yann.lecun.com/exdb/mnist/|url-status=live}}

Deep learning-based image recognition has become "superhuman", producing more accurate results than human contestants. This first occurred in 2011 in recognition of traffic signs, and in 2014, with recognition of human faces.{{Cite journal|last1=Cireşan|first1=Dan|last2=Meier|first2=Ueli|last3=Masci|first3=Jonathan|last4=Schmidhuber|first4=Jürgen|date=August 2012|title=Multi-column deep neural network for traffic sign classification|journal=Neural Networks|series=Selected Papers from IJCNN 2011|volume=32|pages=333–338|doi=10.1016/j.neunet.2012.02.023|pmid=22386783|citeseerx=10.1.1.226.8219}}{{cite arXiv|title=Surpassing Human Level Face Recognition|author1=Chaochao Lu |author2= Xiaoou Tang |year=2014 |class=cs.CV |eprint=1404.3840 }}

Deep learning-trained vehicles now interpret 360° camera views.[http://www.technologyreview.com/news/533936/nvidia-demos-a-car-computer-trained-with-deep-learning/ Nvidia Demos a Car Computer Trained with "Deep Learning"] (6 January 2015), David Talbot, MIT Technology Review Another example is Facial Dysmorphology Novel Analysis (FDNA) used to analyze cases of human malformation connected to a large database of genetic syndromes.

File:Jimmy Wales in France, with the style of Munch's "The Scream" applied using neural style transfer.jpg" applied using neural style transfer]]

Closely related to the progress that has been made in image recognition is the increasing application of deep learning techniques to various visual art tasks. DNNs have proven themselves capable, for example, of

• identifying the style period of a given painting
• Neural Style Transfer{{snd}} capturing the style of a given artwork and applying it in a visually pleasing manner to an arbitrary photograph or video
• generating striking imagery based on random visual input fields.{{cite journal |author1=G. W. Smith|author2=Frederic Fol Leymarie|date=10 April 2017|title=The Machine as Artist: An Introduction|journal=Arts|volume=6|issue=4|pages=5|doi=10.3390/arts6020005|doi-access=free}}{{cite journal |author=Blaise Agüera y Arcas|date=29 September 2017|title=Art in the Age of Machine Intelligence|journal=Arts|volume=6|issue=4|pages=18|doi=10.3390/arts6040018|doi-access=free}}

{{Main|Natural language processing}}

Neural networks have been used for implementing language models since the early 2000s. LSTM helped to improve machine translation and language modeling.

Other key techniques in this field are negative sampling{{cite arXiv|last1=Goldberg|first1=Yoav|last2=Levy|first2=Omar|title=word2vec Explained: Deriving Mikolov et al.'s Negative-Sampling Word-Embedding Method|eprint=1402.3722|class=cs.CL|year=2014}} and word embedding. Word embedding, such as word2vec, can be thought of as a representational layer in a deep learning architecture that transforms an atomic word into a positional representation of the word relative to other words in the dataset; the position is represented as a point in a vector space. Using word embedding as an RNN input layer allows the network to parse sentences and phrases using an effective compositional vector grammar. A compositional vector grammar can be thought of as probabilistic context free grammar (PCFG) implemented by an RNN.{{cite web|last1=Socher|first1=Richard|last2=Manning|first2=Christopher|title=Deep Learning for NLP|url=http://nlp.stanford.edu/courses/NAACL2013/NAACL2013-Socher-Manning-DeepLearning.pdf|access-date=26 October 2014|archive-date=6 July 2014|archive-url=https://web.archive.org/web/20140706040227/http://nlp.stanford.edu/courses/NAACL2013/NAACL2013-Socher-Manning-DeepLearning.pdf|url-status=live}} Recursive auto-encoders built atop word embeddings can assess sentence similarity and detect paraphrasing. Deep neural architectures provide the best results for constituency parsing,{{Cite journal|url = http://aclweb.org/anthology/P/P13/P13-1045.pdf|title = Parsing With Compositional Vector Grammars|last1 = Socher|first1 = Richard|date = 2013|journal = Proceedings of the ACL 2013 Conference|last2 = Bauer|first2 = John|last3 = Manning|first3 = Christopher|last4 = Ng|first4 = Andrew|access-date = 2014-09-03|archive-date = 2014-11-27|archive-url = https://web.archive.org/web/20141127005912/http://www.aclweb.org/anthology/P/P13/P13-1045.pdf|url-status = live}} sentiment analysis,{{cite book| title=Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing| author1=Socher, R.| author2=Perelygin, A.| author3=Wu, J.| author4=Chuang, J.| author5=Manning, C.D.| author6=Ng, A.| author7=Potts, C.| chapter=Recursive Deep Models for Semantic Compositionality over a Sentiment Treebank| chapter-url=https://nlp.stanford.edu/~socherr/EMNLP2013_RNTN.pdf| publisher=Association for Computational Linguistics| date=October 2013| pages=1631–1642| doi=10.18653/v1/D13-1170| access-date=21 December 2023| url-status = live| archive-url=https://web.archive.org/web/20161228100300/http://nlp.stanford.edu/%7Esocherr/EMNLP2013_RNTN.pdf| archive-date=28 December 2016}} information retrieval,{{Cite journal| last1=Shen| first1=Yelong| last2=He| first2=Xiaodong| last3=Gao| first3=Jianfeng|last4=Deng|first4=Li|last5=Mesnil|first5=Gregoire|date=1 November 2014|title=A Latent Semantic Model with Convolutional-Pooling Structure for Information Retrieval|url=https://www.microsoft.com/en-us/research/publication/a-latent-semantic-model-with-convolutional-pooling-structure-for-information-retrieval/|journal=Microsoft Research|access-date=14 June 2017|archive-date=27 October 2017|archive-url=https://web.archive.org/web/20171027050418/https://www.microsoft.com/en-us/research/publication/a-latent-semantic-model-with-convolutional-pooling-structure-for-information-retrieval/|url-status=live}}{{Cite journal|last1=Huang|first1=Po-Sen|last2=He|first2=Xiaodong|last3=Gao|first3=Jianfeng|last4=Deng|first4=Li|last5=Acero|first5=Alex|last6=Heck|first6=Larry|date=1 October 2013|title=Learning Deep Structured Semantic Models for Web Search using Clickthrough Data|url=https://www.microsoft.com/en-us/research/publication/learning-deep-structured-semantic-models-for-web-search-using-clickthrough-data/|journal=Microsoft Research|access-date=14 June 2017|archive-date=27 October 2017|archive-url=https://web.archive.org/web/20171027050414/https://www.microsoft.com/en-us/research/publication/learning-deep-structured-semantic-models-for-web-search-using-clickthrough-data/|url-status=live}} spoken language understanding,{{cite journal | last1 = Mesnil | first1 = G. | last2 = Dauphin | first2 = Y. | last3 = Yao | first3 = K. | last4 = Bengio | first4 = Y. | last5 = Deng | first5 = L. | last6 = Hakkani-Tur | first6 = D.|author6-link=Dilek Hakkani-Tür | last7 = He | first7 = X. | last8 = Heck | first8 = L. | last9 = Tur | first9 = G. | last10 = Yu | first10 = D. | last11 = Zweig | first11 = G. | s2cid = 1317136 | year = 2015 | title = Using recurrent neural networks for slot filling in spoken language understanding | journal = IEEE Transactions on Audio, Speech, and Language Processing | volume = 23 | issue = 3| pages = 530–539 | doi=10.1109/taslp.2014.2383614}} machine translation,{{Cite journal|last1=Sutskever|first1=L.|last2=Vinyals|first2=O.|last3=Le|first3=Q.|date=2014|title=Sequence to Sequence Learning with Neural Networks|url=https://papers.nips.cc/paper/5346-sequence-to-sequence-learning-with-neural-networks.pdf|journal=Proc. NIPS|bibcode=2014arXiv1409.3215S|arxiv=1409.3215|access-date=2017-06-13|archive-date=2021-05-09|archive-url=https://web.archive.org/web/20210509123145/https://papers.nips.cc/paper/2014/file/a14ac55a4f27472c5d894ec1c3c743d2-Paper.pdf|url-status=live}}{{Cite journal|last1=Gao|first1=Jianfeng|last2=He|first2=Xiaodong|last3=Yih|first3=Scott Wen-tau|last4=Deng|first4=Li|date=1 June 2014|title=Learning Continuous Phrase Representations for Translation Modeling|url=https://www.microsoft.com/en-us/research/publication/learning-continuous-phrase-representations-for-translation-modeling/|journal=Microsoft Research|access-date=14 June 2017|archive-date=27 October 2017|archive-url=https://web.archive.org/web/20171027050403/https://www.microsoft.com/en-us/research/publication/learning-continuous-phrase-representations-for-translation-modeling/|url-status=live}} contextual entity linking, writing style recognition,{{Cite journal |doi = 10.1002/dac.3259|title = Authorship verification using deep belief network systems|journal = International Journal of Communication Systems|volume = 30|issue = 12|pages = e3259|year = 2017|last1 = Brocardo|first1 = Marcelo Luiz|last2 = Traore|first2 = Issa|last3 = Woungang|first3 = Isaac|last4 = Obaidat|first4 = Mohammad S.| s2cid=40745740 }} named-entity recognition (token classification),{{Cite journal |last1=Kariampuzha |first1=William |last2=Alyea |first2=Gioconda |last3=Qu |first3=Sue |last4=Sanjak |first4= Jaleal |last5=Mathé |first5=Ewy |last6=Sid |first6=Eric |last7= Chatelaine |first7=Haley |last8=Yadaw |first8=Arjun |last9=Xu |first9=Yanji |last10=Zhu |first10=Qian |date=2023 |title=Precision information extraction for rare disease epidemiology at scale |journal=Journal of Translational Medicine |language=en |volume=21 |issue=1 |page=157 |doi=10.1186/s12967-023-04011-y |pmid=36855134 |pmc=9972634 |doi-access=free }} text classification, and others.{{Cite news|url=https://www.microsoft.com/en-us/research/project/deep-learning-for-natural-language-processing-theory-and-practice-cikm2014-tutorial/|title=Deep Learning for Natural Language Processing: Theory and Practice (CIKM2014 Tutorial) - Microsoft Research|work=Microsoft Research|access-date=14 June 2017|archive-date=13 March 2017|archive-url=https://web.archive.org/web/20170313184253/https://www.microsoft.com/en-us/research/project/deep-learning-for-natural-language-processing-theory-and-practice-cikm2014-tutorial/|url-status=live}}

Recent developments generalize word embedding to sentence embedding.

Google Translate (GT) uses a large end-to-end long short-term memory (LSTM) network.{{cite web|url=https://blog.google/products/translate/found-translation-more-accurate-fluent-sentences-google-translate/|title=Found in translation: More accurate, fluent sentences in Google Translate|last=Turovsky|first=Barak|date=15 November 2016|website=The Keyword Google Blog|access-date=23 March 2017|archive-date=7 April 2017|archive-url=https://web.archive.org/web/20170407071226/https://blog.google/products/translate/found-translation-more-accurate-fluent-sentences-google-translate/|url-status=live}}{{cite web|url=https://research.googleblog.com/2016/11/zero-shot-translation-with-googles.html|title=Zero-Shot Translation with Google's Multilingual Neural Machine Translation System|last1=Schuster|first1=Mike|last2=Johnson|first2=Melvin|date=22 November 2016|website=Google Research Blog|access-date=23 March 2017|last3=Thorat|first3=Nikhil|archive-date=10 July 2017|archive-url=https://web.archive.org/web/20170710183732/https://research.googleblog.com/2016/11/zero-shot-translation-with-googles.html|url-status=live}}{{cite arXiv |eprint=1609.08144|last1=Wu|first1=Yonghui|title=Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation|last2=Schuster|first2=Mike|last3=Chen|first3=Zhifeng|last4=Le|first4=Quoc V|last5=Norouzi|first5=Mohammad|last6=Macherey|first6=Wolfgang|last7=Krikun|first7=Maxim|last8=Cao|first8=Yuan|last9=Gao|first9=Qin|last10=Macherey|first10=Klaus|last11=Klingner|first11=Jeff|last12=Shah|first12=Apurva|last13=Johnson|first13=Melvin|last14=Liu|first14=Xiaobing|last15=Kaiser|first15=Łukasz|last16=Gouws|first16=Stephan|last17=Kato|first17=Yoshikiyo|last18=Kudo|first18=Taku|last19=Kazawa|first19=Hideto|last20=Stevens|first20=Keith|last21=Kurian|first21=George|last22=Patil|first22=Nishant|last23=Wang|first23=Wei|last24=Young|first24=Cliff|last25=Smith|first25=Jason|last26=Riesa|first26=Jason|last27=Rudnick|first27=Alex|last28=Vinyals|first28=Oriol|last29=Corrado|first29=Greg|last30=Hughes|first30=Macduff|display-authors=29|class=cs.CL|year=2016}}{{cite magazine |title=An Infusion of AI Makes Google Translate More Powerful Than Ever |first=Cade |last=Metz |magazine=Wired |date=27 September 2016 |url=https://www.wired.com/2016/09/google-claims-ai-breakthrough-machine-translation/ |access-date=12 October 2017 |archive-date=8 November 2020 |archive-url=https://web.archive.org/web/20201108101324/https://www.wired.com/2016/09/google-claims-ai-breakthrough-machine-translation/ |url-status=live }} Google Neural Machine Translation (GNMT) uses an example-based machine translation method in which the system "learns from millions of examples". It translates "whole sentences at a time, rather than pieces". Google Translate supports over one hundred languages. The network encodes the "semantics of the sentence rather than simply memorizing phrase-to-phrase translations".{{cite web|url=http://www-clips.imag.fr/geta/herve.blanchon/Pdfs/NLP-KE-10.pdf|title=MT on and for the Web|last1=Boitet|first1=Christian|last2=Blanchon|first2=Hervé|date=2010|access-date=1 December 2016|last3=Seligman|first3=Mark|last4=Bellynck|first4=Valérie|archive-date=29 March 2017|archive-url=https://web.archive.org/web/20170329125916/http://www-clips.imag.fr/geta/herve.blanchon/Pdfs/NLP-KE-10.pdf|url-status=dead}} GT uses English as an intermediate between most language pairs.

{{For|more information|Drug discovery|Toxicology}}

A large percentage of candidate drugs fail to win regulatory approval. These failures are caused by insufficient efficacy (on-target effect), undesired interactions (off-target effects), or unanticipated toxic effects.{{Cite journal

| pmid = 23903212

| year = 2013

| last1 = Arrowsmith

| first1 = J

| title = Trial watch: Phase II and phase III attrition rates 2011-2012

| journal = Nature Reviews Drug Discovery

| volume = 12

| issue = 8

| pages = 569

| last2 = Miller

| first2 = P

| s2cid = 20246434

| doi = 10.1038/nrd4090

| doi-access = free

}}{{Cite journal

| pmid = 25582842

| year = 2015

| last1 = Verbist

| first1 = B

| title = Using transcriptomics to guide lead optimization in drug discovery projects: Lessons learned from the QSTAR project

| journal = Drug Discovery Today

| last2 = Klambauer

| first2 = G

| last3 = Vervoort

| first3 = L

| last4 = Talloen

| first4 = W

| last5 = The Qstar

| first5 = Consortium

| last6 = Shkedy

| first6 = Z

| last7 = Thas

| first7 = O

| last8 = Bender

| first8 = A

| last9 = Göhlmann

| first9 = H. W.

| last10 = Hochreiter

| first10 = S

| doi = 10.1016/j.drudis.2014.12.014

| volume=20

| issue = 5

| pages=505–513

| doi-access = free

| hdl = 1942/18723

| hdl-access = free

}} Research has explored use of deep learning to predict the biomolecular targets,{{Cite web |title=Merck Molecular Activity Challenge |url=https://kaggle.com/c/MerckActivity |url-status=live |archive-url=https://web.archive.org/web/20200716190808/https://www.kaggle.com/c/MerckActivity |archive-date=2020-07-16 |access-date=2020-07-16 |website=kaggle.com}}{{Cite web |title=Multi-task Neural Networks for QSAR Predictions {{!}} Data Science Association |url=http://www.datascienceassn.org/content/multi-task-neural-networks-qsar-predictions |url-status=live |archive-url=https://web.archive.org/web/20170430142049/http://www.datascienceassn.org/content/multi-task-neural-networks-qsar-predictions |archive-date=30 April 2017 |access-date=14 June 2017 |website=www.datascienceassn.org}} off-targets, and toxic effects of environmental chemicals in nutrients, household products and drugs."Toxicology in the 21st century Data Challenge"{{cite web |title=NCATS Announces Tox21 Data Challenge Winners |url=https://tripod.nih.gov/tox21/challenge/leaderboard.jsp |url-status=live |archive-url=https://web.archive.org/web/20150908025122/https://tripod.nih.gov/tox21/challenge/leaderboard.jsp |archive-date=2015-09-08 |access-date=2015-03-05}}{{cite web |title=NCATS Announces Tox21 Data Challenge Winners |url=http://www.ncats.nih.gov/news-and-events/features/tox21-challenge-winners.html |url-status=dead |archive-url=https://web.archive.org/web/20150228225709/http://www.ncats.nih.gov/news-and-events/features/tox21-challenge-winners.html |archive-date=28 February 2015 |access-date=5 March 2015}}

AtomNet is a deep learning system for structure-based rational drug design.{{cite arXiv|title = AtomNet: A Deep Convolutional Neural Network for Bioactivity Prediction in Structure-based Drug Discovery|eprint= 1510.02855|date = 9 October 2015|first1 = Izhar|last1 = Wallach|first2 = Michael|last2 = Dzamba|first3 = Abraham|last3 = Heifets|class= cs.LG}} AtomNet was used to predict novel candidate biomolecules for disease targets such as the Ebola virus{{Cite news |title = Toronto startup has a faster way to discover effective medicines |url = https://www.theglobeandmail.com/report-on-business/small-business/starting-out/toronto-startup-has-a-faster-way-to-discover-effective-medicines/article25660419/ |website = The Globe and Mail |access-date = 9 November 2015 |archive-date = 20 October 2015 |archive-url = https://web.archive.org/web/20151020040115/http://www.theglobeandmail.com/report-on-business/small-business/starting-out/toronto-startup-has-a-faster-way-to-discover-effective-medicines/article25660419/ |url-status = live }} and multiple sclerosis.{{Cite web|title = Startup Harnesses Supercomputers to Seek Cures|url = http://ww2.kqed.org/futureofyou/2015/05/27/startup-harnesses-supercomputers-to-seek-cures/|website = KQED Future of You| date=27 May 2015 |access-date = 9 November 2015|archive-date = 24 December 2015|archive-url = https://web.archive.org/web/20151224104721/http://ww2.kqed.org/futureofyou/2015/05/27/startup-harnesses-supercomputers-to-seek-cures/|url-status = live}}

In 2017 graph neural networks were used for the first time to predict various properties of molecules in a large toxicology data set.{{cite arXiv|last1=Gilmer|first1=Justin|last2=Schoenholz|first2=Samuel S.|last3=Riley|first3=Patrick F.|last4=Vinyals|first4=Oriol|last5=Dahl|first5=George E.|date=2017-06-12|title=Neural Message Passing for Quantum Chemistry|class=cs.LG|eprint=1704.01212}} In 2019, generative neural networks were used to produce molecules that were validated experimentally all the way into mice.{{cite journal |last1=Zhavoronkov |first1=Alex|s2cid=201716327|date=2019|title=Deep learning enables rapid identification of potent DDR1 kinase inhibitors |journal=Nature Biotechnology |volume=37|issue=9|pages=1038–1040|doi=10.1038/s41587-019-0224-x |pmid=31477924}}{{cite magazine |last1=Gregory |first1=Barber |title=A Molecule Designed By AI Exhibits 'Druglike' Qualities |url=https://www.wired.com/story/molecule-designed-ai-exhibits-druglike-qualities/ |magazine=Wired |access-date=2019-09-05 |archive-date=2020-04-30 |archive-url=https://web.archive.org/web/20200430143244/https://www.wired.com/story/molecule-designed-ai-exhibits-druglike-qualities/ |url-status=live }}

{{Main|Customer relationship management}}

Deep reinforcement learning has been used to approximate the value of possible direct marketing actions, defined in terms of RFM variables. The estimated value function was shown to have a natural interpretation as customer lifetime value.{{cite arXiv|last=Tkachenko |first=Yegor |title=Autonomous CRM Control via CLV Approximation with Deep Reinforcement Learning in Discrete and Continuous Action Space |date=8 April 2015 |eprint=1504.01840|class=cs.LG }}

{{Main|Recommender system}}

Recommendation systems have used deep learning to extract meaningful features for a latent factor model for content-based music and journal recommendations.{{Cite book|url=http://papers.nips.cc/paper/5004-deep-content-based-music-recommendation.pdf|title=Advances in Neural Information Processing Systems 26|last1=van den Oord|first1=Aaron|last2=Dieleman|first2=Sander|last3=Schrauwen|first3=Benjamin|date=2013|publisher=Curran Associates, Inc.|editor-last=Burges|editor-first=C. J. C.|pages=2643–2651|editor-last2=Bottou|editor-first2=L.|editor-last3=Welling|editor-first3=M.|editor-last4=Ghahramani|editor-first4=Z.|editor-last5=Weinberger|editor-first5=K. Q.|access-date=2017-06-14|archive-date=2017-05-16|archive-url=https://web.archive.org/web/20170516185259/http://papers.nips.cc/paper/5004-deep-content-based-music-recommendation.pdf|url-status=live}}{{cite journal | last1 = Feng | first1 = X.Y. | last2 = Zhang | first2 = H. | last3 = Ren | first3 = Y.J. | last4 = Shang | first4 = P.H. | last5 = Zhu | first5 = Y. | last6 = Liang | first6 = Y.C. | last7 = Guan | first7 = R.C. | last8 = Xu | first8 = D. | year = 2019 | title = The Deep Learning–Based Recommender System "Pubmender" for Choosing a Biomedical Publication Venue: Development and Validation Study | journal = Journal of Medical Internet Research | volume = 21 | issue = 5| page = e12957 | doi = 10.2196/12957 | pmid = 31127715 | pmc = 6555124 | doi-access = free }} Multi-view deep learning has been applied for learning user preferences from multiple domains.{{Cite journal|last1=Elkahky|first1=Ali Mamdouh|last2=Song|first2=Yang|last3=He|first3=Xiaodong|date=1 May 2015|title=A Multi-View Deep Learning Approach for Cross Domain User Modeling in Recommendation Systems|url=https://www.microsoft.com/en-us/research/publication/a-multi-view-deep-learning-approach-for-cross-domain-user-modeling-in-recommendation-systems/|journal=Microsoft Research|access-date=14 June 2017|archive-date=25 January 2018|archive-url=https://web.archive.org/web/20180125134534/https://www.microsoft.com/en-us/research/publication/a-multi-view-deep-learning-approach-for-cross-domain-user-modeling-in-recommendation-systems/|url-status=live}} The model uses a hybrid collaborative and content-based approach and enhances recommendations in multiple tasks.

{{Main|Bioinformatics}}

An autoencoder ANN was used in bioinformatics, to predict gene ontology annotations and gene-function relationships.{{cite book|first1=Davide|last1=Chicco|first2=Peter|last2=Sadowski|first3=Pierre|last3=Baldi|title=Proceedings of the 5th ACM Conference on Bioinformatics, Computational Biology, and Health Informatics |chapter=Deep autoencoder neural networks for gene ontology annotation predictions |s2cid=207217210|date=1 January 2014|publisher=ACM|pages=533–540|doi=10.1145/2649387.2649442|isbn=9781450328944|hdl=11311/964622|url=http://dl.acm.org/citation.cfm?id=2649442|access-date=23 November 2015|archive-date=9 May 2021|archive-url=https://web.archive.org/web/20210509123140/https://dl.acm.org/doi/10.1145/2649387.2649442|url-status=live}}

In medical informatics, deep learning was used to predict sleep quality based on data from wearables{{Cite journal|last=Sathyanarayana|first=Aarti|s2cid=3821594|date=1 January 2016|title=Sleep Quality Prediction From Wearable Data Using Deep Learning|journal=JMIR mHealth and uHealth|volume=4|issue=4|doi=10.2196/mhealth.6562|pmid=27815231|pmc=5116102|pages=e125 |doi-access=free }} and predictions of health complications from electronic health record data.{{Cite journal|last1=Choi|first1=Edward|last2=Schuetz|first2=Andy|last3=Stewart|first3=Walter F.|last4=Sun|first4=Jimeng|date=13 August 2016|title=Using recurrent neural network models for early detection of heart failure onset|journal=Journal of the American Medical Informatics Association|volume=24|issue=2|pages=361–370|doi=10.1093/jamia/ocw112|issn=1067-5027|pmid=27521897|pmc=5391725}}

Deep neural networks have shown unparalleled performance in predicting protein structure, according to the sequence of the amino acids that make it up. In 2020, AlphaFold, a deep-learning based system, achieved a level of accuracy significantly higher than all previous computational methods.{{Cite web |title=DeepMind's protein-folding AI has solved a 50-year-old grand challenge of biology |url=https://www.technologyreview.com/2020/11/30/1012712/deepmind-protein-folding-ai-solved-biology-science-drugs-disease/ |access-date=2024-05-10 |website=MIT Technology Review |language=en}}{{Cite web |last=Shead |first=Sam |date=2020-11-30 |title=DeepMind solves 50-year-old 'grand challenge' with protein folding A.I. |url=https://www.cnbc.com/2020/11/30/deepmind-solves-protein-folding-grand-challenge-with-alphafold-ai.html |access-date=2024-05-10 |website=CNBC |language=en}}

Deep neural networks can be used to estimate the entropy of a stochastic process and called Neural Joint Entropy Estimator (NJEE).{{Cite journal |last1=Shalev |first1=Y. |last2=Painsky |first2=A. |last3=Ben-Gal |first3=I. |title=Neural Joint Entropy Estimation |journal=IEEE Transactions on Neural Networks and Learning Systems |language=en |date=2022 |volume=PP |issue=4 |pages=5488–5500 |url=https://www.iradbengal.sites.tau.ac.il/_files/ugd/901879_d51bc0a620734585b5d3154488b3ae84.pdf |doi=10.1109/TNNLS.2022.3204919 |pmid=36155469 |arxiv=2012.11197 |s2cid=229339809 }} Such an estimation provides insights on the effects of input random variables on an independent random variable. Practically, the DNN is trained as a classifier that maps an input vector or matrix X to an output probability distribution over the possible classes of random variable Y, given input X. For example, in image classification tasks, the NJEE maps a vector of pixels' color values to probabilities over possible image classes. In practice, the probability distribution of Y is obtained by a Softmax layer with number of nodes that is equal to the alphabet size of Y. NJEE uses continuously differentiable activation functions, such that the conditions for the universal approximation theorem holds. It is shown that this method provides a strongly consistent estimator and outperforms other methods in case of large alphabet sizes.

Deep learning has been shown to produce competitive results in medical application such as cancer cell classification, lesion detection, organ segmentation and image enhancement.{{Cite journal|last1=Litjens|first1=Geert|last2=Kooi|first2=Thijs|last3=Bejnordi|first3=Babak Ehteshami|last4=Setio|first4=Arnaud Arindra Adiyoso|last5=Ciompi|first5=Francesco|last6=Ghafoorian|first6=Mohsen|last7=van der Laak|first7=Jeroen A.W.M.|last8=van Ginneken|first8=Bram|last9=Sánchez|first9=Clara I.|s2cid=2088679|date=December 2017|title=A survey on deep learning in medical image analysis|journal=Medical Image Analysis|volume=42|pages=60–88|doi=10.1016/j.media.2017.07.005|pmid=28778026|arxiv=1702.05747|bibcode=2017arXiv170205747L}}{{Cite book|doi=10.1109/ICCVW.2017.18|isbn=9781538610343|chapter=Deep Convolutional Neural Networks for Detecting Cellular Changes Due to Malignancy|title=2017 IEEE International Conference on Computer Vision Workshops (ICCVW)|pages=82–89|year=2017|last1=Forslid|first1=Gustav|last2=Wieslander|first2=Hakan|last3=Bengtsson|first3=Ewert|last4=Wahlby|first4=Carolina|last5=Hirsch|first5=Jan-Michael|last6=Stark|first6=Christina Runow|last7=Sadanandan|first7=Sajith Kecheril|s2cid=4728736|chapter-url=http://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-326160|access-date=2019-11-12|archive-date=2021-05-09|archive-url=https://web.archive.org/web/20210509123157/https://d1bxh8uas1mnw7.cloudfront.net/assets/embed.js|url-status=live}} Modern deep learning tools demonstrate the high accuracy of detecting various diseases and the helpfulness of their use by specialists to improve the diagnosis efficiency.{{Cite journal |last1=Dong |first1=Xin |last2=Zhou |first2=Yizhao |last3=Wang |first3=Lantian |last4=Peng |first4=Jingfeng |last5=Lou |first5=Yanbo |last6=Fan |first6=Yiqun |date=2020 |title=Liver Cancer Detection Using Hybridized Fully Convolutional Neural Network Based on Deep Learning Framework |journal=IEEE Access |volume=8 |pages=129889–129898 |doi=10.1109/ACCESS.2020.3006362 |bibcode=2020IEEEA...8l9889D |s2cid=220733699 |issn=2169-3536|doi-access=free }}{{Cite journal |last1=Lyakhov |first1=Pavel Alekseevich |last2=Lyakhova |first2=Ulyana Alekseevna |last3=Nagornov |first3=Nikolay Nikolaevich |date=2022-04-03 |title=System for the Recognizing of Pigmented Skin Lesions with Fusion and Analysis of Heterogeneous Data Based on a Multimodal Neural Network |journal=Cancers |language=en |volume=14 |issue=7 |pages=1819 |doi=10.3390/cancers14071819 |pmid=35406591 |pmc=8997449 |issn=2072-6694|doi-access=free }}

Finding the appropriate mobile audience for mobile advertising is always challenging, since many data points must be considered and analyzed before a target segment can be created and used in ad serving by any ad server.{{cite book |doi=10.1109/CSCITA.2017.8066548 |isbn=978-1-5090-4381-1|chapter=Predicting the popularity of instagram posts for a lifestyle magazine using deep learning|title=2017 2nd International Conference on Communication Systems, Computing and IT Applications (CSCITA)|pages=174–177|year=2017|last1=De|first1=Shaunak|last2=Maity|first2=Abhishek|last3=Goel|first3=Vritti|last4=Shitole|first4=Sanjay|last5=Bhattacharya|first5=Avik|s2cid=35350962}} Deep learning has been used to interpret large, many-dimensioned advertising datasets. Many data points are collected during the request/serve/click internet advertising cycle. This information can form the basis of machine learning to improve ad selection.

Deep learning has been successfully applied to inverse problems such as denoising, super-resolution, inpainting, and film colorization.{{Cite news|url=https://blog.floydhub.com/colorizing-and-restoring-old-images-with-deep-learning/|title=Colorizing and Restoring Old Images with Deep Learning|date=13 November 2018|website=FloydHub Blog|language=en|access-date=11 October 2019|archive-date=11 October 2019|archive-url=https://web.archive.org/web/20191011162814/https://blog.floydhub.com/colorizing-and-restoring-old-images-with-deep-learning/|url-status=live}} These applications include learning methods such as "Shrinkage Fields for Effective Image Restoration"{{cite conference |url= http://research.uweschmidt.org/pubs/cvpr14schmidt.pdf |first1= Uwe |last1= Schmidt |first2= Stefan |last2= Roth |conference= Computer Vision and Pattern Recognition (CVPR), 2014 IEEE Conference on |title= Shrinkage Fields for Effective Image Restoration |access-date= 2018-01-01 |archive-date= 2018-01-02 |archive-url= https://web.archive.org/web/20180102013217/http://research.uweschmidt.org/pubs/cvpr14schmidt.pdf |url-status= live }} which trains on an image dataset, and Deep Image Prior, which trains on the image that needs restoration.

Deep learning is being successfully applied to financial fraud detection, tax evasion detection,{{cite journal

|first1=Christos |last1=Kleanthous

|first2=Sotirios |last2=Chatzis

|title=Gated Mixture Variational Autoencoders for Value Added Tax audit case selection

|journal=Knowledge-Based Systems

|volume=188

|year=2020

|page=105048

|doi=10.1016/j.knosys.2019.105048

|s2cid=204092079

}} and anti-money laundering.{{cite journal

|first=Tomasz

|last=Czech

|title=Deep learning: the next frontier for money laundering detection

|url=https://www.globalbankingandfinance.com/deep-learning-the-next-frontier-for-money-laundering-detection/

|journal=Global Banking and Finance Review

|date=28 June 2018

|access-date=2018-07-15

|archive-date=2018-11-16

|archive-url=https://web.archive.org/web/20181116082711/https://www.globalbankingandfinance.com/deep-learning-the-next-frontier-for-money-laundering-detection/

|url-status=live

}}

In November 2023, researchers at Google DeepMind and Lawrence Berkeley National Laboratory announced that they had developed an AI system known as GNoME. This system has contributed to materials science by discovering over 2 million new materials within a relatively short timeframe. GNoME employs deep learning techniques to efficiently explore potential material structures, achieving a significant increase in the identification of stable inorganic crystal structures. The system's predictions were validated through autonomous robotic experiments, demonstrating a noteworthy success rate of 71%. The data of newly discovered materials is publicly available through the Materials Project database, offering researchers the opportunity to identify materials with desired properties for various applications. This development has implications for the future of scientific discovery and the integration of AI in material science research, potentially expediting material innovation and reducing costs in product development. The use of AI and deep learning suggests the possibility of minimizing or eliminating manual lab experiments and allowing scientists to focus more on the design and analysis of unique compounds.{{Cite web |last=Nuñez |first=Michael |date=2023-11-29 |title=Google DeepMind's materials AI has already discovered 2.2 million new crystals |url=https://venturebeat.com/ai/google-deepminds-materials-ai-has-already-discovered-2-2-million-new-crystals/ |access-date=2023-12-19 |website=VentureBeat |language=en-US}}{{Cite journal |last1=Merchant |first1=Amil |last2=Batzner |first2=Simon |last3=Schoenholz |first3=Samuel S. |last4=Aykol |first4=Muratahan |last5=Cheon |first5=Gowoon |last6=Cubuk |first6=Ekin Dogus |date=December 2023 |title=Scaling deep learning for materials discovery |journal=Nature |language=en |volume=624 |issue=7990 |pages=80–85 |doi=10.1038/s41586-023-06735-9 |issn=1476-4687|doi-access=free |pmid=38030720 |pmc=10700131 |bibcode=2023Natur.624...80M }}{{Cite journal |last=Peplow |first=Mark |date=2023-11-29 |title=Google AI and robots join forces to build new materials |url=https://www.nature.com/articles/d41586-023-03745-5 |journal=Nature |language=en |doi=10.1038/d41586-023-03745-5|pmid=38030771 |s2cid=265503872 }}

The United States Department of Defense applied deep learning to train robots in new tasks through observation.{{Cite web|url=https://www.eurekalert.org/pub_releases/2018-02/uarl-ard020218.php|title=Army researchers develop new algorithms to train robots|website=EurekAlert!|access-date=29 August 2018|archive-date=28 August 2018|archive-url=https://web.archive.org/web/20180828035608/https://www.eurekalert.org/pub_releases/2018-02/uarl-ard020218.php|url-status=live}}

Physics informed neural networks have been used to solve partial differential equations in both forward and inverse problems in a data driven manner.{{Cite journal|last1=Raissi|first1=M.|last2=Perdikaris|first2=P.|last3=Karniadakis|first3=G. E.|date=2019-02-01|title=Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations|journal=Journal of Computational Physics|language=en|volume=378|pages=686–707|doi=10.1016/j.jcp.2018.10.045|bibcode=2019JCoPh.378..686R|osti=1595805|s2cid=57379996|issn=0021-9991|doi-access=free}} One example is the reconstructing fluid flow governed by the Navier-Stokes equations. Using physics informed neural networks does not require the often expensive mesh generation that conventional CFD methods rely on.{{Cite journal|last1=Mao|first1=Zhiping|last2=Jagtap|first2=Ameya D.|last3=Karniadakis|first3=George Em|date=2020-03-01|title=Physics-informed neural networks for high-speed flows|journal=Computer Methods in Applied Mechanics and Engineering|language=en|volume=360|pages=112789|doi=10.1016/j.cma.2019.112789|bibcode=2020CMAME.360k2789M|s2cid=212755458|issn=0045-7825|doi-access=free}}{{Cite journal|last1=Raissi|first1=Maziar|last2=Yazdani|first2=Alireza|last3=Karniadakis|first3=George Em|date=2020-02-28|title=Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations|journal=Science|volume=367|issue=6481|pages=1026–1030|doi=10.1126/science.aaw4741|pmc=7219083|pmid=32001523|bibcode=2020Sci...367.1026R}}

Deep backward stochastic differential equation method is a numerical method that combines deep learning with Backward stochastic differential equation (BSDE). This method is particularly useful for solving high-dimensional problems in financial mathematics. By leveraging the powerful function approximation capabilities of deep neural networks, deep BSDE addresses the computational challenges faced by traditional numerical methods in high-dimensional settings. Specifically, traditional methods like finite difference methods or Monte Carlo simulations often struggle with the curse of dimensionality, where computational cost increases exponentially with the number of dimensions. Deep BSDE methods, however, employ deep neural networks to approximate solutions of high-dimensional partial differential equations (PDEs), effectively reducing the computational burden.{{cite journal | last1=Han | first1=J. | last2=Jentzen | first2=A. | last3=E | first3=W. | title=Solving high-dimensional partial differential equations using deep learning | journal=Proceedings of the National Academy of Sciences | volume=115 | issue=34 | pages=8505–8510 | year=2018 | doi=10.1073/pnas.1718942115 | doi-access=free | pmid=30082389 | pmc=6112690 | arxiv=1707.02568 | bibcode=2018PNAS..115.8505H }}

In addition, the integration of Physics-informed neural networks (PINNs) into the deep BSDE framework enhances its capability by embedding the underlying physical laws directly into the neural network architecture. This ensures that the solutions not only fit the data but also adhere to the governing stochastic differential equations. PINNs leverage the power of deep learning while respecting the constraints imposed by the physical models, resulting in more accurate and reliable solutions for financial mathematics problems.

Image reconstruction is the reconstruction of the underlying images from the image-related measurements. Several works showed the better and superior performance of the deep learning methods compared to analytical methods for various applications, e.g., spectral imaging {{Cite journal |last1=Oktem |first1=Figen S. |last2=Kar |first2=Oğuzhan Fatih |last3=Bezek |first3=Can Deniz |last4=Kamalabadi |first4=Farzad |date=2021 |title=High-Resolution Multi-Spectral Imaging With Diffractive Lenses and Learned Reconstruction |url=https://ieeexplore.ieee.org/document/9415140 |journal=IEEE Transactions on Computational Imaging |volume=7 |pages=489–504 |doi=10.1109/TCI.2021.3075349 |arxiv=2008.11625 |s2cid=235340737 |issn=2333-9403}} and ultrasound imaging.{{Cite journal |last1=Bernhardt |first1=Melanie |last2=Vishnevskiy |first2=Valery |last3=Rau |first3=Richard |last4=Goksel |first4=Orcun |date=December 2020 |title=Training Variational Networks With Multidomain Simulations: Speed-of-Sound Image Reconstruction |url=https://ieeexplore.ieee.org/document/9144249 |journal=IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control |volume=67 |issue=12 |pages=2584–2594 |doi=10.1109/TUFFC.2020.3010186 |pmid=32746211 |arxiv=2006.14395 |s2cid=220055785 |issn=1525-8955}}

Traditional weather prediction systems solve a very complex system of partial differential equations. GraphCast is a deep learning based model, trained on a long history of weather data to predict how weather patterns change over time. It is able to predict weather conditions for up to 10 days globally, at a very detailed level, and in under a minute, with precision similar to state of the art systems.{{Cite journal |last1=Lam |first1=Remi |last2=Sanchez-Gonzalez |first2=Alvaro |last3=Willson |first3=Matthew |last4=Wirnsberger |first4=Peter |last5=Fortunato |first5=Meire |last6=Alet |first6=Ferran |last7=Ravuri |first7=Suman |last8=Ewalds |first8=Timo |last9=Eaton-Rosen |first9=Zach |last10=Hu |first10=Weihua |last11=Merose |first11=Alexander |last12=Hoyer |first12=Stephan |last13=Holland |first13=George |last14=Vinyals |first14=Oriol |last15=Stott |first15=Jacklynn |date=2023-12-22 |title=Learning skillful medium-range global weather forecasting |journal=Science |language=en |volume=382 |issue=6677 |pages=1416–1421 |doi=10.1126/science.adi2336 |issn=0036-8075|doi-access=free |pmid=37962497 |arxiv=2212.12794 |bibcode=2023Sci...382.1416L }}{{Cite web |last=Sivakumar |first=Ramakrishnan |date=2023-11-27 |title=GraphCast: A breakthrough in Weather Forecasting |url=https://ramkrishna2910.medium.com/graphcast-a-breakthrough-in-weather-forecasting-d70fae9ac365 |access-date=2024-05-19 |website=Medium |language=en}}

{{Main|Epigenetic clock}}

An epigenetic clock is a biochemical test that can be used to measure age. Galkin et al. used deep neural networks to train an epigenetic aging clock of unprecedented accuracy using >6,000 blood samples.{{cite journal | last1 = Galkin | first1 = F. | last2 = Mamoshina | first2 = P. | last3 = Kochetov | first3 = K. | last4 = Sidorenko | first4 = D. | last5 = Zhavoronkov | first5 = A. | year = 2020 | title = DeepMAge: A Methylation Aging Clock Developed with Deep Learning | journal = Aging and Disease | doi = 10.14336/AD | doi-access = free }} The clock uses information from 1000 CpG sites and predicts people with certain conditions older than healthy controls: IBD, frontotemporal dementia, ovarian cancer, obesity. The aging clock was planned to be released for public use in 2021 by an Insilico Medicine spinoff company Deep Longevity.

Deep learning is closely related to a class of theories of brain development (specifically, neocortical development) proposed by cognitive neuroscientists in the early 1990s.{{cite journal | last1 = Utgoff | first1 = P. E. | last2 = Stracuzzi | first2 = D. J. | s2cid = 1119517 | year = 2002 | title = Many-layered learning | journal = Neural Computation | volume = 14 | issue = 10| pages = 2497–2529 | doi=10.1162/08997660260293319| pmid = 12396572 }}{{cite book|url={{google books |plainurl=y |id=vELaRu_MrwoC}}|title=Rethinking Innateness: A Connectionist Perspective on Development|last=Elman|first=Jeffrey L.|publisher=MIT Press|year=1998|isbn=978-0-262-55030-7}}{{cite journal | last1 = Shrager | first1 = J. | last2 = Johnson | first2 = MH | year = 1996 | title = Dynamic plasticity influences the emergence of function in a simple cortical array | journal = Neural Networks | volume = 9 | issue = 7| pages = 1119–1129 | doi=10.1016/0893-6080(96)00033-0| pmid = 12662587 }}{{cite journal | last1 = Quartz | first1 = SR | last2 = Sejnowski | first2 = TJ | year = 1997 | title = The neural basis of cognitive development: A constructivist manifesto | journal = Behavioral and Brain Sciences | volume = 20 | issue = 4| pages = 537–556 | doi=10.1017/s0140525x97001581| pmid = 10097006 | citeseerx = 10.1.1.41.7854 | s2cid = 5818342 }} These developmental theories were instantiated in computational models, making them predecessors of deep learning systems. These developmental models share the property that various proposed learning dynamics in the brain (e.g., a wave of nerve growth factor) support the self-organization somewhat analogous to the neural networks utilized in deep learning models. Like the neocortex, neural networks employ a hierarchy of layered filters in which each layer considers information from a prior layer (or the operating environment), and then passes its output (and possibly the original input), to other layers. This process yields a self-organizing stack of transducers, well-tuned to their operating environment. A 1995 description stated, "...the infant's brain seems to organize itself under the influence of waves of so-called trophic-factors ... different regions of the brain become connected sequentially, with one layer of tissue maturing before another and so on until the whole brain is mature".S. Blakeslee, "In brain's early growth, timetable may be critical", The New York Times, Science Section, pp. B5–B6, 1995.

A variety of approaches have been used to investigate the plausibility of deep learning models from a neurobiological perspective. On the one hand, several variants of the backpropagation algorithm have been proposed in order to increase its processing realism.{{Cite journal|last1=Mazzoni|first1=P.|last2=Andersen|first2=R. A.|last3=Jordan|first3=M. I.|date=15 May 1991|title=A more biologically plausible learning rule for neural networks.|journal=Proceedings of the National Academy of Sciences|volume=88|issue=10|pages=4433–4437|doi=10.1073/pnas.88.10.4433|issn=0027-8424|pmid=1903542|pmc=51674|bibcode=1991PNAS...88.4433M|doi-access=free}}{{Cite journal|last=O'Reilly|first=Randall C.|s2cid=2376781|date=1 July 1996|title=Biologically Plausible Error-Driven Learning Using Local Activation Differences: The Generalized Recirculation Algorithm|journal=Neural Computation|volume=8|issue=5|pages=895–938|doi=10.1162/neco.1996.8.5.895|issn=0899-7667}} Other researchers have argued that unsupervised forms of deep learning, such as those based on hierarchical generative models and deep belief networks, may be closer to biological reality.{{Cite journal|last1=Testolin|first1=Alberto|last2=Zorzi|first2=Marco|s2cid=9868901|date=2016|title=Probabilistic Models and Generative Neural Networks: Towards an Unified Framework for Modeling Normal and Impaired Neurocognitive Functions|journal=Frontiers in Computational Neuroscience|volume=10|pages=73|doi=10.3389/fncom.2016.00073|pmid=27468262|pmc=4943066|issn=1662-5188|doi-access=free}}{{Cite journal|last1=Testolin|first1=Alberto|last2=Stoianov|first2=Ivilin|last3=Zorzi|first3=Marco|s2cid=24504018|date=September 2017|title=Letter perception emerges from unsupervised deep learning and recycling of natural image features|journal=Nature Human Behaviour|volume=1|issue=9|pages=657–664|doi=10.1038/s41562-017-0186-2|pmid=31024135|issn=2397-3374}} In this respect, generative neural network models have been related to neurobiological evidence about sampling-based processing in the cerebral cortex.{{Cite journal|last1=Buesing|first1=Lars|last2=Bill|first2=Johannes|last3=Nessler|first3=Bernhard|last4=Maass|first4=Wolfgang|s2cid=7504633|date=3 November 2011|title=Neural Dynamics as Sampling: A Model for Stochastic Computation in Recurrent Networks of Spiking Neurons|journal=PLOS Computational Biology|volume=7|issue=11|pages=e1002211|doi=10.1371/journal.pcbi.1002211|pmid=22096452|pmc=3207943|issn=1553-7358|bibcode=2011PLSCB...7E2211B |doi-access=free }}

Although a systematic comparison between the human brain organization and the neuronal encoding in deep networks has not yet been established, several analogies have been reported. For example, the computations performed by deep learning units could be similar to those of actual neurons{{Cite journal|last1=Cash|first1=S.|last2=Yuste|first2=R.|date=February 1999|title=Linear summation of excitatory inputs by CA1 pyramidal neurons|journal=Neuron|volume=22|issue=2|pages=383–394|issn=0896-6273|pmid=10069343|doi=10.1016/s0896-6273(00)81098-3|s2cid=14663106|doi-access=free}} and neural populations.{{Cite journal|date=1 August 2004|title=Sparse coding of sensory inputs|journal=Current Opinion in Neurobiology|volume=14|issue=4|pages=481–487|doi=10.1016/j.conb.2004.07.007|pmid=15321069|issn=0959-4388 | last1 = Olshausen | first1 = B | last2 = Field | first2 = D|s2cid=16560320}} Similarly, the representations developed by deep learning models are similar to those measured in the primate visual system{{Cite journal|last1=Yamins|first1=Daniel L K|last2=DiCarlo|first2=James J|s2cid=16970545|date=March 2016|title=Using goal-driven deep learning models to understand sensory cortex|journal=Nature Neuroscience|volume=19|issue=3|pages=356–365|doi=10.1038/nn.4244|pmid=26906502|issn=1546-1726}} both at the single-unit{{Cite journal|last1=Zorzi|first1=Marco|last2=Testolin|first2=Alberto|s2cid=39281431|date=19 February 2018|title=An emergentist perspective on the origin of number sense|journal=Phil. Trans. R. Soc. B|volume=373|issue=1740|pages=20170043|doi=10.1098/rstb.2017.0043|issn=0962-8436|pmid=29292348|pmc=5784047}} and at the population{{Cite journal|last1=Güçlü|first1=Umut|last2=van Gerven|first2=Marcel A. J.|date=8 July 2015|title=Deep Neural Networks Reveal a Gradient in the Complexity of Neural Representations across the Ventral Stream|journal=Journal of Neuroscience|volume=35|issue=27|pages=10005–10014|doi=10.1523/jneurosci.5023-14.2015|pmid=26157000|pmc=6605414|arxiv=1411.6422}} levels.

Facebook's AI lab performs tasks such as automatically tagging uploaded pictures with the names of the people in them.{{cite magazine |first=C. |last=Metz |title=Facebook's 'Deep Learning' Guru Reveals the Future of AI |url=https://www.wired.com/wiredenterprise/2013/12/facebook-yann-lecun-qa/ |magazine=Wired |date=12 December 2013 |access-date=26 August 2017 |archive-date=28 March 2014 |archive-url=https://web.archive.org/web/20140328071226/http://www.wired.com/wiredenterprise/2013/12/facebook-yann-lecun-qa/ |url-status=live }}

Google's DeepMind Technologies developed a system capable of learning how to play Atari video games using only pixels as data input. In 2015 they demonstrated their AlphaGo system, which learned the game of Go well enough to beat a professional Go player.{{Cite journal|title = Google AI algorithm masters ancient game of Go|journal = Nature|year = 2016|doi = 10.1038/529445a|last1 = Gibney|first1 = Elizabeth|volume = 529|issue = 7587|pages = 445–446|pmid = 26819021|bibcode = 2016Natur.529..445G|s2cid = 4460235|doi-access = free}}{{Cite journal|title = Mastering the game of Go with deep neural networks and tree search|journal = Nature| issn= 0028-0836|pages = 484–489|volume = 529|issue = 7587|doi = 10.1038/nature16961|pmid = 26819042|first1 = David|last1 = Silver|author-link1=David Silver (programmer)|first2 = Aja|last2 = Huang|author-link2=Aja Huang|first3 = Chris J.|last3 = Maddison|first4 = Arthur|last4 = Guez|first5 = Laurent|last5 = Sifre|first6 = George van den|last6 = Driessche|first7 = Julian|last7 = Schrittwieser|first8 = Ioannis|last8 = Antonoglou|first9 = Veda|last9 = Panneershelvam|first10= Marc|last10= Lanctot|first11= Sander|last11= Dieleman|first12=Dominik|last12= Grewe|first13= John|last13= Nham|first14= Nal|last14= Kalchbrenner|first15= Ilya|last15= Sutskever|author-link15=Ilya Sutskever|first16= Timothy|last16= Lillicrap|first17= Madeleine|last17= Leach|first18= Koray|last18= Kavukcuoglu|first19= Thore|last19= Graepel|first20= Demis |last20=Hassabis|s2cid = 515925|author-link20=Demis Hassabis|date= 28 January 2016|bibcode = 2016Natur.529..484S}}{{closed access}}{{Cite web|title = A Google DeepMind Algorithm Uses Deep Learning and More to Master the Game of Go {{!}} MIT Technology Review|url = http://www.technologyreview.com/news/546066/googles-ai-masters-the-game-of-go-a-decade-earlier-than-expected/|website = MIT Technology Review|access-date = 30 January 2016|archive-date = 1 February 2016|archive-url = https://web.archive.org/web/20160201140636/http://www.technologyreview.com/news/546066/googles-ai-masters-the-game-of-go-a-decade-earlier-than-expected/|url-status = dead}} Google Translate uses a neural network to translate between more than 100 languages.

In 2017, Covariant.ai was launched, which focuses on integrating deep learning into factories.{{Cite news|url=https://www.nytimes.com/2017/11/06/technology/artificial-intelligence-start-up.html|title=A.I. Researchers Leave Elon Musk Lab to Begin Robotics Start-Up|first=Cade|last=Metz|newspaper=The New York Times|date=6 November 2017|access-date=5 July 2019|archive-date=7 July 2019|archive-url=https://web.archive.org/web/20190707161547/https://www.nytimes.com/2017/11/06/technology/artificial-intelligence-start-up.html|url-status=live}}

As of 2008,{{Cite book|author1=Bradley Knox, W.|author2=Stone, Peter|title=2008 7th IEEE International Conference on Development and Learning |chapter=TAMER: Training an Agent Manually via Evaluative Reinforcement |year=2008|pages = 292–297|doi=10.1109/devlrn.2008.4640845|isbn = 978-1-4244-2661-4|s2cid = 5613334}} researchers at The University of Texas at Austin (UT) developed a machine learning framework called Training an Agent Manually via Evaluative Reinforcement, or TAMER, which proposed new methods for robots or computer programs to learn how to perform tasks by interacting with a human instructor. First developed as TAMER, a new algorithm called Deep TAMER was later introduced in 2018 during a collaboration between U.S. Army Research Laboratory (ARL) and UT researchers. Deep TAMER used deep learning to provide a robot with the ability to learn new tasks through observation. Using Deep TAMER, a robot learned a task with a human trainer, watching video streams or observing a human perform a task in-person. The robot later practiced the task with the help of some coaching from the trainer, who provided feedback such as "good job" and "bad job".{{Cite web|url=https://governmentciomedia.com/talk-algorithms-ai-becomes-faster-learner|title=Talk to the Algorithms: AI Becomes a Faster Learner|website=governmentciomedia.com|date=16 May 2018 |access-date=29 August 2018|archive-date=28 August 2018|archive-url=https://web.archive.org/web/20180828001727/https://governmentciomedia.com/talk-algorithms-ai-becomes-faster-learner|url-status=live}}

Deep learning has attracted both criticism and comment, in some cases from outside the field of computer science.

{{see also|Explainable artificial intelligence}}

A main criticism concerns the lack of theory surrounding some methods.{{Cite web|url=https://medium.com/@GaryMarcus/in-defense-of-skepticism-about-deep-learning-6e8bfd5ae0f1|title=In defense of skepticism about deep learning|last=Marcus|first=Gary|date=14 January 2018|website=Gary Marcus|access-date=11 October 2018|archive-date=12 October 2018|archive-url=https://web.archive.org/web/20181012035405/https://medium.com/@GaryMarcus/in-defense-of-skepticism-about-deep-learning-6e8bfd5ae0f1|url-status=live}} Learning in the most common deep architectures is implemented using well-understood gradient descent. However, the theory surrounding other algorithms, such as contrastive divergence is less clear.{{citation needed|date=July 2016}} (e.g., Does it converge? If so, how fast? What is it approximating?) Deep learning methods are often looked at as a black box, with most confirmations done empirically, rather than theoretically.{{cite web | last=Knight | first=Will | title=DARPA is funding projects that will try to open up AI's black boxes | website=MIT Technology Review | date=14 March 2017 | url=https://www.technologyreview.com/s/603795/the-us-military-wants-its-autonomous-machines-to-explain-themselves/ | access-date=2 November 2017 | archive-date=4 November 2019 | archive-url=https://web.archive.org/web/20191104033107/https://www.technologyreview.com/s/603795/the-us-military-wants-its-autonomous-machines-to-explain-themselves/ | url-status=live }}

In further reference to the idea that artistic sensitivity might be inherent in relatively low levels of the cognitive hierarchy, a published series of graphic representations of the internal states of deep (20-30 layers) neural networks attempting to discern within essentially random data the images on which they were trained{{cite web |url=http://googleresearch.blogspot.co.uk/2015/06/inceptionism-going-deeper-into-neural.html |author1=Alexander Mordvintsev |author2=Christopher Olah |author3=Mike Tyka |date=17 June 2015 |title=Inceptionism: Going Deeper into Neural Networks |publisher=Google Research Blog |access-date=20 June 2015 |archive-date=3 July 2015 |archive-url=https://web.archive.org/web/20150703064823/http://googleresearch.blogspot.co.uk/2015/06/inceptionism-going-deeper-into-neural.html |url-status=live }} demonstrate a visual appeal: the original research notice received well over 1,000 comments, and was the subject of what was for a time the most frequently accessed article on The Guardian's{{cite news|url=https://www.theguardian.com/technology/2015/jun/18/google-image-recognition-neural-network-androids-dream-electric-sheep|title=Yes, androids do dream of electric sheep|date=18 June 2015|newspaper=The Guardian|author=Alex Hern|access-date=20 June 2015|archive-date=19 June 2015|archive-url=https://web.archive.org/web/20150619200845/http://www.theguardian.com/technology/2015/jun/18/google-image-recognition-neural-network-androids-dream-electric-sheep|url-status=live}} website.

Some deep learning architectures display problematic behaviors,{{cite web |first=Ben |last=Goertzel |title=Are there Deep Reasons Underlying the Pathologies of Today's Deep Learning Algorithms? |year=2015 |url=http://goertzel.org/DeepLearning_v1.pdf |access-date=2015-05-10 |archive-date=2015-05-13 |archive-url=https://web.archive.org/web/20150513053107/http://goertzel.org/DeepLearning_v1.pdf |url-status=live }} such as confidently classifying unrecognizable images as belonging to a familiar category of ordinary images (2014){{cite arXiv |eprint=1412.1897|last1=Nguyen|first1=Anh|title=Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images|last2=Yosinski|first2=Jason|last3=Clune|first3=Jeff|class=cs.CV|year=2014}} and misclassifying minuscule perturbations of correctly classified images (2013).{{cite arXiv |eprint=1312.6199|last1=Szegedy|first1=Christian|title=Intriguing properties of neural networks|last2=Zaremba|first2=Wojciech|last3=Sutskever|first3=Ilya|last4=Bruna|first4=Joan|last5=Erhan|first5=Dumitru|last6=Goodfellow|first6=Ian|last7=Fergus|first7=Rob|class=cs.CV|year=2013}} Goertzel hypothesized that these behaviors are due to limitations in their internal representations and that these limitations would inhibit integration into heterogeneous multi-component artificial general intelligence (AGI) architectures. These issues may possibly be addressed by deep learning architectures that internally form states homologous to image-grammar{{cite journal | last1 = Zhu | first1 = S.C. | last2 = Mumford | first2 = D. | year = 2006| title = A stochastic grammar of images | journal = Found. Trends Comput. Graph. Vis. | volume = 2 | issue = 4| pages = 259–362 | doi = 10.1561/0600000018| citeseerx = 10.1.1.681.2190 }} decompositions of observed entities and events. Learning a grammar (visual or linguistic) from training data would be equivalent to restricting the system to commonsense reasoning that operates on concepts in terms of grammatical production rules and is a basic goal of both human language acquisitionMiller, G. A., and N. Chomsky. "Pattern conception". Paper for Conference on pattern detection, University of Michigan. 1957. and artificial intelligence (AI).{{cite web |first=Jason |last=Eisner |title=Deep Learning of Recursive Structure: Grammar Induction |url=http://techtalks.tv/talks/deep-learning-of-recursive-structure-grammar-induction/58089/ |access-date=2015-05-10 |archive-date=2017-12-30 |archive-url=https://web.archive.org/web/20171230010335/http://techtalks.tv/talks/deep-learning-of-recursive-structure-grammar-induction/58089/ |url-status=dead }}

As deep learning moves from the lab into the world, research and experience show that artificial neural networks are vulnerable to hacks and deception.{{Cite web|url=https://gizmodo.com/hackers-have-already-started-to-weaponize-artificial-in-1797688425|title=Hackers Have Already Started to Weaponize Artificial Intelligence|website=Gizmodo|date=11 September 2017 |access-date=11 October 2019|archive-date=11 October 2019|archive-url=https://web.archive.org/web/20191011162231/https://gizmodo.com/hackers-have-already-started-to-weaponize-artificial-in-1797688425|url-status=live}} By identifying patterns that these systems use to function, attackers can modify inputs to ANNs in such a way that the ANN finds a match that human observers would not recognize. For example, an attacker can make subtle changes to an image such that the ANN finds a match even though the image looks to a human nothing like the search target. Such manipulation is termed an "adversarial attack".{{Cite web|url=https://www.dailydot.com/debug/adversarial-attacks-ai-mistakes/|title=How hackers can force AI to make dumb mistakes|date=18 June 2018|website=The Daily Dot|language=en|access-date=11 October 2019|archive-date=11 October 2019|archive-url=https://web.archive.org/web/20191011162230/https://www.dailydot.com/debug/adversarial-attacks-ai-mistakes/|url-status=live}}

In 2016 researchers used one ANN to doctor images in trial and error fashion, identify another's focal points, and thereby generate images that deceived it. The modified images looked no different to human eyes. Another group showed that printouts of doctored images then photographed successfully tricked an image classification system.{{Cite news|url=https://singularityhub.com/2017/10/10/ai-is-easy-to-fool-why-that-needs-to-change|title=AI Is Easy to Fool—Why That Needs to Change|date=10 October 2017|work=Singularity Hub|access-date=11 October 2017|archive-date=11 October 2017|archive-url=https://web.archive.org/web/20171011233017/https://singularityhub.com/2017/10/10/ai-is-easy-to-fool-why-that-needs-to-change/|url-status=live}} One defense is reverse image search, in which a possible fake image is submitted to a site such as TinEye that can then find other instances of it. A refinement is to search using only parts of the image, to identify images from which that piece may have been taken.{{Cite journal|last=Gibney|first=Elizabeth|title=The scientist who spots fake videos|url=https://www.nature.com/news/the-scientist-who-spots-fake-videos-1.22784|journal=Nature|doi=10.1038/nature.2017.22784|year=2017|access-date=2017-10-11|archive-date=2017-10-10|archive-url=https://web.archive.org/web/20171010011017/http://www.nature.com/news/the-scientist-who-spots-fake-videos-1.22784|url-status=live}}

Another group showed that certain psychedelic spectacles could fool a facial recognition system into thinking ordinary people were celebrities, potentially allowing one person to impersonate another. In 2017 researchers added stickers to stop signs and caused an ANN to misclassify them.

ANNs can however be further trained to detect attempts at deception, potentially leading attackers and defenders into an arms race similar to the kind that already defines the malware defense industry. ANNs have been trained to defeat ANN-based anti-malware software by repeatedly attacking a defense with malware that was continually altered by a genetic algorithm until it tricked the anti-malware while retaining its ability to damage the target.

In 2016, another group demonstrated that certain sounds could make the Google Now voice command system open a particular web address, and hypothesized that this could "serve as a stepping stone for further attacks (e.g., opening a web page hosting drive-by malware)".

In "data poisoning", false data is continually smuggled into a machine learning system's training set to prevent it from achieving mastery.

The deep learning systems that are trained using supervised learning often rely on data that is created or annotated by humans, or both.{{Cite journal |last=Tubaro |first=Paola |date=2020 |title=Whose intelligence is artificial intelligence? |url=https://hal.science/hal-03029735 |journal=Global Dialogue |language=en |pages=38–39}} It has been argued that not only low-paid clickwork (such as on Amazon Mechanical Turk) is regularly deployed for this purpose, but also implicit forms of human microwork that are often not recognized as such.{{Cite journal|last=Mühlhoff|first=Rainer|date=6 November 2019|title=Human-aided artificial intelligence: Or, how to run large computations in human brains? Toward a media sociology of machine learning|journal=New Media & Society|language=en|volume=22|issue=10|pages=1868–1884|doi=10.1177/1461444819885334|s2cid=209363848|issn=1461-4448|url=https://depositonce.tu-berlin.de/handle/11303/12510|doi-access=free}} The philosopher Rainer Mühlhoff distinguishes five types of "machinic capture" of human microwork to generate training data: (1) gamification (the embedding of annotation or computation tasks in the flow of a game), (2) "trapping and tracking" (e.g. CAPTCHAs for image recognition or click-tracking on Google search results pages), (3) exploitation of social motivations (e.g. tagging faces on Facebook to obtain labeled facial images), (4) information mining (e.g. by leveraging quantified-self devices such as activity trackers) and (5) clickwork.

• Applications of artificial intelligence
• Comparison of deep learning software
• Compressed sensing
• Differentiable programming
• Echo state network
• List of artificial intelligence projects
• Liquid state machine
• List of datasets for machine-learning research
• Reservoir computing
• Scale space and deep learning
• Sparse coding
• Stochastic parrot
• Topological deep learning

{{Reflist|30em}}

{{refbegin}}

• {{cite book |last1=Bishop |first1=Christopher M. |last2=Bishop |first2=Hugh |title=Deep learning: foundations and concepts |date=2024 |publisher=Springer |isbn=978-3-031-45467-7}}
• {{cite book |last1=Prince |first1=Simon J. D. |title=Understanding deep learning |date=2023 |publisher=The MIT Press |isbn=9780262048644}}
• {{cite book

|title=Deep Learning

|year=2016

|first1=Ian

|last1=Goodfellow

|author-link1=Ian Goodfellow

|first2=Yoshua

|last2=Bengio

|author-link2=Yoshua Bengio

|first3=Aaron

|last3=Courville

|publisher=MIT Press

|url=http://www.deeplearningbook.org

|isbn=978-0-26203561-3

|postscript=, introductory textbook.

|access-date=2021-05-09

|archive-date=2016-04-16

|archive-url=https://web.archive.org/web/20160416111010/http://www.deeplearningbook.org/

|url-status=live

}}

{{Prone to spam|date=June 2015}}

{{refend}}

{{Artificial intelligence navbox}}