Down the deep rabbit hole: Untangling deep learning from machine learning and artificial intelligence
First Monday

Down the deep rabbit hole: Untangling deep learning from machine learning and artificial intelligence by Niel Chah

Interest in deep learning, machine learning, and artificial intelligence from industry and the general public has reached a fever pitch recently. However, these terms are frequently misused, confused, and conflated. This paper serves as a non-technical guide for those interested in a high-level understanding of these increasingly influential notions by exploring briefly the historical context of deep learning, its public presence, and growing concerns over the limitations of these techniques. As a first step, artificial intelligence and machine learning are defined. Next, an overview of the historical background of deep learning reveals its wide scope and deep roots. A case study of a major deep learning implementation is presented in order to analyze public perceptions shaped by companies focused on technology. Finally, a review of deep learning limitations illustrates systemic vulnerabilities and a growing sense of concern over these systems.


Untangling deep learning from artificial intelligence and machine learning
Deep learning is born
Many narratives to explain deep learning
Deep learning on the public stage
Notable concerns




Recently, the phrases artificial intelligence (AI), machine learning (ML), and deep learning (DL) have made frequent appearances outside usual communities of researchers and scientists, becoming part of a vocabulary in the media and business world. The significant attention that has been given to these areas is based in part on unprecedented advances in certain technological capabilities like image recognition, speech recognition, and machine translation. Millions of users encounter these capabilities through their use of personal computing devices and the Internet. Advances in AI, ML, and DL can be seen in action in offerings from Google and Facebook in their implementations of computer vision (letting computers see what is in an image), automatic speech recognition (letting computers hear human speech), and natural language processing (letting computers parse human languages). However, as popular as these terms are, they are also at risk of being misused, confused, and conflated.

This paper briefly explores deep learning in relation to two closely related terms: artificial intelligence and machine learning. First, preliminary definitions and the complexities associated with them are treated. Next, this paper explores DL’s historical context in an overview, its implementation at a major technology company and associated public discourses, and notable concerns associated with its limitations and vulnerabilities. By exploring the wide scope and deep roots of DL, this paper attempts to serve as a preliminary guide for those interested in understanding these increasingly influential terms.



Untangling deep learning from artificial intelligence and machine learning

Deep learning (DL) refers to a recently rebranded set of techniques that draws upon a long history of advances using neural networks. Before this statement can be explained further, it is necessary to untangle how DL is positioned in relation to two other terms in this shared space: artificial intelligence (AI) and machine learning (ML). If these terms can be depicted as occupying a series of nested circles based on the specificity of their definitions (see Figure 1), this paper first defines the outer ring (AI), then the intermediary ring (ML), and focuses on the inner ring (DL). This depiction conveys how the enclosure of a term by another implies that the former is a more specific sub-field within the latter. It will be useful to consider this visualization in order to have a general sense of size and scope. The diagram is not meant to express a precise and exhaustive framework for the relationships between these terms. Throughout the paper, this nested AI/ML/DL relationship, its complexities, and its conflated usages in academic, public, and government contexts will be referenced.


Relationship between definitions of AI, ML, and DL
Figure 1: Relationship between definitions of AI, ML, and DL. (Source: Parloff, 2016)


Before each of the terms in the AI/ML/DL nexus are explained further, it is important to unpack what happens when a new advance in AI, ML, or DL is announced. Since 2006, advances in DL have pushed the state-of-the-art in many fields such as automatic speech recognition, computer vision, and natural language processing. The “state-of-the-art” is often invoked in science and engineering to describe a new feat of performance according to some pre-defined metric or competition. For example, an image recognition system that achieves a new record in the well known ImageNet competition would be considered as achieving state-of-the-art results until the metric is overcome by the next system (Deng, et al., 2009). Through these incremental advances in the state-of-the-art, DL has expanded the capabilities of AI from recognizing handwritten numerical digits and defeating humans in chess to recognizing more complex images of cats and defeating humans in the more complex game of Go. However, although DL and AI are related in this way, it is important to clarify precisely how DL is AI by including certain qualifiers.

It is difficult to provide a definition of AI that is universally accepted. The 2016 Stanford AI Report was released as part of a larger initiative on the University’s One Hundred Year Study on Artificial Intelligence (or AI100). The report recognized the contentious and difficult nature of defining artificial intelligence (Stone, et al., 2016). Computer scientist Nils J. Nilsson offered a general definition for AI in that report: “Artificial intelligence is that activity devoted to making machines intelligent, and intelligence is that quality that enables an entity to function appropriately and with foresight in its environment” (emphasis added) (Stone, et al., 2016). The dynamic elements in this definition emphasize what is considered to be “appropriate foresight” in a particular “environment”. These elements affected how AI was defined over time.

The Stanford AI100 report noted how this dynamic definition determined what was considered as AI at a particular time and place. AI advanced so rapidly that “the field of AI is a continual endeavor to push forward the frontier of machine intelligence” such that “ironically, AI suffers the perennial fate of losing claim to its acquisitions” (Stone, et al., 2016). As a result of this “AI effect”, state-of-the-art AI technologies that are active today will eventually be overtaken by advances in the future, shifting the definition of AI further as well. Recognizing this dynamic characteristic of AI is important when considering how ML and DL, the next layers in our nested definitions, are affected.

Machine learning refers to a set of techniques that implement capabilities commonly thought of as AI by enabling machines to learn from data without explicit programming or hand-written rules. ML was first defined by Arthur Samuel in 1959 as the “field of study that gives computers the ability to learn without being explicitly programmed” (Géron, 2017). There are two important components to consider in this definition of ML: learning and a lack of explicit programming. An AI is “explicitly programmed” if it exhibits many useful functions but is actually operating based on computational rules that were manually written by humans. For instance, ML was commercially successful in filtering spam e-mail messages using a technique known as the naïve Bayes classifier (Géron, 2017). A spam filter that was explicitly programmed would require as many rules as there are kinds of spam. While this explicit programming approach cannot scale properly, ML allows a given machine to learn statistical patterns in spam from data itself. This emphasis on machine learning from data is a recurring theme.

Although the three terms — AI, ML and DL — are intricately linked, nuanced differences in their specific definitions can make the difference between whether the term is used precisely or whether the actual operations on the ground are obfuscated. The dynamic definition of AI affects what state-of-the-art advances are considered as AI for a particular time and place. ML is primarily concerned with training machines to learn from data, following closely the original definition by Arthur Samuel in 1959. To implement the ever changing state-of-the-art techniques that exhibit AI capabilities, DL is one of the most popular set of ML techniques in use today.



Deep learning is born

While deep learning is relatively new, first appearing in 2006, the underlying techniques using artificial neural networks (ANNs) date back to the 1940s (or earlier, depending on the source describing a narrative of DL). For instance, the timeline of neural network ideas by Haohan Wang and Bhiksha Raj in their paper “On the origin of deep learning” traced DL back to Aristotle in 300 BC (Wang and Raj, 2017). In another narrative, Jürgen Schmidhuber described the beginnings of DL to ideas from the 1800s in his article “Deep learning in neural networks: An overview” (Schmidhuber, 2015). For this paper, the first artificial neuron proposed by Warren McCulloch and Walter Pitts at the University of Illinois College of Medicine in Chicago in 1943 is the beginning of the DL story, although other arbitrary starting points may be justified in other accounts (McCulloch and Pitts, 1943).

A highly condensed account of DL’s history is provided here, with more thorough and technical accounts available in Schmidhuber (2015) and Wang and Raj (2017). In 1943, the McCulloch-Pitts artificial neuron was modeled on a biological neuron that receives electrical signals from other neurons and transmits a signal if the signal strength passes a certain threshold. This early artificial neuron could perform simple logical computations. Along with the Perceptron model proposed by Frank Rosenblatt in 1957, these techniques contributed to early successes with ANNs (Géron, 2017). This took a turn in 1969 as Perceptrons by Marvin Minsky and Seymour Papert made the case against neural networks and contributed to a decline in research (Minsky and Papert, 1969). In the 1980s, ANNs improved but they fell behind then state-of-the-art results from other ML techniques such as support vector machines (Géron, 2017). ML using neural networks continued to advance incrementally during the unproductive AI winters of the 1980s. For example, Yann Lecun developed an ANN that could read handwritten digits (Lecun, et al., 1998).

The year 2006 has been considered by some researchers as the start of deep learning. In 2006, Geoffrey Hinton, et al. published a paper on deep belief nets (DBN) that demonstrated the possibility of using “deep” neural networks to achieve state-of-the-art results (1.25 percent error rate) in recognizing handwritten digits (Hinton, et al., 2006). This paper marked the first appearance of the words “deep” in the context of machine learning, where “deep” refers to layers of interconnected artificial neurons stacked together. This definition is found widely throughout research literature. The Stanford AI report defined DL as “a form of machine learning based on layered representations of variables referred to as neural networks ... can be applied widely to an array of applications that rely on pattern recognition” [1]. Schmidhuber also noted “the expression deep learning was actually coined around 2006” [2]. Wang and Raj agreed that the DBN paper “is generally seen as the opening of nowadays [sic] Deep Learning era” [3].

Just as DL emerged from a long history of developments in neural networks, new advances from DL continued to be made rapidly. The history of DL continues to be written with each new advance. DL made significant state-of-the-art advances with a variety of neural network architectures such as convolutional neural networks (CNN), recurrent neural networks (RNN), and long short term memory (LSTM). The most recent novel development from the DL field is the capsule network by Geoffrey Hinton, the godfather of DL, and colleagues proposed in November 2017 (Sabour, et al., 2017). Capsules are inspired by the capabilities of human vision to “ignore irrelevant details” and process “a tiny fraction of the optic array” and achieved state-of-the-art results in recognizing handwritten digits [4]. As a response to certain shortcomings associated with CNNs, it will take time to see how capsule networks are applied and what other novel developments may arise from the DL research.



Many narratives to explain deep learning

Due to the field’s wide scope and deep historical roots, the universal acceptance of a single canonical history of DL is difficult. Aside from wide agreement on the field’s origins in the 2006 DBN paper, there have been various attempts to describe DL’s development. For instance, in one well known narrative, DL was a field initially built by a team of Canadian researchers. There are slightly different narratives presented by others, with interesting decisions on what was included and excluded. These accounts are primarily geared towards an academic and research audience so the technical elements are not readily accessible to those outside the immediate DL community.

According to one account, deep learning was developed by a number of key researchers with links to Canada (Bergen and Wagner, 2015). Geoffrey Hinton, the godfather of deep learning, implemented the deep belief net at the University of Toronto and played a leading role in developing DL methods with support from the Canadian Institute for Advanced Research (CIFAR). Yoshua Bengio, at Université de Montréal, and Yann Lecun, currently at Facebook, were other key figures that were part of this effort. A 2017 Globe and Mail article by CIFAR executive Alan Bernstein, et al. made it clear that DL were stimulated by CIFAR’s help: “Deep learning and related AI techniques were developed by Geoff Hinton ..., Yoshua Bengio ..., and Yann LeCun ..., Richard Sutton ... and a host of other researchers supported by CIFAR and its program in Learning, Machines and Brains” (Bernstein, et al., 2017). This story of DL developing in Canada also referenced the role of CIFAR in funding Hinton, et al. at a time when such financial support was not popular. The significant influence of the Canadian contribution to deep learning by Hinton, Bengio, and Lecun can be seen in their 2015 paper in Nature entitled simply “deep learning” (Lecun, et al., 2015).

Although the Canadian contribution to deep learning was significant, it is critically important to recognize the role of other key figures and their contributions. The narratives by other prominent researchers contributes to a more nuanced understanding of this field. For instance, Jürgen Schmidhuber is a prominent compute scientist who has made significant contributions such as long-short term memory, a type of neural network (Schmidhuber, 2015). According to Schmidhuber’s account in “Deep learning in neural networks: An overview,” DL techniques can be traced back to the nineteenth century, since early neural networks were “essentially variants of linear regression methods going back at least to the early 1800s” [5]. The rest of that report described important milestones in the development of different kinds of neural networks.

In Wang and Raj’s (2017) account, other aspects of the DL timeline were noted. In their account of DL’s evolutionary history, they considered the underlying concepts that were associated with DL. According to their view, DL was the spiritual successor to a long line of ideas that were motivated by connectionism and the “ambition to build a computer system simulating the human brain” [6]. As a result, their paper traced the origins of DL to Aristotle’s associationism from 300 B.C. [7]. The remainder of their paper described associations between different elements in the human mind, tracing those concepts that eventually led to the development of ML and DL.

Although the precise details of each historical narrative differed, there was general agreement on the definition of the field of DL. Regardless of the storyteller, the field’s achievements were presented with an emphasis on learning from the data, unsurprisingly echoing the core definition of ML. As Lecun, Bengio, and Hinton (2015) state, “the key aspect of deep learning is that these layers of features are not designed by human engineers: they are learned from data using a general-purpose learning procedure.” Such an understanding and framing of DL is representative of how the field is presented outside of research circles. ML, by definition, requires machines to learn from data without human intervention, and DL is presented as one of the best performing state-of-the-art methods in ML to implement AI capabilities.



Deep learning on the public stage

Although technical literature on deep learning was written by and for researchers, computer scientists, and programmers, many key concepts on deep learning have filtered into resources such as news articles and government publications. The New York Times, Forbes, and the New Yorker are examples of some outlets that have presented DL to larger audiences. Aside from the mathematical equations, algorithms, and proofs that are essential in research papers, many of the essential concepts are accurately presented at a general level for consumption by the public. Reading these popular accounts in conjunction with press releases from large technology firms implementing DL is instructive. Furthermore, the products from companies such as Google and Facebook have made the wide scale impact of AI/ML/DL apparent stimulating popular accounts.

In 2016, Sundar Pichai, chief executive officer of Google, announced that Google would become an “AI-first company” (Lewis-Kraus, 2016). In the same year, Google overhauled its Translate service with a new neural machine translation (NMT) system based on DL (Lewis-Kraus, 2016). The story of Google’s application of AI/ML/DL was covered in detail by Gideon Lewis-Kraus in a New York Times Magazine story entitled “The great A.I. awakening.” This article weaved a history of deep learning of its own, but echoed many common points: the Canadian and international story (“... the story of deep learning, takes place in a variety of far-flung laboratories — in Scotland, Switzerland, Japan and most of all Canada — over seven decades”); the early setbacks in Perceptrons (“... with an M.I.T. colleague, Minsky published a book that proved that there were painfully simple problems the Perceptron could never solve”); the “deep” multilayers associated with DL (“... Hinton already knew at the time that complex tasks could be carried out if you had recourse to multiple layers”); and, the importance of learning with computation from data (“... the whole point of machine learning, however, is to avoid that kind of explicit mentoring” and “... there just wasn’t enough computer power or enough data [until recently]“) (Lewis-Kraus, 2016).

DL, ML, and AI also made appearances in government discourse. The Canadian government considered DL a significant area for research funding, as demonstrated in the 2017 initiative for CIFAR to lead the Pan-Canadian Artificial Intelligence Strategy, a CAN$125 million commitment to develop three AI institutes in Canadian centres (CIFAR, 2017). In 2016, a CAN$93 million Canada First Research Excellence Fund was awarded to Université de Montréal for Data serving Canadians: Deep learning and optimization for the knowledge revolution (Government of Canada, 2016). In both these initiatives, the layered definitions of AI/ML/DL were invoked, with DL named outright as a focus area from among the ML/AI disciplines for CIFAR (“The strategy funded three AI institutes in Canada’s three major centres for deep learning and reinforcement learning research — in Edmonton, Montréal, and Toronto-Waterloo”) as well as at Université de Montréal (“a transformative and far-reaching strategy that capitalizes on the unique and synergistic combination of machine learning/deep learning”).

The United States government also expressed interest in AI/ML/DL. As of 2016, the Obama White House led a number of workshops and interagency working groups on the “benefits and risk of artificial intelligence” (Felten, 2016). While the Canadian initiatives directly named DL as a field that has advanced AI, the American government was more generally concerned with the outermost layer in the AI/ML/DL relationship, focusing on AI advances and ethics in general. Interestingly, there was an indirect reference to DL but it was not named outright as in the Canadian documents: “... a series of breakthroughs in the research community and industry have recently spurred momentum and investment in the development of this field” (Felten, 2016). The differences in these discourses addressed a difference in scope for the two governments. Canada witnessed first-hand how its research funding had injected energy into DL, while the United States was more interested in the implications of AI from advances in DL and other techniques.



Notable concerns

Thus far, DL had been presented in a largely positive light with little reference to issues associated with modern neural networks. This section briefly presents two notable concerns associated with AI/ML/DL systems. First, at a technical level, the techniques are vulnerable to an “adversarial attack” that attempts to fool the neural network. Although these technical vulnerabilities have been addressed in a number of significant ways, they are worth examining as representative of technical issues. Second, there have been other concerns, examining ethical implications of AI/ML/DL systems.

DL techniques are vulnerable to “adversarial attacks”. A neural network (NN) that has processed thousands of images of cats, dogs, and other animals — and has a high accuracy in classifying these images — will still occasionally misclassify an image. In this case, the NN devised a conceptualization of the particular weights and patterns associated with each animal; however it was still vulnerable to slight perturbations in data that may be imperceptible to humans (see Figure 2). As explained in a Y Combinator blog post, an adversarial attack is conducted by deliberately adjusting an image at the pixel level to be indistinguishable for a human but consistently misinterpreted by a trained neural network (Mikhailov and Trusov, 2017).


An example of an adversarial attack
Figure 2: An example of an adversarial attack. (Source: Mikhailov and Trusov, 2017)


The vulnerabilities associated with adversarial attacks are indicative of imperfect representations adopted by DL systems. While humans are able to reason, infer, and extrapolate from observations, these capabilities have not been well developed in DL systems. Despite state-of-the-art advances in performance, real-world instances where machine learning falls short have appeared at times. In 2016 an individual died for the first time in a self-driving car accident involving a Tesla operating in autopilot mode when it failed to distinguish a “large white 18-wheel truck and trailer crossing the highway” and “attempted to drive full speed under the trailer” (Yadron and Tynan, 2016).

In addition to technical vulnerabilities with adversarial attacks, there have been significant ethical concerns with AI implemented through DL. What human biases are inherent in data that NNs learn from? How are these biases identified and what actions should be taken in response? What does it mean that a state-of-the-art result has been achieved? How meaningful is the new state-of-the-art result as an indicator of progress in the real world with its messy complexities? At all levels of the AI/ML/DL relationship, these kinds of ethical questions and issues have been at the forefront. Attention to the ethical dimension has attracted increasingly prominent voices in industry, government, and academia. In 2017, Université de Montréal released the Montréal Declaration for a Responsible Development of Artificial Intelligence. The Declaration is a consideration of how DL and ML methods have ethical implications for AI in the areas of personal well-being, autonomy, privacy, and responsibility. Many of the responses to these ethical questions for AI/ML/DL continue to be debated as technology develops.

An overview of the most significant vulnerabilities and the ethical concerns of DL have revealed how the underlying technique is an algorithmically driven system that simply learns inconsistently. Machine learning is inconsistent in that it is susceptible to vulnerabilities like adversarial attacks. It has also learned from data marked with biases of humans. Once deconstructed in this way, DL can be readily linked to a growing literature at the confluence of social and technical issues on human bias in algorithms, algorithmic identity, the machine’s conceptualization of data, and issues with the platforms where these DL systems have been implemented (Cheney-Lippold, 2017; Finn, 2017).




By reviewing the definitions of AI, ML, and DL, this paper has offered a very introductory and extremely brief guide to distinguish between artificial intelligence, machine learning, or deep learning. Through a chain of events that began with the incremental advance of a metric, the implications of the new state-of-the-art propagate to improved ML and AI capabilities. Eventually, these advances may be applied in production environments at technology companies, allowing the changes to reach a large number of users connected through the Internet. Coupled with the dynamic nature of AI, or the “AI effect”, incremental advances in the underlying techniques have the possibility to accumulate over time.

Although a variety of different stories on the history of DL were presented, these accounts are still being written with each new advance. An overview of DL’s historical development showed that new developments such as capsule networks are continuously being made to address shortcomings in previous architectures. Over time, the deeper implications and applications of these new methods will become clearer.

Many of the prominent researchers in the DL community have been vocal about current limitations of the state-of-the-art as well as the future work that will be required in order to lead to major progress in AI. LeCun, et al. (2015) noted that “Ultimately, major progress in artificial intelligence will come about through systems that combine representation learning with complex reasoning.” [8]. The best DL systems have been frequently compared to the best human performance on equivalent tasks. In this way, significant inspiration is drawn from how the human brain operates in an effective and efficient manner. Adversarial attacks that intentionally trick machines but cannot mislead humans suggest that “humans learn to actively perceive patterns by sequentially directing attention to relevant parts of the available data” [9]. The lure towards simpler models is also a motivating factor for research in DL methods [10].

As DL researchers continue to work on research problems, ethical, political, and societal issues involving these technical systems will be considered. These pressing concerns and questions are not necessarily considered by the researchers and computer scientists that work on the technical matters. The nature of academic research and “fields”, where research is conducted in separate communities implies there is the danger of homophily, where experts from the same specialization are in agreement while they “talk past” those from different specializations. The lack of a shared and collaborative discussion on these pressing matters is concerning, particularly when incidents such as the 2016 autopilot accident enter the realm of reality. Increasingly over time, the cross-pollination of ideas and discussion from people in technical areas and other fields will be a significant factor advancing AI, ML, and DL. End of article


About the author

Niel Chah is a Ph.D. student at the University of Toronto in the Faculty of Information (iSchool). His research interests include knowledge graphs, data science, and machine learning.
E-mail: niel [dot] chah [at] mail [dot] utoronto [dot] ca



1. Stone, et al., 2016, p. 4.

2. Schmidhuber, 2015, p. 106.

3. Wang and Raj, 2017, p. 29.

4. Sabour, et al., 2017, p. 1.

5. Schmidhuber, 2015, p. 94.

6. Wang and Raj, 2017, p. 5.

7. Ibid.

8. LeCun, et al., p. 442.

9. Schmidhuber, 2015, p. 117.

10. Wang and Raj, 2017, p. 59.



Alan Bernstein, Pierre Boivin, and David McKay, 2017. “It’ls time for Canada to invest in developing artificial intelligence,” Globe and Mail (26 March), at, accessed 8 January 2018.

Mark Bergen and Kurt Wagner, 2015. “Welcome to the AI conspiracy: The ‘Canadian mafia’ behind tech’s latest craze,” Recode (15 July), at, accessed 8 January 2018.

John Cheney-Lippold, 2017. We are data: Algorithms and the making of our digital selves. New York: New York University Press.

Canadian Institute for Advanced Research (CIFAR), 2017. “Pan-Canadian artificial intelligence strategy overview,” at, accessed 8 January 2018.

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li, 2009. “Imagenet: A large-scale hierarchical image database,” 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp. 248–255.
doi:, accessed 19 January 2019.

Ed Felten, 2016. “Preparing for the future of artificial intelligence” (3 May), at, accessed 8 January 2018.

Ed Finn, 2017. What algorithms want: Imagination in the age of computing. Cambridge, Mass.: MIT Press.

Aurélien Géron, 2017. Hands-on machine learning with scikit-learn and tensorflow: Concepts, tools, and techniques to build intelligent systems. Sebastopol, Calif.: O’Reilly Media.

Government of Canada, 2016. “Canada First Research Excellence Fund. Competition results” (6 September), at, accessed 8 January 2018.

Geoffrey E. Hinton, Simon Osindero, and Yee-Whye Teh, 2006. “A fast learning algorithm for deep belief nets,” Neural Computation, volume 18, number 7, pp. 1,527–1,554.
doi:, accessed 19 January 2019.

Yann LeCun, Yoshua Bengio, and Geoffrey Hinton, 2015. “Deep learning,” Nature, volume 521, number 7553 (27 May), pp. 436–444.
doi:, accessed 19 January 2019.

Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner, 1998. “Gradient-based learning applied to document recognition,” Proceedings of the IEEE, volume 86, number 11, pp. 2,278–2,324.
doi:, accessed 19 January 2019.

Gideon Lewis-Kraus, 2016. “The great AI awakening,” New York Times Magazine (14 December), at, accessed 8 January 2018.

Warren S. McCulloch and Walter Pitts, 1943. “A logical calculus of the ideas immanent in nervous activity,” Bulletin of Mathematical Biophysics, volume 5, number 4, pp. 115–133.
doi:, accessed 19 January 2019.

Emil Mikhailov and Roman Trusov, 2017. “How adversarial attacks work,” Y Combinator (2 November), at, accessed 8 January 2018.

Marvin Minsky and Seymour Papert, 1969. Perceptrons: An introduction to computational geometry. Cambridge, Mass.: M.I.T. Press.

Montréal Declaration for a Responsible Development of Artificial Intelligence, 2017. “Montréal Declaration,” at, accessed 8 January 2018.

Roger Parloff, 2016. “Why deep learning is suddenly changing your life,” Fortune (28 September), at, accessed 8 January 2018.

Jürgen Schmidhuber, 2015. “Deep learning in neural networks: An overview,” Neural Networks, volume 61, pp. 85–117.
doi:, accessed 19 January 2019.

Peter Stone, Rodney Brooks, Erik Brynjolfsson, Ryan Calo, Oren Etzioni, Greg Hager, Julia Hirschberg, Shivaram Kalyanakrishnan, Ece Kamar, Sarit Kraus, Kevin Leyton-Brown, David Parkes, William Press, AnnaLee Saxenian, Julie Shah, Milind Tambe, and Astro Teller, 2016. “Artificial intelligence and life in 2030,” One Hundred Year Study on Artificial Intelligence. Report of the 2015–2016 Study Panel, at, accessed 19 January 2019.

Sara Sabour, Nicholas Frosst, and Geoffrey E. Hinton, 2017. “Dynamic routing between capsules,” Advances in Neural Information Processing Systems 30 (NIPS 2017), at, accessed 19 January 2019.

Haohan Wang and Bhiksha Raj, 2017. “On the origin of deep learning,” arXiv (24 February), at, accessed 19 January 2019.

Danny Yadron and Dan Tynan, 2016. “Tesla driver dies in first fatal crash while using autopilot mode,” Guardian (30 June), at accessed 8 January 2018.


Editorial history

Received 9 January 2018; accepted 18 January 2019.

Creative Commons License
This paper is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.

Down the deep rabbit hole: Untangling deep learning from machine learning and artificial intelligence
by Niel Chah.
First Monday, Volume 24, Number 2 - 4 February 2019

A Great Cities Initiative of the University of Illinois at Chicago University Library.

© First Monday, 1995-2019. ISSN 1396-0466.