Gary F. Marcus is a scientist, author, and entrepreneur, focusing o...
Imagenet is an image database with over 14 million hand-annotated i...
Here is a curated collection of many of the classic deep learning p...
At this point, the "diminishing returns" that Chollet is quoted abo...
And the definition of "multiple" has change over time too. Now, the...
Here is a tutorial video on backpropagation: https://www.youtube.co...
Link to paper: http://yann.lecun.com/exdb/publis/pdf/lecun-89.pdf
Here is a progress report on "One-Shot learning" which is related t...
However, unsupervised learning (does not require labels), has achie...
However, transfer learning has had some very impressive results...
This is a big drawback, and there are a lot of people trying to add...
Hence the rise of Bayesian Deep Learning: http://bayesiandeeplearni...
Hence the rise of interest in the field of causal inference in mach...
Deep Learning:
A Critical Appraisal
Gary Marcus
1
New York University
Abstract
Although deep learning has historical roots going back decades, neither the term “deep
learning” nor the approach was popular just over five years ago, when the field was
reignited by papers such as Krizhevsky, Sutskever and Hinton’s now classic 2012
(Krizhevsky, Sutskever, & Hinton, 2012)deep net model of Imagenet.
What has the field discovered in the five subsequent years? Against a background of
considerable progress in areas such as speech recognition, image recognition, and game
playing, and considerable enthusiasm in the popular press, I present ten concerns for deep
learning, and suggest that deep learning must be supplemented by other techniques if we
are to reach artificial general intelligence.!
! Departments of Psychology and Neural Science, New York University, gary.marcus at nyu.edu. I thank Christina 1
Chen, François Chollet, Ernie Davis, Zack Lipton, Stefano Pacifico, Suchi Saria, and Athena Vouloumanos for
sharp-eyed comments, all generously supplied on short notice during the holidays at the close of 2017.
Page ! of !1 27
For most problems where deep learning has enabled
transformationally better solutions (vision, speech), we've
entered diminishing returns territory in 2016-2017.
François Chollet, Google, author of Keras
neural network library
December 18, 2017
‘Science progresses one funeral at a time.' The future
depends on some graduate student who is deeply suspicious
of everything I have said.
Geoff Hinton, grandfather of deep learning
September 15, 2017
1. Is deep learning approaching a wall?
Although deep learning has historical roots going back decades(Schmidhuber, 2015), it
attracted relatively little notice until just over five years ago. Virtually everything
changed in 2012, with the publication of a series of highly influential papers such as
Krizhevsky, Sutskever and Hinton’s 2012 ImageNet Classification with Deep
Convolutional Neural Networks (Krizhevsky, Sutskever, & Hinton, 2012), which
achieved state-of-the-art results on the object recognition challenge known as ImageNet
(Deng et al., ). Other labs were already working on similar work (Cireşan, Meier, Masci,
& Schmidhuber, 2012). Before the year was out, deep learning made the front page of
The New York Times , and it rapidly became the best known technique in artificial
2
intelligence, by a wide margin. If the general idea of training neural networks with
multiple layers was not new, it was, in part because of increases in computational power
and data, the first time that deep learning truly became practical.
Deep learning has since yielded numerous state of the art results, in domains such as
speech recognition, image recognition , and language translation and plays a role in a
wide swath of current AI applications. Corporations have invested billions of dollars
fighting for deep learning talent. One prominent deep learning advocate, Andrew Ng, has
gone so far to suggest that “If a typical person can do a mental task with less than one
second of thought, we can probably automate it using AI either now or in the near
http://www.nytimes.com/2012/11/24/science/scientists-see-advances-in-deep-learning-a-part-of-artificial-
2
intelligence.html
Page ! of !2 27
future.” (A, 2016). A recent New York Times Sunday Magazine article , largely about
3
deep learning, implied that the technique is “poised to reinvent computing itself.”
Yet deep learning may well be approaching a wall, much as I anticipated earlier, at
beginning of the resurgence (Marcus, 2012), and as leading figures like Hinton (Sabour,
Frosst, & Hinton, 2017) and Chollet (2017) have begun to imply in recent months.
What exactly is deep learning, and what has its shown about the nature of intelligence?
What can we expect it to do, and where might we expect it to break down? How close or
far are we from “artificial general intelligence”, and a point at which machines show a
human-like flexibility in solving unfamiliar problems? The purpose of this paper is both
to temper some irrational exuberance and also to consider what we as a field might need
to move forward.
This paper is written simultaneously for researchers in the field, and for a growing set of
AI consumers with less technical background who may wish to understand where the
field is headed. As such I will begin with a very brief, nontechnical introduction aimed at
4
elucidating what deep learning systems do well and why (Section 2), before turning to an
assessment of deep learning’s weaknesses (Section 3) and some fears that arise from
misunderstandings about deep learning’s capabilities (Section 4), and closing with
perspective on going forward (Section 5).
Deep learning is not likely to disappear, nor should it. But five years into the field’s
resurgence seems like a good moment for a critical reflection, on what deep learning has
and has not been able to achieve.
2. What deep learning is, and what it does well
Deep learning, as it is primarily used, is essentially a statistical technique for classifying
patterns, based on sample data, using neural networks with multiple layers.
5
https://www.nytimes.com/2016/12/14/magazine/the-great-ai-awakening.html
3
For more technical introduction, there are many excellent recent tutorials on deep learning including (Chollet,
4
2017) and (Goodfellow, Bengio, & Courville, 2016), as well as insightful blogs and online resources from Zachary
Lipton, Chris Olah, and many others.
Other applications of deep learning beyond classification are possible, too, though currently less popular, and
5
outside of the scope of the current article. These include using deep learning as an alternative to regression, as a
component in generative models that create (e.g.,) synthetic images, as a tool for compressing images, as a tool for
learning probability distributions, and (relatedly) as an important technique for approximation known as variational
inference.
Page ! of !3 27
Neural networks in the deep learning literature typically consist of a set of input units that
stand for things like pixels or words, multiple hidden layers (the more such layers, the
deeper a network is said to be) containing hidden units (also known as nodes or neurons),
and a set output units, with connections running between those nodes. In a typical
application such a network might be trained on a large sets of handwritten digits (these
are the inputs, represented as images) and labels (these are the outputs) that identify the
categories to which those inputs belong (this image is a 2, that one is a 3, and so forth).
Over time, an algorithm called back-propagation allows a process called gradient descent
to adjust the connections between units using a process, such that any given input tends to
produce the corresponding output.
Collectively, one can think of the relation between inputs and outputs that a neural
network learns as a mapping. Neural networks, particularly those with multiple hidden
layers (hence the term deep) are remarkably good at learning input-output mappings,
Such systems are commonly described as neural networks because the input nodes,
hidden nodes, and output nodes can be thought of as loosely analogous to biological
neurons, albeit greatly simplified, and the connections between nodes can be thought of
as in some way reflecting connections between neurons. A longstanding question, outside
the scope of the current paper, concerns the degree to which artificial neural networks are
biologically plausible.
Most deep learning networks make heavy use of a technique called convolution (LeCun,
1989), which constrains the neural connections in the network such that they innately
capture a property known as translational invariance. This is essentially the idea that an
object can slide around an image while maintaining its identity; a circle in the top left can
be presumed, even absent direct experience) to be the same as a circle in the bottom right.
Page ! of !4 27
Input layer
Output layer
Hidden layers
...
...
...
...
Deep learning is also known for its ability to self-generate intermediate representations,
such as internal units that may respond to things like horizontal lines, or more complex
elements of pictorial structure.
In principle, given infinite data, deep learning systems are powerful enough to represent
any finite deterministic “mapping” between any given set of inputs and a set of
corresponding outputs, though in practice whether they can learn such a mapping
depends on many factors. One common concern is getting caught in local minima, in
which a systems gets stuck on a suboptimal solution, with no better solution nearby in the
space of solutions being searched. (Experts use a variety of techniques to avoid such
problems, to reasonably good effect). In practice, results with large data sets are often
quite good, on a wide range of potential mappings.
In speech recognition, for example, a neural network learns a mapping between a set of
speech sounds, and set of labels (such as words or phonemes). In object recognition, a
neural network learns a mapping between a set of images and a set of labels (such that,
for example, pictures of cars are labeled as cars). In DeepMind’s Atari game system
(Mnih et al., 2015), neural networks learned mappings between pixels and joystick
positions.
Deep learning systems are most often used as classification system in the sense that the
mission of a typical network is to decide which of a set of categories (defined by the
output units on the neural network) a given input belongs to. With enough imagination,
the power of classification is immense; outputs can represent words, places on a Go
board, or virtually anything else.
In a world with infinite data, and infinite computational resources, there might be little
need for any other technique.
3. Limits on the scope of deep learning
Deep learning’s limitations begin with the contrapositive: we live in a world in which
data are never infinite. Instead, systems that rely on deep learning frequently have to
generalize beyond the specific data that they have seen, whether to a new pronunciation
of a word or to an image that differs from one that the system has seen before, and where
data are less than infinite, the ability of formal proofs to guarantee high-quality
performance is more limited.
Page ! of !5 27
As discussed later in this article, generalization can be thought of as coming in two
flavors, interpolation between known examples, and extrapolation, which requires going
beyond a space of known training examples (Marcus, 1998a).
For neural networks to generalize well, there generally must be a large amount of data,
and the test data must be similar to the training data, allowing new answers to be
interpolated in between old ones. In Krizhevsky et al’s paper (Krizhevsky, Sutskever, &
Hinton, 2012), a nine layer convolutional neural network with 60 million parameters and
650,000 nodes was trained on roughly a million distinct examples drawn from
approximately one thousand categories.
6
This sort of brute force approach worked well in the very finite world of ImageNet, into
which all stimuli can be classified into a comparatively small set of categories. It also
works well in stable domains like speech recognition in which exemplars are mapped in
constant way onto a limited set of speech sound categories, but for many reasons deep
learning cannot be considered (as it sometimes is in the popular press) as a general
solution to artificial intelligence.
Here are ten challenges faced by current deep learning systems:
3.1. Deep learning thus far is data hungry
Human beings can learn abstract relationships in a few trials. If I told you that a schmister
was a sister over the age of 10 but under the age of 21, perhaps giving you a single
example, you could immediately infer whether you had any schmisters, whether your best
friend had a schmister, whether your children or parents had any schmisters, and so forth.
(Odds are, your parents no longer do, if they ever did, and you could rapidly draw that
inference, too.)
In learning what a schmister is, in this case through explicit definition, you rely not on
hundreds or thousands or millions of training examples, but on a capacity to represent
abstract relationships between algebra-like variables.
Humans can learn such abstractions, both through explicit definition and more implicit
means (Marcus, 2001). Indeed even 7-month old infants can do so, acquiring learned
abstract language-like rules from a small number of unlabeled examples, in just two
Using a common technique known as data augmentation, each example was actually presented along with its label
6
in a many different locations, both in its original form and in mirror reversed form. A second type of data
augmentation varied the brightness of the images, yielding still more examples for training, in order to train the
network to recognize images with different intensities. Part of the art of machine learning involves knowing what
forms of data augmentation will and won’t help within a given system.
Page ! of !6 27
minutes (Marcus, Vijayan, Bandi Rao, & Vishton, 1999). Subsequent work by Gervain
and colleagues (2012) suggests that newborns are capable of similar computations.
Deep learning currently lacks a mechanism for learning abstractions through explicit,
verbal definition, and works best when there are thousands, millions or even billions of
training examples, as in DeepMind’s work on board games and Atari. As Brenden Lake
and his colleagues have recently emphasized in a series of papers, humans are far more
efficient in learning complex rules than deep learning systems are (Lake, Salakhutdinov,
& Tenenbaum, 2015; Lake, Ullman, Tenenbaum, & Gershman, 2016). (See also related
work by George et al (2017), and my own work with Steven Pinker on children’s
overregularization errors in comparison to neural networks (Marcus et al., 1992).)
Geoff Hinton has also worried about deep learning’s reliance on large numbers of labeled
examples, and expressed this concern in his recent work on capsule networks with his
coauthors (Sabour et al., 2017) noting that convolutional neural networks (the most
common deep learning architecture) may face “exponential inefficiencies that may lead to
their demise. A good candidate is the difficulty that convolutional nets have in
generalizing to novel viewpoints [ie perspectives on object in visual recognition tasks].
The ability to deal with translation[al invariance] is built in, but for the other ... [common
type of] transformation we have to chose between replicating feature detectors on a grid
that grows exponentially ... or increasing the size of the labelled training set in a similarly
exponential way.”
In problems where data are limited, deep learning often is not an ideal solution.
3.2.Deep learning thus far is shallow and has limited capacity for
transfer
Although deep learning is capable of some amazing things, it is important to realize that
the word “deep” in deep learning refers to a technical, architectural property (the large
number of hidden layers used in a modern neural networks, where there predecessors
used only one) rather than a conceptual one (the representations acquired by such
networks don’t, for example, naturally apply to abstract concepts like “justice”,
“democracy” or “meddling”).
Even more down-to-earth concepts like “ball” or “opponent” can lie out of reach.
Consider for example DeepMind’s Atari game work (Mnih et al., 2015) on deep
reinforcement learning, which combines deep learning with reinforcement learning (in
which a learner tries to maximize reward). Ostensibly, the results are fantastic: the system
meets or beats human experts on a large sample of games using a single set of
“hyperparameters” that govern properties such as the rate at which a network alters its
weights, and no advance knowledge about specific games, or even their rules. But it is
Page ! of !7 27
easy to wildly overinterpret what the results show. To take one example, according to a
widely-circulated video of the system learning to play the brick-breaking Atari game
Breakout, “after 240 minutes of training, [the system] realizes that digging a tunnel
thought the wall is the most effective technique to beat the game”.
But the system has learned no such thing; it doesn’t really understand what a tunnel, or
what a wall is; it has just learned specific contingencies for particular scenarios. Transfer
tests — in which the deep reinforcement learning system is confronted with scenarios
that differ in minor ways from the one ones on which the system was trained show that
deep reinforcement learning’s solutions are often extremely superficial. For example, a
team of researchers at Vicarious showed that a more efficient successor technique,
DeepMind’s Atari system [Asynchronous Advantage Actor-Critic; also known as A3C],
failed on a variety of minor perturbations to Breakout (Kansky et al., 2017) from the
training set, such as moving the Y coordinate (height) of the paddle, or inserting a wall
midscreen. These demonstrations make clear that it is misleading to credit deep
reinforcement learning with inducing concept like wall or paddle; rather, such remarks
are what comparative (animal) psychology sometimes call overattributions. It’s not that
the Atari system genuinely learned a concept of wall that was robust but rather the system
superficially approximated breaking through walls within a narrow set of highly trained
circumstances.
7
My own team of researchers at a startup company called Geometric Intelligence (later
acquired by Uber) found similar results as well, in the context of a slalom game, In 2017,
a team of researchers at Berkeley and OpenAI has shown that it was not difficult to
construct comparable adversarial examples in a variety of games, undermining not only
DQN (the original DeepMind algorithm) but also A3C and several other related
techniques (Huang, Papernot, Goodfellow, Duan, & Abbeel, 2017).
Recent experiments by Robin Jia and Percy Liang (2017) make a similar point, in a
different domain: language. Various neural networks were trained on a question
answering task known as SQuAD (derived from the Stanford Question Answering
Database), in which the goal is to highlight the words in a particular passage that
correspond to a given question. In one sample, for instance, a trained system correctly,
and impressively, identified the quarterback on the winning of Super Bowl XXXIII as
John Elway, based on a short paragraph. But Jia and Liang showed the mere insertion of
distractor sentences (such as a fictional one about the alleged victory of Google’s Jeff
In the same paper, Vicarious proposed an alternative to deep learning called schema networks (Kansky et al., 2017)
7
that can handle a number of variations in the Atari game Breakout, albeit apparently without the multi-game
generality of DeepMind’s Atari system.
Page ! of !8 27
Dean in another Bowl game ) caused performance to drop precipitously. Across sixteen
8
models, accuracy dropped from a mean of 75% to a mean of 36%.
As is so often the case, the patterns extracted by deep learning are more superficial than
they initially appear.
3.3.Deep learning thus far has no natural way to deal with
hierarchical structure
To a linguist like Noam Chomsky, the troubles Jia and Liang documented would be
unsurprising. Fundamentally, most current deep-learning based language models
represent sentences as mere sequences of words, whereas Chomsky has long argued that
language has a hierarchical structure, in which larger structures are recursively
constructed out of smaller components. (For example, in the sentence the teenager who
previously crossed the Atlantic set a record for flying around the world, the main clause is
the teenager set a record for flying around the world, while the embedded clause who
previously crossed the Atlantic is an embedded clause that specifies which teenager.)
In the 80’s Fodor and Pylyshyn (1988)expressed similar concerns, with respect to an
earlier breed of neural networks. Likewise, in (Marcus, 2001), I conjectured that single
recurrent neural networks (SRNs; a forerunner to today’s more sophisticated deep
learning based recurrent neural networks, known as RNNs; Elman, 1990) would have
trouble systematically representing and extending recursive structure to various kinds of
unfamiliar sentences (see the cited articles for more specific claims about which types).
Earlier this year, Brenden Lake and Marco Baroni (2017) tested whether such pessimistic
conjectures continued to hold true. As they put it in their title, contemporary neural nets
were “Still not systematic after all these years”. RNNs could “generalize well when the
differences between training and test ... are small [but] when generalization requires
systematic compositional skills, RNNs fail spectacularly”.
Similar issues are likely to emerge in other domains, such as planning and motor control,
in which complex hierarchical structure is needed, particular when a system is likely to
encounter novel situations. One can see indirect evidence for this in the struggles with
transfer in Atari games mentioned above, and more generally in the field of robotics, in
which systems generally fail to generalize abstract plans well in novel environments.
Here’s the full Super Bowl passage; Jia and Liang’s distractor sentence that confused the model is at the end.
8
Peyton Manning became the first quarterback ever to lead two different teams to multiple Super Bowls. He is also
the oldest quarterback ever to play in a Super Bowl at age 39. The past record was held by John Elway, who led the
Broncos to victory in Super Bowl XXXIII at age 38 and is currently Denver’s Executive Vice President of Football
Operations and General Manager. Quarterback Jeff Dean had jersey number 37 in Champ Bowl XXXIV.
Page ! of !9 27
The core problem, at least at present, is that deep learning learns correlations between
sets of features that are themselves “flat” or nonhierachical, as if in a simple, unstructured
list, with every feature on equal footing. Hierarchical structure (e.g., syntactic trees that
distinguish between main clauses and embedded clauses in a sentence) are not inherently
or directly represented in such systems, and as a result deep learning systems are forced
to use a variety of proxies that are ultimately inadequate, such as the sequential position
of a word presented in a sequences.
Systems like Word2Vec (Mikolov, Chen, Corrado, & Dean, 2013) that represent
individuals words as vectors have been modestly successful; a number of systems that
have used clever tricks
try to represent complete sentences in deep-learning compatible vector spaces (Socher,
Huval, Manning, & Ng, 2012). But, as Lake and Baroni’s experiments make clear.
recurrent networks continue limited in their capacity to represent and generalize rich
structure in a faithful manner.
3.4.Deep learning thus far has struggled with open-ended inference
If you can’t represent nuance like the difference between “John promised Mary to leave”
and “John promised to leave Mary”, you can’t draw inferences about who is leaving
whom, or what is likely to happen next. Current machine reading systems have achieved
some degree of success in tasks like SQuAD, in which the answer to a given
question is explicitly contained within a text, but far less success in tasks in which
inference goes beyond what is explicit in a text, either by combining multiple sentences
(so called multi-hop inference) or by combining explicit sentences with background
knowledge that is not stated in a specific text selection. Humans, as they read texts,
frequently derive wide-ranging inferences that are both novel and only implicitly
licensed, as when they, for example, infer the intentions of a character based only on
indirect dialog.
Altough Bowman and colleagues (Bowman, Angeli, Potts, & Manning, 2015; Williams,
Nangia, & Bowman, 2017) have taken some important steps in this direction, there is, at
present, no deep learning system that can draw open-ended inferences based on real-
world knowledge with anything like human-level accuracy.
3.5.Deep learning thus far is not sufficiently transparent
The relative opacity of “black box” neural networks has been a major focus of discussion
in the last few years (Samek, Wiegand, & Müller, 2017; Ribeiro, Singh, & Guestrin,
2016). In their current incarnation, deep learning systems have millions or even billions
of parameters, identifiable to their developers not in terms of the sort of human
Page ! of !10 27
interpretable labels that canonical programmers use (“last_character_typed”) but only in
terms of their geography within a complex network (e.g., the activity value of the i
th
node
in layer j in network module k). Although some strides have been in visualizing the
contributions of individuals nodes in complex networks (Nguyen, Clune, Bengio,
Dosovitskiy, & Yosinski, 2016), most observers would acknowledge that neural networks
as a whole remain something of a black box.
How much that matters in the long run remains unclear (Lipton, 2016). If systems are
robust and self-contained enough it might not matter; if it is important to use them in the
context of larger systems, it could be crucial for debuggability.
The transparency issue, as yet unsolved, is a potential liability when using deep learning
for problem domains like financial trades or medical diagnosis, in which human users
might like to understand how a given system made a given decision. As Catherine
O’Neill (2016) has pointed out, such opacity can also lead to serious issues of bias.
3.6.Deep learning thus far has not been well integrated with prior
knowledge
The dominant approach in deep learning is hermeneutic, in the sense of being self-
contained and isolated from other, potentially usefully knowledge. Work in deep learning
typically consists of finding a training database, sets of inputs associated with respective
outputs, and learn all that is required for the problem by learning the relations between
those inputs and outputs, using whatever clever architectural variants one might devise,
along with techniques for cleaning and augmenting the data set. With just a handful of
exceptions, such as LeCun’s convolutional constraint on how neural networks are
wired(LeCun, 1989), prior knowledge is often deliberately minimized.
Thus, for example, in a system like Lerer et al’s (2016) efforts to learn about the physics
of falling towers, there is no prior knowledge of physics (beyond what is implied in
convolution). Newton’s laws, for example, are not explicitly encoded; the system instead
(to some limited degree) approximates them by learning contingencies from raw, pixel
level data. As I note in a forthcoming paper in innate (Marcus, in prep) researchers in
deep learning appear to have a very strong bias against including prior knowledge even
when (as in the case of physics) that prior knowledge is well known.
It also not straightforward in general how to integrate prior knowledge into a deep
learning system:, in part because the knowledge represented in deep learning systems
pertains mainly to (largely opaque) correlations between features, rather than to
abstractions like quantified statements (e.g. all men are mortal), see discussion of
universally-quantified one-to-one-mappings in Marcus (2001), or generics (violable
Page ! of !11 27
statements like dogs have four legs or mosquitos carry West Nile virus (Gelman, Leslie,
Was, & Koch, 2015)).
A related problem stems from a culture in machine learning that emphasizes competition
on problems that are inherently self-contained, without little need for broad general
knowledge. This tendency is well exemplified by the machine learning contest platform
known as Kaggle, in which contestants vie for the best results on a given data set.
Everything they need for a given problem is neatly packaged, with all the relevant input
and outputs files. Great progress has been made in this way; speech recognition and some
aspects of image recognition can be largely solved in the Kaggle paradigm.
The trouble, however, is that life is not a Kaggle competition; children don’t get all the
data they need neatly packaged in a single directory. Real-world learning offers data
much more sporadically, and problems aren’t so neatly encapsulated. Deep learning
works great on problems like speech recognition in which there are lots of labeled
examples, but scarcely any even knows how to apply it to more open-ended problems.
What’s the best way to fix a bicycle that has a rope caught in its spokes? Should I major
in math or neuroscience? No training set will tell us that.
Problems that have less to do with categorization and more to do with commonsense
reasoning essentially lie outside the scope of what deep learning is appropriate for, and so
far as I can tell, deep learning has little to offer such problems. In a recent review of
commonsense reasoning, Ernie Davis and I (2015) began with a set of easily-drawn
inferences that people can readily answer without anything like direct training, such as
Who is taller, Prince William or his baby son Prince George? Can you make a salad out
of a polyester shirt? If you stick a pin into a carrot, does it make a hole in the carrot or in the
pin?
As far as I know, nobody has even tried to tackle this sort of thing with deep learning.
Such apparently simple problems require humans to integrate knowledge across vastly disparate
sources, and as such are a long way from the sweet spot of deep learning-style perceptual
classification. Instead, they are perhaps best thought of as a sign that entirely different
sorts of tools are needed, along with deep learning, if we are to reach human-level
cognitive flexibility.
3.7.Deep learning thus far cannot inherently distinguish causation
from correlation
If it is a truism that causation does not equal correlation, the distinction between the two
is also a serious concern for deep learning. Roughly speaking, deep learning learns
complex correlations between input and output features, but with no inherent
Page ! of !12 27
representation of causality. A deep learning system can easily learn that height and
vocabulary are, across the population as a whole, correlated, but less easily represent the
way in which that correlation derives from growth and development (kids get bigger as
they learn more words, but that doesn’t mean that growing tall causes them to learn more
words, nor that learning new words causes them to grow). Causality has been central
strand in some other approaches to AI (Pearl, 2000) but, perhaps because deep learning is
not geared towards such challenges, relatively little work within the deep learning
tradition has tried to address it.
9
3.8.Deep learning presumes a largely stable world, in ways that may
be problematic
The logic of deep learning is such that it is likely to work best in highly stable worlds,
like the board game Go, which has unvarying rules, and less well in systems such as
politics and economics that are constantly changing. To the extent that deep learning is
applied in tasks such as stock prediction, there is a good chance that it will eventually
face the fate of Google Flu Trends, which initially did a great job of predicting
epidemological data on search trends, only to complete miss things like the peak of the
2013 flu season (Lazer, Kennedy, King, & Vespignani, 2014).
3.9. Deep learning thus far works well as an approximation, but its
answers often cannot be fully trusted
In part as a consequence of the other issues raised in this section, deep learning systems
are quite good at some large fraction of a given domain, yet easily fooled.
An ever-growing array of papers has shown this vulnerability, from the linguistic
examples of Jia and Liang mentioned above to a wide range of demonstrations in the
domain of vision, where deep learning systems have mistaken yellow-and-black patterns
of stripes for school buses (Nguyen, Yosinski, & Clune, 2014) and sticker-clad parking
signs for well-stocked refrigerators (Vinyals, Toshev, Bengio, & Erhan, 2014) in the
context of a captioning system that otherwise seems impressive.
More recently, there have been real-world stop signs, lightly defaced, that have been
mistaken for speed limit signs (Evtimov et al., 2017) and 3d-printed turtles that have been
mistake for rifles (Athalye, Engstrom, Ilyas, & Kwok, 2017). A recent news story
One example of interesting recent work is (Lopez-Paz, Nishihara, Chintala, Schölkopf, & Bottou, 2017), albeit
9
focused specifically on an rather unusual sense of the term causation as it relates to the presence or absence of
objects (e.g., “the presence of cars cause the presence of wheel[s]). This strikes me as quite different from the sort of
causation one finds in the relation between a disease and the symptoms it causes.
Page ! of !13 27
recounts the trouble a British police system has had in distinguishing nudes from sand
dunes.
10
The “spoofability” of deep learning systems was perhaps first noted by Szegedy et
al(2013). Four years later, despite much active research, no robust solution has been
found.
11
3.10. Deep learning thus far is difficult to engineer with
Another fact that follows from all the issues raised above is that is simply hard to do
robust engineering with deep learning. As a team of authors at Google put it in 2014, in
the title of an important, and as yet unanswered essay (Sculley, Phillips, Ebner,
Chaudhary, & Young, 2014), machine learning is “the high-interest credit card of
technical debt”, meaning that is comparatively easy to make systems that work in some
limited set of circumstances (short term gain), but quite difficult to guarantee that they
will work in alternative circumstances with novel data that may not resemble previous
training data (long term debt, particularly if one system is used as an element in another
larger system).
In an important talk at ICML, Leon Bottou (2015) compared machine learning to the
development of an airplane engine, and noted that while the airplane design relies on
building complex systems out of simpler systems for which it was possible to create
sound guarantees about performance, machine learning lacks the capacity to produce
comparable guarantees. As Google’s Peter Norvig (2016) has noted, machine learning as
yet lacks the incrementality, transparency and debuggability of classical programming,
trading off a kind of simplicity for deep challenges in achieving robustness.
Henderson and colleagues have recently extended these points, with a focus on deep
reinforcement learning, noting some serious issues in the field related to robustness and
replicability (Henderson et al., 2017).
Although there has been some progress in automating the process of developing machine
learning systems (Zoph, Vasudevan, Shlens, & Le, 2017), there is a long way to go.
https://gizmodo.com/british-cops-want-to-use-ai-to-spot-porn-but-it-keeps-m-1821384511/amp
10
Deep learning’s predecessors were vulnerable to similar problems, as Pinker and Prince (1988)pointed out, in a
11
discussion of neural networks that produced bizarre past tense forms for a subset of its inputs. The verb to mail, for
example, was inflected in the past tense as membled, the verb tour as toureder. Children rarely if ever make mistakes
like these.
Page ! of !14 27
3.11. Discussion
Of course, deep learning, is by itself, just mathematics; none of the problems identified
above are because the underlying mathematics of deep learning are somehow flawed. In
general, deep learning is a perfectly fine way of optimizing a complex system for
representing a mapping between inputs and outputs, given a sufficiently large data set.
The real problem lies in misunderstanding what deep learning is, and is not, good for. The
technique excels at solving closed-end classification problems, in which a wide range of
potential signals must be mapped onto a limited number of categories, given that there is
enough data available and the test set closely resembles the training set.
But deviations from these assumptions can cause problems; deep learning is just a
statistical technique, and all statistical techniques suffer from deviation from their
assumptions.
Deep learning systems work less well when there are limited amounts of training data
available, or when the test set differs importantly from the training set, or when the space
of examples is broad and filled with novelty. And some problems cannot, given real-
world limitations, be thought of as classification problems at all. Open-ended natural
language understanding, for example, should not be thought of as a classifier mapping
between a large finite set of sentences and large, finite set of sentences, but rather a
mapping between a potentially infinite range of input sentences and an equally vast array
of meanings, many never previously encountered. In a problem like that, deep learning
becomes a square peg slammed into a round hole, a crude approximation when there
must be a solution elsewhere.
One clear way to get an intuitive sense of why something is amiss to consider a set of
experiments I did long ago, in 1997, when I tested some simplified aspects of language
development on a class of neural networks that were then popular in cognitive science.
The 1997-vintage networks were, to be sure, simpler than current models — they used no
more than three layers (inputs nodes connected to hidden nodes connected to outputs
node), and lacked Lecun’s powerful convolution technique. But they were driven by
backpropagation just as today’s systems are, and just as beholden to their training data.
In language, the name of the game is generalization — once I hear a sentence like John
pilked a football to Mary, I can infer that is also grammatical to say John pilked Mary the
football, and Eliza pilked the ball to Alec; equally if I can infer what the word pilk means,
I can infer what the latter sentences would mean, even if I had not hear them before.
Page ! of !15 27
Distilling the broad-ranging problems of language down to a simple example that I
believe still has resonance now, I ran a series of experiments in which I trained three-
layer perceptrons (fully connected in today’s technical parlance, with no convolution) on
the identity function, f(x) = x, e.g, f(12)=12.
Training examples were represented by a set of input nodes (and corresponding output
nodes) that represented numbers in terms of binary digits. The number 7 for example,
would be represented by turning on the input (and output) nodes representing 4, 2, and 1.
As a test of generalization, I trained the network on various sets of even numbers, and
tested it all possible inputs, both odd and even.
Every time I ran the experiment, using a wide variety of parameters, the results were the
same: the network would (unless it got stuck in local minimum) correctly apply the
identity function to the even numbers that it had seen before (say 2, 4, 8 and 12), and to
some other even numbers (say 6 and 14) but fail on all the odds numbers, yielding, for
example f(15) = 14.
In general, the neural nets I tested could learn their training examples, and interpolate to a
set of test examples that were in a cloud of points around those examples in n-
dimensional space (which I dubbed the training space), but they could not extrapolate
beyond that training space.
Odd numbers were outside the training space, and the networks could not generalize
identity outside that space. Adding more hidden units didn’t help, and nor did adding
12
more hidden layers. Simple multilayer perceptrons simply couldn’t generalize outside
their training space (Marcus, 1998a; Marcus, 1998b; Marcus, 2001). (Chollet makes quite
similar points in the closing chapters of his his (Chollet, 2017) text.)
What we have seen in this paper is that challenges in generalizing beyond a space of
training examples persist in current deep learning networks, nearly two decades later.
Many of the problems reviewed in this paper — the data hungriness, the vulnerability to
fooling, the problems in dealing with open-ended inference and transfer — can be seen
as extension of this fundamental problem. Contemporary neural networks do well on
challenges that remain close to their core training data, but start to break down on cases
further out in the periphery.
Of course, the network had never seen an odd number before, but pretraining the network on odd numbers in a
12
different context didn’t help. And of course people, in contrast, readily generalize to novel words immediately upon
hearing them. Likewise, the experiments I did with seven-month-olds consisted entirely of novel words.
Page ! of !16 27
The widely-adopted addition of convolution guarantees that one particular class of
problems that are akin to my identity problem can be solved: so-called translational
invariances, in which an object retains its identity when it is shifted to a location. But the
solution is not general, as for example Lake’s recent demonstrations show. (Data
augmentation offers another way of dealing with deep learning’s challenges in
extrapolation, by trying to broaden the space of training examples itself, but such
techniques are more useful in 2d vision than in language).
As yet there is no general solution within deep learning to the problem of generalizing
outside the training space. And it is for that reason, more than any other, that we need to
look to different kinds of solutions if we want to reach artificial general intelligence.
4. Potential risks of excessive hype
One of the biggest risks in the current overhyping of AI is another AI winter, such as the
one that devastated the field in the 1970’s, after the Lighthill report (Lighthill, 1973),
suggested that AI was too brittle, too narrow and too superficial to be used in practice.
Although there are vastly more practical applications of AI now than there were in the
1970s, hype is still a major concern. When a high-profile figure like Andrew Ng writes in
the Harvard Business Review promising a degree of imminent automation that is out of
step with reality, there is fresh risk for seriously dashed expectations. Machines cannot in
fact do many things that ordinary humans can do in a second, ranging from reliably
comprehending the world to understanding sentences. No healthy human being would
ever mistake a turtle for a rifle or parking sign for a refrigerator.
Executives investing massively in AI may turn out to be disappointed, especially given
the poor state of the art in natural language understanding. Already, some major projects
have been largely abandoned, like Facebook’s M project, which was launched in August
2015 with much publicity as a general purpose personal assistant, and then later
13
downgraded to a significantly smaller role, helping users with a vastly small range of
well-defined tasks such as calendar entry.
It is probably fair to say that chatbots in general have not lived up to the hype they
received a couple years ago. If, for example, driverless car should also, disappoint,
relative to their early hype, by proving unsafe when rolled out at scale, or simply not
achieving full autonomy after many promises, the whole field of AI could be in for a
sharp downturn, both in popularity and funding. We already may be seeing hints of this,
https://www.wired.com/2015/08/how-facebook-m-works/
13
Page ! of !17 27
as in a just published Wired article that was entitled “After peak hype, self-driving cars
14
enter the trough of disillusionment.”
There are other serious fears, too, and not just of the apocalyptic variety (which for now
to still seem to be stuff of science fiction). My own largest fear is that the field of AI
could get trapped in a local minimum, dwelling too heavily in the wrong part of
intellectual space, focusing too much on the detailed exploration of a particular class of
accessible but limited models that are geared around capturing low-hanging fruit —
potentially neglecting riskier excursions that might ultimately lead to a more robust path.
I am reminded of Peter Thiel’s famous (if now slightly outdated) damning of an often
too-narrowly focused tech industry: “We wanted flying cars, instead we got 140
characters”. I still dream of Rosie the Robost, a full-service domestic robot that take of
my home; but for now, six decades into the history of AI, our bots do little more than play
music, sweep floors, and bid on advertisements.
If didn’t make more progress, it would be a shame. AI comes with risk, but also great
potential rewards. AI’s greatest contributions to society, I believe, could and should
ultimately come in domains like automated scientific discovery, leading among other
things towards vastly more sophisticated versions of medicine than are currently possible.
But to get there we need to make sure that the field as whole doesn’t first get stuck in a
local minimum.
5. What would be better?
Despite all of the problems I have sketched, I don’t think that we need to abandon deep
learning.
Rather, we need to reconceptualize it: not as a universal solvent, but simply as one tool
among many, a power screwdriver in a world in which we also need hammers, wrenches,
and pliers, not to mentions chisels and drills, voltmeters, logic probes, and oscilloscopes.
In perceptual classification, where vast amounts of data are available, deep learning is a
valuable tool; in other, richer cognitive domains, it is often far less satisfactory.
The question is, where else should we look? Here are four possibilities.
https://www.wired.com/story/self-driving-cars-challenges/
14
Page ! of !18 27
5.1.Unsupervised learning
In interviews, deep learning pioneers Geoff Hinton and Yann LeCun have both recently
pointed to unsupervised learning as one key way in which to go beyond supervised, data-
hungry versions of deep learning.
To be clear, deep learning and unsupervised learning are not in logical opposition. Deep
learning has mostly been used in a supervised context with labeled data, but there are
ways of using deep learning in an unsupervised fashion. But there is certainly reasons in
many domains to move away from the massive demands on data that supervised deep
learning typically requires.
Unsupervised learning, as the term is commonly used, tends to refer to several kinds of
systems. One common type of system “clusters” together inputs that share properties,
even without having them explicitly labeled. Google’s cat detector model (Le et al., 2012)
is perhaps the most publicly prominent example of this sort of approach.
Another approach, advocated researchers such as Yann LeCun (Luc, Neverova, Couprie,
Verbeek, & LeCun, 2017), and not mutually exclusive with the first, is to replace labeled
data sets with things like movies that change over time. The intuition is that systems
trained on videos can use each pair of successive frames as a kind of ersatz teaching
signal, in which the goal is to predict the next frame; frame t becomes a predictor for
frame t1, without the need for any human labeling.
My view is that both of these approaches are useful (and so are some others not discussed
here), but that neither inherently solve the sorts of problems outlined in section 3. One is
still left with data hungry systems that lack explicit variables, and I see no advance there
towards open-ended inference, interpretability or debuggability.
That said, there is a different notion of unsupervised learning, less discussed, which I find
deeply interesting: the kind of unsupervised learning that human children do. Children
often y set themselves a novel task, like creating a tower of Lego bricks or climbing
through a small aperture, as my daughter recently did in climbing through a chair, in the
space between the seat and the chair back . Often, this sort of exploratory problem
solving involves (or at least appears to involve) a good deal of autonomous goal setting
(what should I do?) and high level problem solving (how do I get my arm through the
chair, now that the rest of my body has passed through?), as well the integration of
abstract knowledge (how bodies work, what sorts of apertures and affordances various
objects have, and so forth). If we could build systems that could set their own goals and
do reasoning and problem-solving at this more abstract level, major progress might
quickly follow.
Page ! of !19 27
5.2.Symbol-manipulation, and the need for hybrid models
Another place that we should look is towards classic, “symbolic” AI, sometimes referred
to as GOFAI (Good Old-Fashioned AI). Symbolic AI takes its name from the idea, central
to mathematics, logic, and computer science, that abstractions can be represented by
symbols. Equations like f = ma allow us to calculate outputs for a wide range of inputs,
irrespective of whether we have seen any particular values before; lines in computer
programs do the same thing (if the value of variable x is greater than the value of variable
y, perform action a).
By themselves, symbolic systems have often proven to be brittle, but they were largely
developed in era with vastly less data and computational power than we have now. The
right move today may be to integrate deep learning, which excels at perceptual
classification, with symbolic systems, which excel at inference and abstraction. One
might think such a potential merger on analogy to the brain; perceptual input systems,
like primary sensory cortex, seem to do something like what deep learning does, but there
are other areas, like Broca’s area and prefrontal cortex, that seem to operate at much
higher level of abstraction. The power and flexibility of the brain comes in part from its
capacity to dynamically integrate many different computations in real-time. The process
of scene perception, for instance, seamlessly integrates direct sensory information with
complex abstractions about objects and their properties, lighting sources, and so forth.
Some tentative steps towards integration already exist, including neurosymbolic
modeling (Besold et al., 2017) and recent trend towards systems such as differentiable
neural computers (Graves et al., 2016), programming with differentiable interpreters
(Bošnjak, Rocktäschel, Naradowsky, & Riedel, 2016), and neural programming with
discrete operations (Neelakantan, Le, Abadi, McCallum, & Amodei, 2016). While none
of this work has yet fully scaled towards anything like full-service artificial general
intelligence, I have long argued (Marcus, 2001) that more on integrating microprocessor-
like operations into neural networks could be extremely valuable.
To the extent that the brain might be seen as consisting of “a broad array of reusable
computational primitives—elementary units of processing akin to sets of basic
instructions in a microprocessor—perhaps wired together in parallel, as in the
reconfigurable integrated circuit type known as the field-programmable gate array”, as I
have argued elsewhere(Marcus, Marblestone, & Dean, 2014), steps towards enriching the
instruction set out of which our computational systems are built can only be a good thing.
Page ! of !20 27
5.3.More insight from cognitive and developmental psychology
Another potential valuable place to look is human cognition (Davis & Marcus, 2015;
Lake et al., 2016; Marcus, 2001; Pinker & Prince, 1988). There is no need for machines
to literally replicate the human mind, which is, after all, deeply error prone, and far from
perfect. But there remain many areas, from natural language understanding to
commonsense reasoning, in which humans still retain a clear advantage; learning the
mechanisms underlying those human strengths could lead to advances in AI, even the
goal is not, and should not be, an exact replica of human brain.
For many people, learning from humans means neuroscience; in my view, that may be
premature. We don’t yet know enough about neuroscience to literally reverse engineer the
brain, per se, and may not for several decades, possibly until AI itself gets better. AI can
help us to decipher the brain, rather than the other way around.
Either way, in the meantime, it should certainly be possible to use techniques and insights
drawn from cognitive and developmental and psychology, now, in order to build more
robust and comprehensive artificial intelligence, building models that are motivated not
just by mathematics but also by clues from the strengths of human psychology.
A good starting point might be to first to try understand the innate machinery in humans
minds, as a source of hypotheses into mechanisms that might be valuable in developing
artificial intelligences; in companion article to this one (Marcus, in prep) I summarize a
number of possibilities, some drawn from my own earlier work (Marcus, 2001) and
others from Elizabeth Spelke’s (Spelke & Kinzler, 2007). Those drawn from my own
work focus on how information might be represented and manipulated, such as by
symbolic mechanisms for representing variables and distinctions between kinds and
individuals from a class; those drawn from Spelke focus on how infants might represent
notions such as space, time, and object.
A second focal point might be on common sense knowledge, both in how it develops
(some might be part of our innate endowment, much of it is learned), how it is
represented, and how it is integrated on line in the process of our interactions with the
real world (Davis & Marcus, 2015). Recent work by Lerer et al (2016), Watters and
colleagues (2017), Tenenbaum and colleagues(Wu, Lu, Kohli, Freeman, & Tenenbaum,
2017) and Davis and myself (Davis, Marcus, & Frazier-Logue, 2017) suggest some
competing approaches to how to think about this, within the domain of everyday physical
reasoning.
Page ! of !21 27
A third focus might be on human understanding of narrative, a notion long ago suggested
by Roger Schank and Abelson (1977) and due for a refresh (Marcus, 2014; Kočiský et al.,
2017).
5.4.Bolder challenges
Whether deep learning persists in current form, morphs into something new, or gets
replaced altogether, one might consider a variety of challenge problems that push systems
to move beyond what can be learned in supervised learning paradigms with large
datasets. Drawing in part of from a recent special issue of AI Magazine devoted to
moving beyond the Turing Test that I edited with Francesca Rossi, Manuelo Veloso
(Marcus, Rossi, Veloso - AI Magazine, & 2016, 2016), here are a few suggestions:
•
A comprehension challenge (Paritosh & Marcus, 2016; Kočiský et al., 2017)] which
would require a system to watch an arbitrary video (or read a text, or listen to a
podcast) and answer open-ended questions about what is contained therein. (Who is the
protagonist? What is their motivation? What will happen if the antagonist succeeds in
her mission?) No specific supervised training set can cover all the possible
contingencies; infererence and real-world knowledge integration are necessities.
•
Scientific reasoning and understanding, as in the Allen AI institute’s 8th grade science
challenge (Schoenick, Clark, Tafjord, P, & Etzioni, 2017; Davis, 2016). While the
answers to many basic science questions can simply be retrieved from web searches,
others require inference beyond what is explicitly stated, and the integration of general
knowledge.
•
General game playing (Genesereth, Love, & Pell, 2005), with transfer between games
(Kansky et al., 2017), such that, for example, learning about one first-person shooter
enhances performance on another with entirely different images, equipment and so
forth. (A system that can learn many games, separately, without transfer between them,
such as DeepMind’s Atari game system, would not qualify; the point is to acquire
cumulative, transferrable knowledge).
•
A physically embodied test an AI-driven robot that could build things (Ortiz Jr, 2016),
ranging from tents to IKEA shelves, based on instructions and real-world physical
interactions with the objects parts, rather than vast amounts trial-and-error.
No one challenge is likely to be sufficient. Natural intelligence is multi-dimensional
(Gardner, 2011), and given the complexity of the world, generalized artificial intelligence
will necessarily be multi-dimensional as well.
By pushing beyond perceptual classification and into a broader integration of inference
and knowledge, artificial intelligence will advance, greatly.
Page ! of !22 27
6. Conclusions
As a measure of progress, it is worth considering a somewhat pessimistic piece I wrote
for The New Yorker five years ago , conjecturing that “deep learning is only part of the
15
larger challenge of building intelligent machines” because “such techniques lack ways of
representing causal relationships (such as between diseases and their symptoms), and are
likely to face challenges in acquiring abstract ideas like “sibling” or “identical to.” They
have no obvious ways of performing logical inferences, and they are also still a long way
from integrating abstract knowledge, such as information about what objects are, what
they are for, and how they are typically used.”
As we have seen, many of these concerns remain valid, despite major advances in
specific domains like speech recognition, machine translation, and board games, and
despite equally impressive advances in infrastructure and the amount of data and compute
available.
Intriguingly, in the last year, a growing array of other scholars, coming from an
impressive range of perspectives, have begun to emphasize similar limits. A partial list
includes Brenden Lake and Marco Baroni (2017), François Chollet (2017), Robin Jia and
Percy Liang (2017), Dileep George and others at Vicarious (Kansky et al., 2017) and
Pieter Abbeel and colleagues at Berkeley (Stoica et al., 2017).
Perhaps most notably of all, Geoff Hinton has been courageous enough to reconsider has
own beliefs, revealing in an August interview with the news site Axios that he is
16
“deeply suspicious” of back-propagation, a key enabler of deep learning that he helped
pioneer, because of his concern about its dependence on labeled data sets.
Instead, he suggested (in Axios’ paraphrase) that “entirely new methods will probably
have to be invented.”
I share Hinton’s excitement in seeing what comes next.
https://www.newyorker.com/news/news-desk/is-deep-learning-a-revolution-in-artificial-intelligence
15
https://www.axios.com/ai-pioneer-advocates-starting-over-2485537027.html
16
Page ! of !23 27
References
Athalye, A., Engstrom, L., Ilyas, A., & Kwok, K. (2017). Synthesizing Robust Adversarial
Examples. arXiv, cs.CV.
Besold, T. R., Garcez, A. D., Bader, S., Bowman, H., Domingos, P., Hitzler, P. et al. (2017).
Neural-Symbolic Learning and Reasoning: A Survey and Interpretation. arXiv, cs.AI.
Bošnjak, M., Rocktäschel, T., Naradowsky, J., & Riedel, S. (2016). Programming with a
Differentiable Forth Interpreter. arXiv.
Bottou, L. (2015). Two big challenges in machine learning. Proceedings from 32nd International
Conference on Machine Learning.
Bowman, S. R., Angeli, G., Potts, C., & Manning, C. D. (2015). A large annotated corpus for
learning natural language inference. arXiv, cs.CL.
Chollet, F. (2017). Deep Learning with Python. Manning Publications.
Cireşan, D., Meier, U., Masci, J., & Schmidhuber, J. (2012). Multi-column deep neural network
for traffic sign classification. Neural networks.
Davis, E., & Marcus, G. (2015). Commonsense reasoning and commonsense knowledge in
artificial intelligence. Communications of the ACM, 58(9)(9), 92-103.
Davis, E. (2016). How to Write Science Questions that Are Easy for People and Hard for
Computers. AI magazine, 37(1)(1), 13-22.
Davis, E., Marcus, G., & Frazier-Logue, N. (2017). Commonsense reasoning about containers
using radically incomplete information. Artificial Intelligence, 248, 46-84.
Deng, J., Dong, W., Socher, R., Li, L. J., Li - Computer Vision and, K., & 2009 Imagenet: A
large-scale hierarchical image database. Proceedings from Computer Vision and Pattern
Recognition, 2009. CVPR 2009. IEEE Conference on.
Elman, J. L. (1990). Finding structure in time. Cognitive science, 14(2)(2), 179-211.
Evtimov, I., Eykholt, K., Fernandes, E., Kohno, T., Li, B., Prakash, A. et al. (2017). Robust
Physical-World Attacks on Deep Learning Models. arXiv, cs.CR.
Fodor, J. A., & Pylyshyn, Z. W. (1988). Connectionism and cognitive architecture: a critical
analysis. Cognition, 28(1-2)(1-2), 3-71.
Gardner, H. (2011). Frames of mind: The theory of multiple intelligences. Basic books.
Gelman, S. A., Leslie, S. J., Was, A. M., & Koch, C. M. (2015). Children’s interpretations of
general quantifiers, specific quantifiers, and generics. Lang Cogn Neurosci, 30(4)(4),
448-461.
Genesereth, M., Love, N., & Pell, B. (2005). General game playing: Overview of the AAAI
competition. AI magazine, 26(2)(2), 62.
George, D., Lehrach, W., Kansky, K., Lázaro-Gredilla, M., Laan, C., Marthi, B. et al. (2017). A
generative vision model that trains with high data efficiency and breaks text-based
CAPTCHAs. Science, 358(6368)(6368).
Gervain, J., Berent, I., & Werker, J. F. (2012). Binding at birth: the newborn brain detects identity
relations and sequential position in speech. J Cogn Neurosci, 24(3)(3), 564-574.
Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT press.
Page ! of !24 27
Graves, A., Wayne, G., Reynolds, M., Harley, T., Danihelka, I., Grabska-Barwińska, A. et al.
(2016). Hybrid computing using a neural network with dynamic external memory. Nature,
538(7626)(7626), 471-476.
Henderson, P., Islam, R., Bachman, P., Pineau, J., Precup, D., & Meger, D. (2017). Deep
Reinforcement Learning that Matters. arXiv, cs.LG.
Huang, S., Papernot, N., Goodfellow, I., Duan, Y., & Abbeel, P. (2017). Adversarial Attacks on
Neural Network Policies. arXiv, cs.LG.
Jia, R., & Liang, P. (2017). Adversarial Examples for Evaluating Reading Comprehension
Systems. arXiv.
Kahneman, D. (2013). Thinking, fast and slow (1st pbk. ed. ed.). New York: Farrar, Straus and
Giroux.
Kansky, K., Silver, T., Mély, D. A., Eldawy, M., Lázaro-Gredilla, M., Lou, X. et al. (2017).
Schema Networks: Zero-shot Transfer with a Generative Causal Model of Intuitive
Physics. arXIv, cs.AI.
Kočiský, T., Schwarz, J., Blunsom, P., Dyer, C., Hermann, K. M., Melis, G. et al. (2017). The
NarrativeQA Reading Comprehension Challenge. arXiv, cs.CL.
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep
convolutional neural networks. In (pp. 1097-1105).
Lake, B. M., Salakhutdinov, R., & Tenenbaum, J. B. (2015). Human-level concept learning
through probabilistic program induction. Science, 350(6266)(6266), 1332-1338.
Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2016). Building Machines
That Learn and Think Like People. Behav Brain Sci, 1-101.
Lake, B. M., & Baroni, M. (2017). Still not systematic after all these years: On the compositional
skills of sequence-to-sequence recurrent networks. arXiv.
Lazer, D., Kennedy, R., King, G., & Vespignani, A. (2014). Big data. The parable of Google Flu:
traps in big data analysis. Science, 343(6176)(6176), 1203-1205.
Le, Q. V., Ranzato, M.-A., Monga, R., Devin, M., Chen, K., Corrado, G. et al. (2012). Building
high-level features using large scale unsupervised learning. Proceedings from International
Conference on Machine Learning.
LeCun, Y. (1989). Generalization and network design strategies. Technical Report CRG-TR-89-4.
Lerer, A., Gross, S., & Fergus, R. (2016). Learning Physical Intuition of Block Towers by
Example. arXiv, cs.AI.
Lighthill, J. (1973). Artificial Intelligence: A General Survey. Artificial Intelligence: a paper
symposium.
Lipton, Z. C. (2016). The Mythos of Model Interpretability. arXiv, cs.LG.
Lopez-Paz, D., Nishihara, R., Chintala, S., Schölkopf, B., & Bottou, L. (2017). Discovering
causal signals in images. Proceedings from Proceedings of Computer Vision and Pattern
Recognition (CVPR).
Luc, P., Neverova, N., Couprie, C., Verbeek, J., & LeCun, Y. (2017). Predicting Deeper into the
Future of Semantic Segmentation. International Conference on Computer Vision (ICCV
2017).
Page ! of !25 27
Marcus, G., Rossi, F., Veloso - AI Magazine, M., & 2016. (2016). Beyond the Turing Test. AI
Magazine, Whole issue.
Marcus, G., Marblestone, A., & Dean, T. (2014). The atoms of neural computation. Science,
346(6209)(6209), 551-552.
Marcus, G. (in prep). Innateness, AlphaZero, and Artificial Intelligence.
Marcus, G. (2014). What Comes After the Turing Test? The New Yorker.
Marcus, G. (2012). Is “Deep Learning” a Revolution in Artificial Intelligence? The New Yorker.
Marcus, G. F. (2008). Kluge : the haphazard construction of the human mind. Boston: Houghton
Mifflin.
Marcus, G. F. G. F. (2001). The Algebraic Mind: Integrating Connectionism and cognitive
science. Cambridge, Mass.: MIT Press.
Marcus, G. F. (1998a). Rethinking eliminative connectionism. Cogn Psychol, 37(3)(3), 243-282.
Marcus, G. F. (1998b). Can connectionism save constructivism? Cognition, 66(2)(2), 153-182.
Marcus, G. F., Pinker, S., Ullman, M., Hollander, M., Rosen, T. J., & Xu, F. (1992).
Overregularization in language acquisition. Monogr Soc Res Child Dev, 57(4)(4), 1-182.
Marcus, G. F., Vijayan, S., Bandi Rao, S., & Vishton, P. M. (1999). Rule learning by seven-
month-old infants. Science, 283(5398)(5398), 77-80.
Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient Estimation of Word
Representations in Vector Space. arXiv.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G. et al. (2015).
Human-level control through deep reinforcement learning. Nature, 518(7540)(7540),
529-533.
Neelakantan, A., Le, Q. V., Abadi, M., McCallum, A., & Amodei, D. (2016). Learning a Natural
Language Interface with Neural Programmer. arXiv.
Ng, A. (2016). What Artificial Intelligence Can and Can’t Do Right Now. Harvard Business
Review.
Nguyen, A., Clune, J., Bengio, Y., Dosovitskiy, A., & Yosinski, J. (2016). Plug & Play
Generative Networks: Conditional Iterative Generation of Images in Latent Space. arXiv,
cs.CV.
Nguyen, A., Yosinski, J., & Clune, J. (2014). Deep Neural Networks are Easily Fooled: High
Confidence Predictions for Unrecognizable Images. arXiv, cs.CV.
Norvig, P. (2016). State-of-the-Art AI: Building Tomorrow’s Intelligent Systems. Proceedings
from EmTech Digital, San Francisco.
O’Neil, C. (2016). Weapons of math destruction : how big data increases inequality and threatens
democracy.
Ortiz Jr, C. L. (2016). Why we need a physically embodied Turing test and what it might look
like. AI magazine, 37(1)(1), 55-63.
Paritosh, P., & Marcus, G. (2016). Toward a comprehension challenge, using crowdsourcing as a
tool. AI Magazine, 37(1)(1), 23-31.
Pearl, J. (2000). Causality : models, reasoning, and inference /. Cambridge, U.K.; New York :
Cambridge University Press.
Page ! of !26 27
Pinker, S., & Prince, A. (1988). On language and connectionism: analysis of a parallel distributed
processing model of language acquisition. Cognition, 28(1-2)(1-2), 73-193.
Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). “Why Should I Trust You?”: Explaining the
Predictions of Any Classifier. arXiv, cs.LG.
Sabour, S., Frosst, N., & Hinton, G. E. (2017). Dynamic Routing Between Capsules. arXiv,
cs.CV.
Samek, W., Wiegand, T., & Müller, K.-R. (2017). Explainable Artificial Intelligence:
Understanding, Visualizing and Interpreting Deep Learning Models. arXiv, cs.AI.
Schank, R. C., & Abelson, R. P. (1977). Scripts, Plans, Goals and Understanding: an Inquiry into
Human Knowledge Structures. Hillsdale, NJ: L. Erlbaum.
Schmidhuber, J. (2015). Deep learning in neural networks: An overview. Neural networks.
Schoenick, C., Clark, P., Tafjord, O., P, T., & Etzioni, O. (2017). Moving beyond the Turing Test
with the Allen AI Science Challenge. Communications of the ACM, 60 (9)(9), 60-64.
Sculley, D., Phillips, T., Ebner, D., Chaudhary, V., & Young, M. (2014). Machine learning: The
high-interest credit card of technical debt. Proceedings from SE4ML: Software
Engineering for Machine Learning (NIPS 2014 Workshop).
Socher, R., Huval, B., Manning, C. D., & Ng, A. Y. (2012). Semantic compositionality through
recursive matrix-vector spaces. Proceedings from Proceedings of the 2012 joint conference
on empirical methods in natural language processing and computational natural language
learning.
Spelke, E. S., & Kinzler, K. D. (2007). Core knowledge. Dev Sci, 10(1)(1), 89-96.
Stoica, I., Song, D., Popa, R. A., Patterson, D., Mahoney, M. W., Katz, R. et al. (2017). A
Berkeley View of Systems Challenges for AI. arXiv, cs.AI.
Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I. et al. (2013).
Intriguing properties of neural networks. arXiv, cs.CV.
Vinyals, O., Toshev, A., Bengio, S., & Erhan, D. (2014). Show and Tell: A Neural Image Caption
Generator. arXiv, cs.CV.
Watters, N., Tacchetti, A., Weber, T., Pascanu, R., Battaglia, P., & Zoran, D. (2017). Visual
Interaction Networks. arXiv.
Williams, A., Nangia, N., & Bowman, S. R. (2017). A Broad-Coverage Challenge Corpus for
Sentence Understanding through Inference. arXiv, cs.CL.
Wu, J., Lu, E., Kohli, P., Freeman, B., & Tenenbaum, J. (2017). Learning to See Physics via
Visual De-animation. Proceedings from Advances in Neural Information Processing
Systems.
Zoph, B., Vasudevan, V., Shlens, J., & Le, Q. V. (2017). Learning Transferable Architectures for
Scalable Image Recognition. arXiv, cs.CV.
Page ! of !27 27
At this point, the "diminishing returns" that Chollet is quoted above as saying have begun to set in...(we can clearly see that in the human's performance in the below graph ;) )
Source: https://arxiv.org/pdf/1803.01164.pdf
Table with results:
Source: https://medium.com/zylapp/review-of-deep-learning-algorithms-for-image-classification-5fdbca4a05e2
And the definition of "multiple" has change over time too. Now, the number of layers in the best performing models keeps increasing as we keep making computational advancements that were not feasible in the past...
Here is a tutorial video on backpropagation: https://www.youtube.com/watch?v=Ilg3gGewQ5U
However, unsupervised learning (does not require labels), has achieved incredible milestones in natural language understanding.
Nice article on the topic: https://deepmind.com/blog/article/unsupervised-learning
Here is a curated collection of many of the classic deep learning papers: https://github.com/terryum/awesome-deep-learning-papers
Link to paper: http://yann.lecun.com/exdb/publis/pdf/lecun-89.pdf
Here is a progress report on "One-Shot learning" which is related to what Marcus talks about here. It is by Brendan Lake and his colleagues and published in 2019: https://www.sciencedirect.com/science/article/pii/S2352154619300051
However, transfer learning has had some very impressive results...
This is a big drawback, and there are a lot of people trying to address explainability in AI (i.e. saliency maps and other visualization techniques, among many other areas). https://en.wikipedia.org/wiki/Explainable_artificial_intelligence
Hence the rise of Bayesian Deep Learning: http://bayesiandeeplearning.org/
Hence the rise of interest in the field of causal inference in machine learning: https://why19.causalai.net/
Gary F. Marcus is a scientist, author, and entrepreneur, focusing on natural and artificial intelligence, and teaching at NYU. Some of his writing for wired on this topic: https://www.wired.com/author/gary-marcus/
Imagenet is an image database with over 14 million hand-annotated images in 20 thousand different categories that computer vision/machine learning researchers use to train/test/benchmark the performance of their object recognition algorithms.
Many computer vision algorithms are trained on Imagenet, and the balance/distribution of images in the database has only recently come under scrutiny. Researchers have recently exposed biases that the unbalanced Imagenet database leads to. Here is Imagenet roulette: https://www.nytimes.com/2019/09/20/arts/design/imagenet-trevor-paglen-ai-facial-recognition.html.
The Imagenet database: http://www.image-net.org/
For more on imagenet's history: I️https://en.wikipedia.org/wiki/ImageNet