Research-Technology Management

ISSN: 0895-6308 (Print) 1930-0166 (Online) Journal homepage: http://www.tandfonline.com/loi/urtm20

The Power and Limits of Deep Learning

Yann LeCun

To cite this article: Yann LeCun (2018) The Power and Limits of Deep Learning, Research-
Technology Management, 61:6, 22-27, DOI: 10.1080/08956308.2018.1516928

To link to this article: https://doi.org/10.1080/08956308.2018.1516928

Published online: 05 Nov 2018.

Submit your article to this journal

View Crossmark data

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=urtm20
IRI MEDAL

The Power and Limits of Deep Learning

In his IRI Medal address, Yann LeCun maps the development of machine learning techniques and suggests what the future may
hold.

Yann LeCun

Artificial intelligence (AI) is advancing very rapidly. I’ve had typically proceeds by trial and error: the algorithm is shown
a front-row seat for a lot of the recent progress—first at Bell examples and corrects itself when it makes mistakes, instead
Labs (which was renamed AT&T Labs in 1996, while I was of being explicitly programmed to follow a predetermined
there) and then at the NEC Research Institute in Princeton. I set of instructions.
joined academia after these stints in industrial research. Now,
I have one foot in each world, or at least half a foot in academia Supervised Learning and Deep Learning
and one-and-a-half feet in industry. I joined Facebook at the Almost all practical applications of machine learning are
end of 2013 for the purpose of building Facebook AI Research based on supervised learning. Supervised learning is a process
(FAIR), which was created because Mark Zuckerberg knew in which you train the system by telling it what the correct
that AI was going to be a critical piece of technology for output is from a given set of input examples. You collect a
Facebook and for connecting people with each other. I spend large dataset—let’s say of images of cars and airplanes—and
most of my time at Facebook, but I still teach, and I have a lab you feed that data to the machine. You show the machine a
with PhD students at NYU. I think the two are complementary car, for instance, and if it doesn’t say, “Car,” you tell it, “You
worlds, and I find that very enriching. got it wrong. The correct answer is car.” The machine then
FAIR was founded to focus on research in AI. In the past, adjusts its internal parameters so that next time you show
the field of AI focused on tasks that are actually difficult for the same image, it gets closer to the correct answer. You do
people to do: playing chess and other strategy games, plan- this again with a picture of an airplane, then a car again, and
ning a path, performing long chains of logical reasoning. an airplane again, and so on. You do this millions of times
These tasks turned out to be relatively easy for computers. with millions of examples. The machine eventually gets the
What is more challenging for computers is tasks that any idea of what the difference is between a car and an airplane.
human (and many animals) can do effortlessly: perceiving The wonderful thing about this process is that it doesn’t work
the world through vision, hearing, and touch; navigating in just for associating images with labels; it works for turning
an unknown environment; and grasping and manipulating speech into words, turning portraits into names, recognizing
objects. These tasks require machines to take in large amounts faces, turning photos into captions, turning text into topics,
of input data and process them in ways that cannot easily be translating text from one language to another. There is a
reduced to programming instructions. wealth of applications.
There is a lot of talk about AI nowadays, but what AI Supervised learning requires a lot of data, however, and
means, in simple terms, is machine learning. Machine learn- the early machine learning systems had to be hand-engi-
ing is what allows a machine to learn some of these more neered. You had to figure out a way to massage the data to
complex tasks, such as recognizing objects in images. It make it digestible by a simple learning algorithm. You did
this by extracting features from the image, or the text, or the
speech or audio signal—whatever it was—that would be
Yann LeCun is vice president and Chief AI Scientist at Facebook and Silver
Professor at New York University, affiliated with the Courant Institute and the characteristic of what you wanted to detect, and then you
Center for Data Science. He was the founding director of Facebook AI fed this information into a simple trainable classifier.
Research and of the NYU Center for Data Science. He received a PhD in Deep learning replaces this handcrafted process with a
computer science from Université P&M Curie (Paris). After a postdoc at the
University of Toronto, he joined AT&T Bell Labs, where he became head of
system built of a cascade of trainable modules. You train the
Image Processing Research in 1996. He joined NYU in 2003 and Facebook modules end to end by showing the system an input and
in 2013. His current interests include AI, machine learning, computer vision, telling it what the answer is. Each of the modules then
mobile robotics, and computational neuroscience. He is a member of the adjusts itself to produce the correct answer. This approach
National Academy of Engineering. [email protected]
works better than the hand-engineering approach of super-
DOI: 10.1080/08956308.2018.1516928
Copyright © 2018, Innovation Research Interchange.
vised learning, and it requires considerably less manual labor.
Published by Taylor & Francis. All rights reserved. You don’t have to preprocess the images in any way; you can

22 | Research-Technology Management • November–December 2018

 learning procedure adjusts the coefficients in all the layers
so that the system automatically learns to detect the right
features or combinations of features.
That is the magic of deep learning. In a convolutional net,
the computation is organized in such a way that you make
it easy for the system to compute functions that correspond
to applications you are interested in, like vision or signal
recognition. This idea of hierarchy and multiple layers comes
from the fact that the natural signals are compositional and
there is an advantage in processing them in a hierarchical
fashion.
The first applications of convolutional networks at Bell
Labs were in character recognition. We realized very early
on, however, that we could use this process not just to rec-
ognize individual objects but to recognize multiple objects in
Yann LeCun is Chief AI Scientist at Facebook and the founding director a single image. This is a very important development if you
of Facebook AI Research. Throughout his career, he has repeatedly
redefined the limits of machine learning.
want to apply the system to real photos, for example, where
the object is never very easily separable from the background.
feed raw data to the system and it learns how to represent Pretty quickly, we realized we could build entire systems to
it and figures out how to crunch it. Deep learning makes it read checks or zip codes. By the mid-1990s, we had built a
economically feasible to apply machine learning techniques system that was deployed by banks; it was reading 10 to 20
to a much wider range of applications. percent of all the checks in the United States by the late
Each module in a deep learning system carries out one 1990s.
operation, either a linear operation (computing weighted Unfortunately, by that time, the machine learning research
sums of input components) or a pointwise nonlinear oper- community had turned away from neural network methods,
ation. If you stack multiple layers of alternated linear and which had a reputation for being finicky—they could be
nonlinear modules, you can build a system that can approx- difficult to make work, and they required very complex soft-
imate any function as closely as you want by changing the ware. We were able to build these systems at Bell Labs
coefficients in the linear modules, as long as you have enough because we had spent years building an infrastructure to
layers that are big enough. train the systems and manipulate them, which nobody else
had done. Even so, the only applications for which we had
Deep Learning with Convolutional Networks enough data were essentially character recognition. There
The question is, how do you structure those operations so were very few other applications for which datasets were
that they actually do something useful? One answer to that available. That changed with the Internet, of course, but it
question is an architecture I developed at Bell Labs in the took about 10 years. In the meantime, the community had
late 1980s, a convolutional network. If you’ve ever talked to lost interest. My colleague and I even stopped working on
your phone or talked to Alexa, or if you’ve let Facebook tag machine learning from the mid 1990s to the early 2000s and
your friends automatically or search for images, or if you’re focused on image compression and other topics.
lucky enough to have a car that has some self-driving capa-
bility, like a lane-keeping system, you’ve engaged with a Applying Convolutional Networks
convolutional network. A series of developments in the early 2000s revived that
It’s called a convolutional network because it uses the interest. In 2002, I left AT&T, went briefly to the NEC
mathematical operation of convolution, but it’s really inspired Research Institute, and then joined NYU, where I started
by biology—in fact, by Hubel and Wiesel’s Nobel Prize–win- working on computer vision and robotics. I began with a
ning work in neuroscience from the early 1960s. The basic small experiment, funded by DARPA, to prove that machine
idea is that each unit in this big neural network is connected learning could be used to train a small, radio-controlled truck
only to local receptive fields in the input or in the previous
layer. Each unit in a layer gets activated when it detects a
particular motif in its receptive field, for example, an edge
at a particular orientation. Units in subsequent layers will
We could use this process not just to
learn to detect combinations of motifs from the previous
layer. For example, the combination of a horizontal and ver- recognize individual objects but to
tical edge can form a corner or a cross. The upper layers of
recognize multiple objects in a single
a convolutional net learn to detect more abstract concepts
such as object parts or object categories. The best aspect of image.
training a deep convolutional net is that you don’t have to
specify what motifs need to be detected at each layer: the

The Power and Limits of Deep Learning November–December 2018 | 23

to drive itself. The idea was very simple: Have someone drive the record on a standard dataset called ImageNet. Typical
a truck loaded with cameras and record the images and the performance on that dataset had been an error rate of 26
steering angle from the human driver; instruct the driver to percent; Geoff Hinton’s team at the University of Toronto
drive straight until the truck gets within two meters of an produced an error rate of 16.4 percent. That was a watershed
obstacle during these training runs, then veer off to one side moment—the computer vision community now realized that
or the other. We trained a convolutional net to associate the convolutional nets worked really well when they were
steering angle with the images, and then loaded that net onto trained with enough data. A similar watershed moment
the small truck, which could now drive itself. occurred in speech recognition a couple years earlier and in
DARPA was impressed enough with our results to fund a natural language processing a few years later.
big program—LAGR, or Learning Applied to Ground Robots. Since then, the architecture of convolutional nets has
The goal was to get robots to drive themselves in nature using continued to develop. There’s been a huge inflation in the
machine learning. My team figured out that we could train number of layers, the size of the layers, and how the layers
a convolutional net to recognize an obstacle, not in the same are optimized. There’s a whole industry of people designing
way we trained the little truck, but by labeling every pixel the details of those architectures so that they work better
in an image as to whether it corresponded to a traversable and faster. The error rate on the ImageNet dataset is now
area or a non-traversable area—an obstacle. down to less than 3 percent. Simultaneously, the number
One nice thing about this approach was that the labels of layers in those convolutional nets has gone up, to about
didn’t have to be produced manually; we could collect them 150 layers or so now. Some of those convolutional nets with
from a stereo imaging system. We would collect images of 50 or 100 layers are used routinely in production at com-
the environment and use stereo reconstruction to tell if a panies like Facebook and Google and Microsoft. The perfor-
point in the environment was sticking out of the ground or mance of computer vision systems today is astounding. A
on the ground. Stereo reconstruction allowed us to label the system developed at FAIR called Detectron can not only
entire image, so the robot could do long-range planning. That locate and recognize objects but also draw a mask of every
analysis provided labels to train the convolutional net. object it picks up.
Building on that project, which ended in 2008, we fig- What’s really interesting about convolutional nets is that
ured out that we could extend this functionality to label they’re applicable to text as well as images. Tons of applica-
not just the traversability of an individual pixel, but to tions are popping up every day. Facebook has been using a
label every pixel in the image with the category it belonged system based on a gated convolutional net—a slightly more
to. In computer vision, this is called semantic segmentation. complicated version of a pure convolutional net—to translate
We built a convolutional net to do this and implemented text from one language to another. Convolutional nets are
a version of it on a field-programmable gate array (FPGA) used to filter trajectories at the Large Hadron Collider in
that could run it at 20 frames per second. We submitted a Geneva. Other applications of convolutional networks are
paper to one of the most prominent computer vision con- expected to appear in transportation and medicine. Those
ferences in 2010—and it was soundly rejected. Most of are the obvious applications, beyond information access,
the reviewers had never heard of convolutional nets. The search ranking, and things like that.
reviewers just could not believe that a completely different
method, one they’d never heard of, could beat records on Reinforcement Learning
a standard dataset. That project eventually got presented There’s another type of learning that’s common in machine
at a machine learning conference in 2011, and it inspired learning, called reinforcement learning. Reinforcement learning
a bunch of people in industry. There was interest in using algorithms are being used to get machines to train themselves
it for self-driving cars, for instance. to play chess or Go or various other games. Some machines
Then, at the end of 2012, our friends at the University of are learning to play complex real-time strategy games, as
Toronto used convolutional nets—very large versions of well. Facebook has twice won an annual competition of bots
them, implemented on graphic processing units (GPUs), playing Doom, and DeepMind and Facebook are both work-
which are very good for fast numerical computation—to beat ing on bots that can play StarCraft. That’s a much more dif-
ficult task than Go or chess.
In reinforcement learning, very powerful machines run
in parallel, playing games against themselves. Whenever one
That was a watershed moment—the of the instances wins, it reinforces that strategy and deem-
phasizes the one that lost. AlphaGo is a system developed in
computer vision community now this way to play the game of Go. In the latest version of
realized that convolutional nets worked AlphaGo, called AlphaGo Zero, DeepMind demonstrated that
it could train a system to reach superhuman performance in
really well when they were trained with just a few hours using this kind of competition. At Facebook,
enough data. we have developed a similar system called ELF OpenGo; it
also reached superhuman performance in a relatively short
time. Unlike AlphaGo and AlphaGo Zero, ELF OpenGo is

24 | Research-Technology Management The Power and Limits of Deep Learning

open source, so that the research community can build on
our work.
Here is the problem with reinforcement learning: To learn The next revolution in AI will be
to play even a very simple game—Space Invaders, for
instance—a machine will have to play hundreds or thou- machines that learn not through
sands of repetitions to reach a decent level of skill. It will supervised learning but through
eventually overtake human performance, but it will require
lots and lots of trials to do so. Humans, on the other hand, observation, a paradigm I call
reach a base level after a few minutes of training. For games, self-supervised learning.
it doesn’t matter, because you can run the games very quickly
on many computers in parallel for as long as you want to.
But if you want to train the machine to drive a car, you don’t
want to have to run off a cliff 50,000 times to figure out how psychologists can test whether babies have internal models of
not to run off a cliff. Most humans can learn to drive a car the world by showing them scenes that violate the dominant
with about 20 hours of training without ever causing an model and seeing how surprised they are.
accident. Animals also have models of the world. A baby orangutan
One of the big challenges in the next few years will be to presented a magic trick—one in which he is shown an object
understand how to get machines to be much more efficient in a cup, and the object is removed without him seeing it—
at reinforcement learning, to learn with fewer trials, like will react when seeing the empty cup. In one video, the baby
humans do. orangutan shown the trick rolls on the floor, laughing.1 His
model of the world is violated by the trick. When your model
The Road to Predictive Learning of the world is broken, you have two possible reactions: you
What are we missing to build truly intelligent machines? laugh, or you’re scared. But in all cases, you pay attention,
We’re missing a lot, clearly; animals and humans learn much because you can learn from it. In that last mode, you adjust
more efficiently than machines do, and the applications of your internal model of the world. Those internal models,
AI today are widely varied but still limited. AI today can give whether they’re developed by babies or by animals, are
us better cars, maybe even autonomous cars, better medical trained in a task-independent way—by observation, with
image analysis, personalized medicine, adequate language only a little bit of interaction. People and animals can then
translation, useful but stupid chatbots, search retrieval fil- use those representations to learn very quickly, modifying
tering, and lots and lots of applications in energy, finance, them when something doesn’t fit the model.
manufacturing, environmental protection, and artistic pro- It’s an incredibly powerful approach. Humans can learn
duction. We cannot have machines with common sense that to drive a car or even fly an airplane with about 20 hours of
you can talk to without being frustrated—your house cat has training—for most of us—without ever crashing—for most
more common sense than the smartest AI systems we have. of us. Machines can’t do this with yet. How do we know that
We cannot have intelligent personal assistants or smart chat- we shouldn’t run off a cliff? Because we have a good physical
bots. We cannot have household robots. We don’t have dex- model that tells us that if we turn the wheel in that direction,
terous robots; it’s still beyond the state of the art to have a we’re going to run off the cliff, and the car’s going to fall, and
machine that can actually empty or fill your dishwasher. it’s not going to be good. Our prediction of what’s likely to
There are really two big questions that we still have to happen allows us to not do stupid things and to plan ahead.
answer on the way to real AI. First, how can machines learn What researchers are trying to do today is build AI systems
as efficiently as humans and animals do? And second, how that have a predictive world model. This is a complex task
can we train machines to plan and act and reason, not just because the world is not entirely predictable. A complete AI
perceive? I think the answer to both of these questions is agent will have to have some kind of world simulator that
learning using predictive models. The next revolution in AI allows it to see what all of the potential effects of a given
will be machines that learn not through supervised learning action plan would be. But every possible consequence may
but through observation, a paradigm I call self-supervised not be easily predictable. For instance, if the entire world of
learning. a machine consists of someone putting a pen on the table
As humans, we build models of the world by observing the and letting it go, and you ask the machine to predict what
world. If you show a baby a little cart on a platform, and you the world is going to look like half a second after the pen is
push the cart off the platform and it doesn’t fall, very young let go, the machine will have a whole bunch of possible
babies—less than 6 months old—will look at this and say, answers based on the directions the pen could fall. There’s a
“Yeah, sure, that’s the way the world works. No problem.” whole set of possible answers that are all correct.
After about 10 months, though, they’ve learned the concept You can visualize the prediction space—the variety of
of gravity, and they have some good notions of intuitive phys- answers that could be correct—as a ribbon. When you train
ics. Presented the same scene, they look very surprised, as if the machine, you don’t want to tell it it’s wrong if it generates
asking themselves, “What’s going on?” They’ve developed an
internal model of this part of the world. Developmental 1
See https://www.youtube.com/watch?v=ka_IlYO3HKM/

The Power and Limits of Deep Learning November–December 2018 | 25

a prediction that could have occurred but actually did not; Montreal. The one in Paris is actually the biggest, with about
you only want to correct the machine if its prediction falls 50 people; New York City and Menlo Park have around 40
outside of the ribbon of possible outcomes. The problem is each, and Montreal about 15. We have recently announced
that you may not have any real idea where this ribbon is. new satellite sites in Tel Aviv, Pittsburgh, Seattle, and London.
The answer may be a second neural net designed to learn Research at FAIR is scientist-driven, bottom-up, open
where the ribbon is—that is, what the range of possibilities research. It’s mostly exploratory, very much like what Bell
is. That’s the idea behind generative adversarial networks, which Labs used to do and like what certain pockets of Microsoft
have gotten a lot of people in the field excited. The idea is Research practice. Most projects involve a few scientists,
that you have two neural nets, a generator and a discrimi- sometimes students, sometimes engineers, and it’s very bot-
nator. The discriminator classifies whether the prediction is tom-up; scientists identify and develop their own projects.
something that actually occurred in the world or whether it We hire research scientists who have a nose for good projects.
is something that has been produced by the generator; it We also have a few larger projects that involve more engi-
learns to discriminate between real input and fake input. neering resources. These are more organized, more managed,
Simultaneously, the generator trains itself to fool the dis- more top-down. We organize like this only when a project
criminator into thinking that what it produces is real. The requires that scale of resources.
generator learns how to cheat—how to change its output so The basic principle of FAIR is open research—all of our
that the discriminator can’t tell the difference—through research is published. Most of our code is released as open
something called gradient competition. source. Results are systematically posted on http://arxiv.org.
That’s adversarial training, and the results with this tech- In classic industry research, the first thing you do with some-
nique have been stunning. We’ve been using adversarial thing new is talk to a patent lawyer and make a preliminary
training to get a machine to predict what’s going to happen patent filing. By contrast, what we do at FAIR is write a paper
in a video by having it reconstruct frames—showing it four and post it on arxiv.org. We have a very fluid publication
frames in a video and asking it to predict the next two. The process. Almost all the papers we produce are submitted to
applications of this are significant. When you’re driving on conferences or journals, and most of the code associated with
the street, for example, you predict what people around you them is released as open source. We don’t tell people, “You
are going to do so you can drive defensively. You know that can publish”; we tell people, “You must publish.”
when pedestrians start crossing the street, they’re going to Why do open research? Why require scientists to publish
keep crossing the street. Autonomous cars will need this and open-source their code? There are several good reasons
same capability. At this point, one of these prediction systems for this. First, it’s good for a scientist’s career and self-image,
can predict up to a couple seconds into the future fairly so it makes it easier to attract good researchers to the lab.
accurately. You don’t get the best people by telling scientists they can’t
talk about what they’re doing; that will basically kill their
Open Research at Facebook career. You’ve got to let people brag about what they do,
Facebook has developed a unique model to advance this because that’s how they build a career, and also because they
research, with FAIR. It’s a model that’s very open and con- see their mission as contributing to the world’s knowledge.
nected with the global research community. That’s a great motivation for research excellence.
I had some view of how a high-profile industrial research Second, when a result is published, it’s more believable,
lab should be organized from my experience at Bell Labs it’s more reliable, and it’s more applicable. When they know
(later AT&T Labs) and NEC, and also a little experience as an they’re going to publish, people go through the exercise of
intern at Xerox PARC. So, when Mark Zuckerberg asked me figuring out the methodology so that their work passes the
to build FAIR, I had some idea about how to organize it. The peer review sieve. It’s also more reproducible and applicable
mission of FAIR is to advance the science of AI and develop internally because you have results that you know have been
the technology that applies it. Nearly five years in, we have obtained correctly.
about 175 scientists and engineers, plus some postdocs and Third, believe it or not, publication sometimes makes it
resident PhD students (who are mostly in Europe). We have easier to convince product groups to develop and deploy
four main sites—Paris, New York City, Menlo Park, and technology derived from our research. You can show the
product group the significance of the work. You can tell
them, “This paper has 100 citations and the code has 8,000
stars on Github. Everybody else in the world is using it. Why
aren’t you using it?” Often, the main problem in technology
In AI, the problem isn’t who is ahead; transfer is convincing people in product groups that the tech-
the problem is that the entire field nology is worth something. That’s the biggest obstacle, actu-
ally. Having published results reduces that barrier.
requires technology that doesn’t exist Fourth, it also makes it easy for other labs to improve on
yet. our results, and that’s great because it means the entire field
progresses faster. In AI, the problem isn’t who is ahead; the
problem is that the entire field requires technology that

26 | Research-Technology Management The Power and Limits of Deep Learning

doesn’t exist yet. We’d like to have intelligent virtual assis-
tants, for instance, but the technology, the science, just does
not exist for this. And so, one motivation behind FAIR’s open- Being first to invent has value, but it
ness is to make the field progress faster. Understanding intel-
ligence and making real progress in AI is probably one of the is being first to deploy that brings
greatest scientific and technological challenges of our times; market value and prestige.
it’s one of the big questions of science. Answering it will take
the efforts of the entire research community. Nobody has a
monopoly on good ideas, not even the biggest companies.
And finally, we need to be flexible to get the scientists we intellectual property, people can move back and forth, and
need. There are a lot of industrial research labs focusing on we can collaborate easily with academic labs without wor-
AI who want to hire the best researchers from academia. rying too much about who owns what.
That’s changing the relationship between academia and And, of course, it’s also very good for the reputation of
industry. Industry cannot just wipe the field clean and hire the company.
everybody from academia. That would be a disaster because People in industry are sometimes surprised by this. Aren’t
there wouldn’t be the next generation of PhD and grad stu- we giving away all of our best secrets? Being first to invent
dents coming to our industry. has value, but it is being first to deploy that brings market
So we need to be flexible; we need to let people have a value and prestige. You don’t necessarily need to be first to
foot in academia and a foot in industry. Scientists are very invent in order to be first to deploy, though it helps a lot.
attracted to industry for a lot of reasons—the chance to work What you need is to be very good at doing technology trans-
on real applications of their research, better compensation, fer very quickly. So that’s what we are working on. You can
engineering support, access to data, a chance to work with take advantage of inventions by locking them up through
a wider community of colleagues. But academia also has a patents, which is very difficult, or you can be smart enough
lot to offer in terms of the different style of research, the to deploy them fast and be first to market, which is our goal.
opportunity to work with students, things like that. So we
have researchers at FAIR who spend 20, 50, or 80 percent Conclusion
of their time at Facebook but are still professors at their home Let me conclude with a few words about the future impact
university, teaching and supervising grad students. That’s my of AI. What has happened very often in the history of sci-
case: I have four PhD students, and I teach one course per ence and technology is that the creation of a technological
year. We have some who split their time—one semester here, artifact precedes the science that goes with it. The lens and
one semester there—and some who are mostly at their uni- the telescope were invented before people figured out the
versity, spending 20 percent of their time at Facebook. We theory of optics, for example. The steam engine was
also work with PhD students who are finding their foothold invented before thermodynamics was understood; thermo-
in the field. There’s a very successful, 30-year-old program dynamics was invented to explain the steam engine, essen-
in France called CIFRE, which allows PhD students to essen- tially, or the limitations of it and other thermal cycles.
tially do their PhD in industry and be co-advised by a pro- Aerodynamics was largely developed after the invention of
fessor and someone in industry. We have 15 of those students the airplane. Computer science, of course, was a spinoff of
in Paris. We’re trying to reproduce it in the United States via building computers and calculators. And information theory
bilateral agreements with various universities some of our grew out of trying to communicate through radio. What
researchers are affiliated with. we are after, really, is the equivalent of thermodynamics
This is all made possible by the open research approach. for intelligence. We are trying to understand the science of
Because Facebook is not particularly possessive about intelligence. That’s my research program.

Share your expertise . . .

Contribute to RTM as an author or peer reviewer. Visit us at http://www.iriweb.org/rtm for announcements
of upcoming special issues, author guidelines, and information about volunteering as a peer reviewer.

The Power and Limits of Deep Learning November–December 2018 | 27