The real-world
potential and limitations of artificial intelligence PART I
Artificial
intelligence has the potential to create trillions of dollars of value across
the economy—if business leaders work to understand what AI can and cannot do.
In this episode of the McKinsey
Podcast, McKinsey Global Institute partner Michael Chui and MGI chairman
and director James Manyika speak with McKinsey Publishing’s David Schwartz
about the cutting edge of artificial intelligence.
Podcast transcript
David Schwartz: Hello, and welcome to
the McKinsey Podcast. I’m David Schwartz with McKinsey Publishing.
Today, we’re going to be journeying to the frontiers of artificial
intelligence. We’ll touch on what AI’s impact could be across multiple
industries and functions. We’ll also explore limitations that, at least for
now, stand in the way.
I’m joined by two
McKinsey leaders who are at the point of the spear, Michael Chui, based in San
Francisco and a partner with the McKinsey Global Institute, and James Manyika, the
chairman of the McKinsey Global Institute and a senior partner in our San
Francisco office. Michael and James, welcome.
James Manyika: Thanks for having us.
Michael Chui: Great to be here.
David Schwartz: Michael, where do we see the
most potential from AI?
Michael Chui: The number-one thing that we
know is just the widespread potential applicability. That said, we’re quite
early in terms of the adoption of these technologies, so there’s a lot of
runway to go. One of the other things that we’ve discovered is that one way to
think about where the potential for AI is, is just follow the money.
If you’re a company
where marketing and sales is what drives the value, that’s actually where AI can create the most value. If you’re a company where operational excellence
matters the most to you, that’s where you can create the most value with AI. If
you’re an insurance company, or if you’re a bank, then risk is really important
to you, and that’s another place where AI can add value. It goes through
everything from managing human capital and analyzing your people’s performance
and recruitment, et cetera, all through the entire business system. We see the
potential for trillions of dollars of value to be created annually across the
entire economy.
David Schwartz: Well, it certainly sounds
like there’s a lot of potential and a lot of value yet to be unleashed. James,
can you come at it from the other direction? What are the big limitations of AI
today? And what do these mean in practical terms for business leaders?
James Manyika: When we think about the limitations of AI, we
have to keep in mind that this is still a very rapidly evolving set of
techniques and technologies, so the science itself and the techniques
themselves are still going through development.
When you think about
the limitations, I would think of them in several ways. There are limitations
that are purely technical. Questions like, can we actually explain what the
algorithm is doing? Can we interpret why it’s making the choices and the
outcomes and predictions that it’s making? Then you’ve also got a set of
practical limitations. Questions like, is the data actually available? Is it
labeled? We’ll get into that in a little bit.
But I’d also add a
third limitation. These are limitations that you might call limitations in use.
These are what lead you to questions around, how transparent are the
algorithms? Is there any bias in the data? Is there any bias in the way the
data was collected?
David Schwartz: Michael, let’s drill down on
a first key limitation, data labeling. Can you describe the challenge and some
possible ways forward?
Michael Chui: One of the things that’s a
little bit new about the current generations of AI is what we call machine learning—in the sense that
we’re not just programming computers, but we’re training them; we’re teaching
them.
The way we train them
is to give them this labeled data. If you’re trying to teach a computer to
recognize an object within an image, or if you’re trying to teach your computer
to recognize an anomaly within a data stream that says a piece of machinery is
about to break down, the way you do that is to have a bunch of labeled data and
say, “Look, in these types of images, the object is present. In these types of
images, the object’s not present. In these types of data streams, the machine’s
about to break, and in these types of data streams, the machine’s not about to
break.”
Play Video
Video
A minute with the McKinsey Global Institute:
What AI can and can’t (yet) do
There have been many exciting
breakthroughs in AI recently—but significant challenges remain. Partner Michael
Chui explains five limitations to AI that must be overcome.
We have this idea that
machines will train themselves. Actually, we’ve generated a huge amount of work
for people to do. Take, for example, self-driving cars. These self-driving cars
have cameras on them, and one of the things that they’re trying to do is
collect a bunch of data by driving around.
It turns out, there is
an army of people who are taking the video inputs from this data and then just
tracing out where the other cars are—where the lane markers are as well. So,
the funny thing is, we talk about these AI systems automating what people do.
In fact, it’s generating a whole bunch of manual labor for people to do.
James Manyika: I know this large public museum
where they get students to literally label pieces of art—that’s a cat, that’s a
dog, that’s a tree, that’s a shadow. They just label these different pieces of
art so that algorithms can then better understand them and be able to make
predictions.
In older versions of
this, people were identifying cats and dogs. There have been teams, for
example, in the UK that were going to identify different breeds of dogs for the
purposes of labeling data images for dogs so that when algorithms use that
data, they know what it is. The same thing is happening in a lot of medical
applications, where people have been labeling different kinds of tumors, for
example, so that when machines read those images, they can better understand
what’s a tumor and what kind of tumor is it. But it has taken people to label
those different tumors for that to then be useful for the machines.
Michael Chui: A medical diagnosis is the
perfect example. So, for this idea of having a system that looks at X-rays and
decides whether or not people have pneumonia, you need the data to tell whether
or not this X-ray was associated with somebody who had pneumonia or didn’t have
pneumonia. Collecting that data is an incredibly important thing, but labeling
it is absolutely necessary.
David Schwartz: Let’s talk about ways to
possibly solve it. I know that there are two techniques in supervised learning
that we’re hearing a lot about. One is reinforcement learning, and the other is
GANs [generative adversarial networks]. Could you speak about those?
Companies and
organizations that are taking AI seriously are playing these multiyear games to
acquire the data that they need.
Michael Chui: A number of these techniques
are meant to basically create more examples that allow you to teach the
machine, or have it learn.
Reinforcement learning
has been used to train robots, in the sense that if the robot does the behavior
that you want it to, you reward the robot for doing it. If it does a behavior
you don’t want it to do, you give it negative reinforcement. In that case, what
you have is a function that says whether you did something good or bad. Rather
than having a huge set of labeled data, you just have a function that says you
did good or you did the wrong thing. That’s one way to get around label data—by
having a function that tells you whether you did the right thing.
With GANs, which stands
for generative adversarial networks, you basically have two networks, one
that’s trying to generate the right thing; the other one is trying to
discriminate whether you’re generating the right thing. Again, it’s another way
to get around one potential limitation of having huge amounts of label data in
the sense that you have two systems that are competing against each other in an
adversarial way. It’s been used for doing all kinds of things. The
generative—the “G” part of it—is what’s remarkable. You can generate art in the
style of another artist. You can generate architecture in the style of other
things that you’ve observed. You can generate designs that look like other things
that you might have observed before.
James Manyika: The one thing I would add
about GANs is that, in many respects, they’re a form of semisupervised learning
techniques in the sense that they typically start with some initial labeling
but then, in a generative way, build on it—in this adversarial, kind of a
contest way.
There’s also a whole
host of other techniques that people are experimenting with. One of the things,
for example, is researchers at Microsoft Research Lab have been working on
instream labeling, where you’ll actually label the data through use. You’re
trying to interpret based on how the data’s being used, what it actually means.
This idea of instream labeling has been around for quite a while, but in recent
years, it has started to demonstrate some quite remarkable results. This
problem of labeling is one we’re going to be with for quite a while.
David Schwartz: What about limitations when
there is not enough data?
Michael Chui: One of the things that we’ve
heard from Andrew Ng, who’s one of the
leaders in machine learning and AI, is that companies and organizations that
are taking AI seriously are playing these multiyear games to acquire the data that they need.
In the physical world,
whether you’re doing self-driving cars or drones, it takes time to go out and
drive a whole bunch of streets or fly a whole bunch of things. To try to
improve the speed at which you can learn some of those things, one of the
things you can do is simulate environments. By creating these virtual
environments—basically within a data center, basically within a computer—you
can run a whole bunch more trials and learn a whole bunch more things through
simulation. So, when you actually end up in the physical world, you’ve come to the
physical world with your AI already having learned a bunch of things in
simulation.
That’s the holy-grail
question: How do you build generalizable systems that can learn anything?
James Manyika: A good example of that is
some of the demonstrations, for example, that the team at DeepMind Technologies
has done. They’ve done a lot of simulated training for robotic arms, where much
of the manipulation techniques that these robotic arms have been able to
develop and learn was from having actually been done in simulation—way before
the robot arm was even applied to the real world. When it shows up in the real
world, it comes with these prelearned data sets that have come out of
simulation as a way to get around the limitations of data.
David Schwartz: It sounds like we may be
considering a deeper issue—what machine intelligence actually means. How can we
move from a process of rote inputs and set outputs to something more along the
lines of the ways that humans learn?
James Manyika: That’s, in some ways, the
holy-grail question, which is: How do you build generalizable systems that can
learn anything? Humans are remarkable in the sense that we can take things
we’ve learned over here and apply them to totally different problems that we
may be seeing for the first time. This has led to one big area of research
that’s typically referred to as transfer learning, the idea of, how do you take
models or learnings or insights from one arena and apply them to another? While
we’re making progress in transfer learning, it’s actually one of the harder
problems to solve. And there, you’re finding new techniques.
This idea of simulating
learning where you generate data sets and simulations is one way to do that.
AlphaGo Zero, which is a more interesting version, if you like, of AlphaGo, has
learned to play three different games but has just a generalized structure of
games. Through that, it’s been able to learn chess and Go—by having a
generalized structure. But even that is limited in the sense that it’s still
limited to games that take a certain form.
Michael Chui: In the AI field, what we’re
relearning, which neurologists have known for a long time, is that as people,
we don’t come as tabula rasa. We actually have a number of structures in our
brain that are optimized for certain things, whether it’s understanding
language or behavior, physical behavior, et cetera. People like Geoff Hinton
are using capsules and other types of concepts. This idea of embedding some
learning in the structure of the systems that we’re using is something that
we’ve seen as well. And so, you wonder whether for transfer learning, part of
the solution is understanding that we don’t start from nothing. We start from
systems that have some configuration already, and that helps us be able to take
certain learnings from one place to another because, actually, we’re set up to
do that.
James Manyika: In fact, Steve Wozniak has
come out with certain suggestions, and this has led to all kinds of questions
about what’s the right Turing test or the kind of test you can come up with
generalized learning. One version that he has is the so-called “coffee test,”
which is, the day we can get a system that could walk into an unknown American
household and make a cup of coffee. That’s pretty remarkable, because that requires
being able to interpret a totally unknown environment, being able to discover
things in a totally unknown place, and being able to make something with
unknown equipment in a particular household.
There are a lot of
general problems that need to be solved along the way of making a cup of coffee
in an unknown household, which may sound trivial compared to solving very
narrow, highly technical, specific problems which we think of as remarkable.
The more we can then look to solving what are generalized often as, quite
frankly, garden-variety, real-world problems, those might actually be the true
tests of whether we have generalized systems or not.
And it is important to
remember, by the way, as we think about all the exciting stuff that’s going on
in AI and machine learning, that the vast majority—whether it’s the techniques
or even the applications—are mostly solving very specific things. They’re
solving natural-language processing; they’re solving image recognition; they’re
doing very, very specific things. There’s a huge flourishing of that, whereas
the work going toward solving the more generalized problems, while it’s making
progress, is proceeding much, much more slowly. We shouldn’t confuse the
progress we’re making on these more narrow, specific problem sets to mean,
therefore, we have created a generalized system.
There’s another
limitation, which we should probably discuss, David—and it’s an important one
for lots of reasons. This is the question of “explainability.” Essentially,
neural networks, by their structure, are such that it’s very hard to pinpoint
why a particular outcome is what it is and where exactly in the structure of it
something led to a particular outcome.
CONTINUES IN PART II
No comments:
Post a Comment