OK.
It's amazing to be here.
I'm Josh Tenenbaum.
And I'm going to be talking about scaling AI the human way.
I am a cognitive scientist.
I'm also an AI engineer.
And both sides of our work are motivated by this question,
why do we have this boom in AI technologies, but no real AI?
Increasingly, we have amazing machines
that do things we used to think only humans could do.
But none of them are truly intelligent.
None of them really have common sense.
Each of them has to be built by really smart engineers,
and each does just one thing.
So I want to know, what's missing?
What would it take to build a machine
with the kind of flexible, general purpose intelligence
of a human that each of you has to learn for yourselves?
Each of these things, and infinitely more.
So it won't be a surprise now see here
about what I think, is that the answer comes from what you just
heard from my colleagues, Laura Schultz, Rebecca
Sachs, and the whole fields they represent, much broader fields.
Imagine if we could build a machine that
grows into intelligence the way a human being does,
that starts like a baby and learns like a child.
Now that would be AI that is really intelligent.
If we could build it, do you think we could?
Probably not in 10 years.
Almost certainly not.
Probably not even in 20.
Quite possibly, not even in our lifetimes.
But that's OK.
That's why we call it a quest.
And like any good quest, it's not so important
when or even whether you get to the end,
it's what you discover along the way.
And one thing we've discovered along the way
is that even small steps towards this goal can be big.
So let me take you back one more time to the history
of deep learning.
Consider the math, the mathematics
of deep learning and reinforcement learning
was introduced in these papers, published decades ago
in the 1960s, 70s, and 80s.
They were published mostly in journals
of psychology or cognitive science
as models of human learning.
Each of these papers at the time was a small step,
a very simple model.
But it turns out that these small steps when
engineered at scale, that's become today's
AI that's transforming our world.
So think what we could do with just a few small steps
towards this goal.
Now, as you've already heard, the idea
of building AI machines that learn like a child
is perhaps the oldest idea in AI.
Again, let's go back in history.
Alan Turing introduced this idea in the same paper in 1950
when he introduced the Turing test.
It was championed by Marvin Minsky when
he started the AI lab at MIT.
And Patrick Winston, longtime director of the AI lab,
and many others.
For good reason, because it's the only known scaling path
in the universe.
Human children are the only system
we know that demonstrably, reliably, and reproducibly
grows into human level intelligence.
So why not build AI this way?
Well, why haven't we yet?
Well, again, as I think you've already heard, in some sense,
I think it's only now that the field of the science
of human learning is mature enough that it can offer
some useful guidance to AI.
And maybe AI is mature enough to know what to do with it.
And I think it could be extremely valuable.
So if we go back to Turing here, right, Turing
could only presume what a child's brain was like,
and he presumed wrong.
He said, maybe it's like a notebook
that you buy from the stationer.
Rather little mechanism, lots of blank sheets.
But what you've just seen is that babies start out
with so much more.
And children's learning is so much more than just copying
down what a teacher writes on the blackboard.
So my part in this quest is trying
to capture these insights in engineering.
And we start with the most basic common sense
that's in every young child, but no AI.
So the intuitive physics for example,
when a kid is playing with blocks
or stacking cups like this, imagine if we
could build a robot to do that.
Or the intuitive psychology, when someone that
lets young kids--
like the one in the back corner here
in this classic experiment-- figure out
what someone is doing and why, what their goal is, even
when performing an action they've never seen.
These kids in these videos are just 1 and 1/2 years old,
but watch what this kid does here,
and imagine if we could build a robot
with this kind of common sense, this intelligence, this skill
and helpfulness that could help you out around the house.
Yeah, that would be amazing.
OK?
So in my group we're trying to develop the tools that could
someday make this possible.
And I'll tell you a little bit about them.
One is that we use new AI technologies called
probabilistic programs that our group helped to invent, along
with many others in AI.
These combined deep learning's ability
to see patterns and data with other great AI
ideas that don't fit yet in today's neural networks.
So symbols for representing real knowledge, not just patterns.
Causal reasoning, not just statistical inference
or correlation.
And hierarchical inference for learning to learn,
so knowledge builds progressively, and learning
makes you smarter at learning new things.
Some of the abilities-- just some
of the abilities that Laura showed us
make kids the smartest learners in the universe.
Or if you want to capture the common sense
that even the youngest babies start out with,
we use another kind of program that you might not
have thought of as AI.
So these are video game engines, programs
for writing video games.
Physics engines to simulate a player's real time interaction
with the 3D world of objects, or graphics engines
to render the viewpoint as a player
moves around, or planning engines
to model the non-player characters at the game.
These engines we think may be something
like the software description of the high level brain
architectures that Rebecca finds is built into babies' brains,
where kids start in learning.
So with these tools in our lab, and with many collaborators,
we've built something that we call for instance,
an intuitive physics engine to explain how your brain can
answer an endless range of questions about a physical
scene-- like these blocks here--
without any special training.
Which way will the blocks fall?
Rather, it's let's start will the blocks fall?
Which ones of these scenes are more
likely to fall than others?
How confident are you?
Our model can predict that.
If they do fall, which way will they fall?
How far will they fall?
What if I tell you that the gray stuff is 10 times heavier
than the green stuff?
Or what about in these scenes, is red or yellow heavier,
given that these towers appear surprisingly stable?
Our system can answer--
like your brain-- questions you've
never thought about before, for which you
don't have the data needed for pattern recognition.
But you do have a causal model.
In other work-- a lot of this started in collaboration with
Rebecca--
we've built what we call intuitive psychology engine,
or a model of theory of mind.
This is a model of how agents plan actions
based on their beliefs and desires
to maximize expected utility.
It's what Laura called the naive utility calculus
that together we have studied in young children.
Here, goals give rewards, while physics determines costs.
And this model, we think in your brain can let
you observe someone's actions--
like this woman here--
and compute her goals.
When can you tell what she's reaching for?
I think maybe just about now.
That's what our model predicts also.
The same system can explain how you
might compute in this scene in the lower left
that one person is helping another person.
While the next scene here is not helping,
but quite the opposite, what we call hindering.
And recently, we're excited that the same kind of systems
give some of the first quantitatively testable
models of young children's common sense.
Sorry, the youngest babies.
So we test infants intuitive physics 12-month-old in scenes
like this.
Objects are bouncing around, and when
an object appears that might be a little bit
surprising like this one, infants will look longer at it.
And our model can predict quantitatively
how long infants look at an object based
on how probable it is under our models computation.
Or we can study here the intuitive psychology
of even younger babies with cartoons like this.
The more work a character is willing to do,
the more cost they're willing to pay to achieve a goal.
Jumping over a wall, which they didn't do before,
but which they're about to do here,
that tells you how much an agent wants something.
And it's not just jumping over a wall,
it's rolling up or down a hill, leaping across a gap.
The more you the more you see an agent work,
the more you think it wants that goal.
Our model predicts that quantitatively,
and it turns out infants do too.
So we can capture some of what babies
know what 10 months or 12 months, a little bit of how
they use their intuitive physics or their intuitive psychology
to learn about new objects or agents.
But how does a baby build these engines in the first place?
That's the real learning problem.
Evolution might have given us something like a proto game
engine in your head, but learning for a baby
then means learning to program those engines to describe
the game of your life.
Learning algorithms really have to be programming algorithms.
We need program learning programs.
So how are we going to do that?
Well, we like to call this the hard problem of learning,
because it's really, really hard.
It's much harder than learning in neural networks,
where you have a smooth error landscape
and you're just following the gradient
to roll downhill until you hit bottom, the point in weight
space with the lowest error.
The space of programs has no such nice simple geometry.
And you have a really hard search problem,
but somehow children solve it.
So to try to understand how they solve it and build machines
that might learn programs, we've been
working on an easier warm up problem of how people learn
the programs that capture simple visual concepts,
such as a new tool or a new handwritten character.
So you can learn the concept of a cam--
it's a piece of rock climbing equipment
down on the lower left there--
and pick out the other cams, like this.
Likewise, you can learn any of thousands
of new handwritten characters, even in alphabet's you've
never seen before.
You don't need hundreds or thousands
of examples like a deep learning system, you just need one.
Now, we can capture this ability in machines with something
called Bayesian program learning,
another kind of problematic program
where you have a mental model of the causal processes that
produce ink on the page, like the motor programs or action.
It's a probabilistic program that
captures the different ways characters can be drawn.
And you can invert it using a kind of Bayes' rule
to work backwards to figure out what action sequence is most
likely to have produced it.
That's your concept, and you can use
it to imagine other instances of the same character.
To see it in action here with a little kind of Turing test,
we asked both our machine and people, given just one example,
to imagine other instances of the same character.
Can you tell which is the human or the machine?
Anybody think they can get them here
in each of these six cases?
Here's the right answer.
Did anybody get them all right?
Yeah, probably most of you got about three right, I think.
Basically, people can't tell.
We've passed a very simple Turing test
in concept learning, in very simple concepts learned.
Pretty cool.
We like it.
But it's just of course, a small step.
Could this scale up to learn all the knowledge that a child
acquires over her lifetime?
Well, we've recently made another small step towards this
goal with a new form of Bayesian program learning inspired
by some of the science showing that some of the deepest
learning underlying children's everyday activities--
like these--
might not happen during the day, but during the night
while you're asleep.
So inspired by consolidation and replay during human sleep,
our dream coder algorithm learns to learn new programs
to get better and faster each day
by abstracting out the most central concepts from what
it learned while it was awake, and then
testing out and practicing these in its dreams
to get better the next day.
This system might be able to discover
on its own whole new libraries of concepts,
each like its own programming language.
Will this be the answer to AI's oldest dream?
A machine that truly learns like a child?
Probably not.
Almost certainly not, but it might be the next small step.
Could it be the next form of deep learning, deeper learning?
Might be.
Stay tuned.
We're just at the beginning.
But we're at the end of this first session kicking off
the Intelligence Quest.
And I just want to close by coming back
to the moonshot slide that Jim showed you at the start
and leave you with a few thoughts.
First, what I told you about here
is just a little about just one of these projects.
We're just scratching the surface.
And you can see that the work we do, even
in this one project, touches on all these big questions.
And it touches deeply on both science and engineering.
That's just in the nature of intelligence,
and why we're so excited to be kicking off
this big initiative that brings together so many
of the students and faculty across MIT
and what they're excited about.
But most of all, notice on this slide
there's so much that isn't on there yet.
The dot dot dots on the bottom are the most important part,
even if they're the smallest.
And we're just at the beginning of each of these questions.
So if you feel like these questions might
be your questions, or even better,
if you know where the dots are going,
I hope you'll join us on this quest.
Please, join us.
[APPLAUSE]
Không có nhận xét nào:
Đăng nhận xét