23. Model Merging; Cross-Modal Coupling; Course Summary Video Lecture -

PATRICK WINSTON: Well,
don't stop.
Shoot.
I guess we've got to stop.
I will soon go into withdrawal
symptoms that will
last about six weeks.
But, on the other hand, we all
are beginning to develop a
sort of tired and
desperate look.
And perhaps it's a good thing to
get the semester behind us
and go into solstice
hibernation.
Anyway, there's a
lot to do today.
I want to wrap up a couple
things, talk about what's
next, and maybe get into some
big issues, perhaps a little
Genesis demo, that
sort of thing.
So here's what we're
going to do first.
Last time, I talked about
this whole idea
of structure discovery.
And really, the whole reason I
cracked and started talking
about basic methods is because
of the potential utility of
taking that idea one step
further and finding structure
in situations where you might
not otherwise find it.
It's still an open question
about whether that's best way
to think about it.
But here it goes.
Imagine you've got a
couple of stories.
And these circles represent
the events in the story.
And now what you'd like to get
out of these stories is some
kind of finite state graph that
describes the collection
of stories.
So you might discover, for
example, that these two events
are quite similar.
And these two events
are quite similar.
So you might use that as a basis
for speculating that
maybe a more compact way of
representing the stuff in the
story would look like this.
Where this one--
let's see.
This one goes with this one,
this one goes with this one
and there's a possibility of
another state in between.
So that's the notion of Bayesian
story merging.
Now I'd like to show you a
little bit more convincing
demonstration of that.
Here's how it goes.
So there are the two stories.
This is just a classroom
demonstration, no big deal.
But you can see there's
a sort of parallel
structure in there.
So this is the work of a
graduate student, Mark
Finlayson, who processed those
stories to produce those kinds
of events and those kinds of
events that get assembled into
two story graphs.
And the question is, is the
most probable way of
explaining that corpus
of stories?
And of course, the
answer is no.
If you merge some things like
chase and stop, then you get a
simpler graph, one that is
more probable in the same
sense that we discussed
last time.
Then you can merge run and flee
because they're similar
kinds of events.
And finally, you've got
think and decide.
Boom.
There is your story graph.
And this is the same idea taken
several levels higher
that produce the capacity to
discover, in these two
stories, the concept of revenge,
as promised at the
beginning of our last
discussion.
So sometimes the Bayesian stuff
is the right thing to
do, especially if you
don't know anything.
But sometimes you
do know stuff.
And when you do know stuff it's
possible that you can do
something very much
more efficient.
This sort of thing
takes clouds of
computers to process.
But we learned a lot of stuff
in the course of our
development that we don't
use a cloud of
computers to figure out.
We learned how to associate the
gestures of our mouth with
the sounds that we make,
for example.
So I want to spend a minute or
two talking about some work
that someday will be the
subject of a couple of
lectures, I think.
But it's the question of how to
use multiple modalities and
correspondences between them
to sort out both of the
contributing modalities.
That sounds contradictory.
Let me show you an example.
This is the example,
a zebra finch.
And it's showing you the result
of a program written by
Michael Coen, now a
professor at the
University of Wisconsin.
So the male zebra finch learns
to sing a nice mating song
from its daddy.
And this is what one such
zebra finch sounds like.
[BIRD SONG PLAYING]
Nice, don't you think?
And here's what was learned
by a program that uses no
probabilistic stuff at all,
but rather the notion of
cross-modal coupling.
[BIRD SONG PLAYING]
Can you tell the difference?
It's not known if this
particular song turns on the
female zebra finch.
But to the untrained human
ear, they sure
sound a whole lot alike.
So how does that work?
Here's how that works.
Well, I'm not going to show
you how that works.
What I'm going to show you is
how the classroom example
works, the first chapter
example in
Coen's Ph.D. Thesis.
Here's what happens.
When we talk, we produce a
Fourier transform that moves
along with our speech.
And if we say a vowel like aah,
you get a fairly constant
Fourier spectrum.
And you can say, well, where
are the peaks in
that Fourier spectrum?
And how do they correspond to
the appearance of my mouth
when I say the vowel?
So here's how that works.
So here's the Fourier spectrum
of a particular vowel.
And when you smooth that, those
peaks are called formants.
And so we're just going to keep
track of the first and
second formant.
But when I say those things,
I also can form an ellipse
around my mouth when
I say them.
And when I form an ellipse
around my mouth when I say
them, that gives me this
second modality.
So the question is, is there
a way of associating the
gestures that produce the sound
with the sound itself?
Well, there's the human data
conveniently provided by a
variety of sources, including
Michael Coen's wife who
produced the lip contour
data on the right.
So that's all marked up and
color coded according to the
particular vowels in English.
I guess there are ten of them.
So we humans all learn that.
But guess what?
We don't learn it from this.
Because we don't get to work
with any marked up data.
We learn it from that.
Somehow we're exposed to the
natural world, and we dig the
vowel sounds out.
It's fantastic how we do that.
But we do have cross modal
coupling data and maybe that's
got something to do with it.
So here is a particular
cluster of sounds.
And what I want to know is, can
I merge any of these two
clusters to form a bigger
cluster with a
corresponding meaning?
So what I can do is, I can say,
well, I can watch these.
I know what the lip
form is when a
particular sound is made.
So I have these correspondences.
So maybe there are four
of those, like so.
And maybe this same guy
projects a couple
of tones into that.
And the question is, can this
guy be combined with any of
these guys?
And the answer is yes.
If they're close together on one
side, maybe that suggests
you ought to cluster them
on the other side.
But there's a question about
what close means.
So let's suppose that we also
look at how these guys, these
other two guys, project.
And suppose this guy projects
twice up here
and once over here.
And this guy down here
just projects like
crazy into that guy.
Which are closer?
These two or these two?
Well, my diagram's getting
a little closer.
But if you paid attention when
I was drawing it, you would
see that this guy projects in
proportion to that guy.
So if we look at the-- if we
take each of these projections
as the components of a vector,
then those two vectors are in
the same direction.
And the cosine between
them is zero.
So these are the two that are
closest together from that
sort of perspective of
that kind of metric.
And those guys are the ones
who get combined.
Would you like to see
a demonstration?
Yeah.
OK, here's a demonstration
based on Coen's work.
So here we have two sides.
We could think of one side as
being vowel sounds and the
other side as being lip
contours or something.
But you don't see anything in
the diagram so far about how
these things ought to be
sorted out into groups.
So if I just take one step, why,
it discovers that those
two guys had the
same projection
pattern as each other.
So if I take another step and do
the same thing on the other
side, and now in the third step,
the two areas that were
formerly combined now
form a super area.
And they're seen to project in
the same way as the blue area.
So using this kind of projection
idea, I can
gradually build up an
understanding of how those
regions go together.
And I discover, in this
contrived example, that
there's a vertical arrangement
on the right side that
corresponds to a horizontal
arrangement on the left side.
Now, you say to me, I'd like to
see something a little bit
more like the lip
contour data.
I'm just stepping through
here until I get
something I kind of like.
Oh, that sounds good.
That seems good.
So there's a correspondence
here.
This is all made up data,
Gaussians of various shapes
and orientations.
Let's see what happens when
I run the clustering
algorithm on that.
Something definite was learned
at every step.
We find the correspondence
between the pink region on the
right and the pink region
on the left.
In some cases, where the regions
are rather blurred
together, the other side is the
one that helps the system
figure out how things
are organized.
So I cite this as an example
of something I
think is very important.
Number one, it's possible to
discover regularity without
being obsessively concerned
with Bayesian probability.
And also, that there's very
likely a whole lot of this
going on in human
intelligence.
When we emerge and begin to
examine and explore the world
around us, we're presented with
a lot of unlabeled data
that we've got to
make sense of.
And I believe that this kind of
cross modal coupling idea
is very likely to be bound up
in our understanding of that
world that's presented to us.
It's fast, it's direct.
It doesn't take thousands
of data points.
It just happens.
And it happens effortlessly.
And if this isn't built in--
if this isn't determined to be
built in, you can come back to
MIT in 15 years and
put me and jail.
Because I think this is really
the way it works.
So there it is.
There's a couple of things
to have wrapped up.
And now the next thing I want to
do for the rest of our last
time together in this format
is talk to you about
a variety of things.
And I'll depart from my
usual practice and
move to some slides.
So Dave, could we have the
center screen, please?
So first, a brief review of
where we've been and where
we've come.
I think in the very first class,
I talked about what
artificial intelligence was.
And I talked about how you could
view it from either an
engineering perspective or
a scientific perspective.
I'm on the scientific
perspective side.
And I think, nothing against
applications, but I think
we'll be able to make much
more sophisticated and
wonderful applications if we
have not only the engineering
perspective about building stuff
but also the scientific
perspective about understanding
the stuff to
begin with.
So both perspectives
are important.
And in this case you can see
that they all involve
representations, methods,
and architectures.
Dave, I've changed my mind.
Could you also give me
the side screen so
I can see it too?
So, that's that.
What's next?
The business perspective,
which we talked about on
Thanksgiving.
The important idea being that
the knee-jerk expectation that
the commercial value of
something is in replacing
people is something that is
not sensible in the first
instance, and demonstrated to be
unlikely and untrue in the
second instance.
The thing that turns people on
from the point of view of
applications is not replacing
people but making new revenue,
making new capability.
And that at once licenses you
to not have something done
exclusively by a computer but
something that can be done in
partnership with a person.
So all the important
applications of artificial
intelligence involve people
and computers working in
tandem, with each doing
what they do best--
not with replacing people.
So that's that.
Here's what AI does that makes
AI different from the rest of
the fields that attempt to
contribute to an understanding
of intelligence.
Now, we have the benefit of
having a language for
procedures.
We have all the metaphors that
we are the custodians of in
consequences of knowing
about programming.
We have the metaphor of
garbage collection.
We can talk about all sorts
of things with programming
metaphors that are unavailable
to people in other fields that
are interested in psychology.
We have a way to make models
because we can write programs.
And when we write programs,
there's no question of
sweeping things under the rug.
We have to work out
the details.
And once we've done all that,
then we have opportunities to
experiment that are beyond the
ability to experiment in most
other fields.
Oh, magnificent experiments
are done these days in
developmental psychology
and all the rest of--
all the other branches of
psychology, including MRI
studies that probe into your
brain and see how it's
consuming sugar.
But it's very difficult to
ablate or take away some piece
of your knowledge and see
how you work without it.
I can take the line-drawing
program that we talked about
and say, how will it do if it
doesn't know anything about
fork junctions?
And we can determine
an answer.
But I can't reach into
Sebastian's head here with a
surgical procedure
and take out his
knowledge of fork junctions.
It just can't be done.
And finally, another reason why
we're different is because
we can put upper bounds on how
much knowledge is needed in
order to perform a certain
kind of task.
These days, with bulldozer
computing, the question most
often asked is, how can you get
billions of the things off
the web and use them.
We--
I especially-- sometimes ask the
opposite question, which
is how little knowledge
can you have and still
understand a story?
That's what's interesting
to me.
So there's a methodological
slide that talks to the
question of how you do
artificial intelligence in
particular and, I suppose,
science in general,
engineering in general.
There's a great tendency in
this field to fall in love
with particular methods.
And we've had people who've
devoted entire careers to
neural nets, genetic algorithms,
Bayesian
probability.
And that's mechanism envy.
And a better way, in my
judgment, is to say, what is
the problem?
Scientific method-- what's
the problem?
And then bring the right
machinery to bear on the
problem, rather than looking
for things to do with a
particular kind of machinery.
So this is the methodology that
first articulated in a
forceful way, by David Marr.
You want to start with the
competence you're trying to
understand, then bring a
representation to bear on it,
a representation that exposes
the constraints and
regularities.
Because without those, you
can't make models.
And without those models you
can't understand it, explain
it, predict it, or control it.
So it seems to make sense from
a kind of MIT, model-centered
point of view.
And only when you've got all
that straight do you start
working on your methods and
implement an experiment and
then go around that loop.
So that's all I want to say by
way of review, I suppose.
I want to take a minute or two
and just remind you of what's
on the final.
And there's nothing to remind
because you all know what's on
the final already.
We'll have four sections
corresponding with four exams.
Then we'll have a fifth and
final question that will be
everything else.
All that stuff you slept through
will be featured
there, as well a little problem
on Bayesian inference.
We rearranged the subject,
mostly so I could write the
demonstrations.
So that the Bayesian
stuff didn't come
before the fourth quiz.
Therefore the Bayesian stuff
that you see on those previous
quizzes is likely to be harder
than the stuff that we'll ask
on the final, because you
haven't had as much experience
with it as people
did last year.
I've got a few icons on
there to remind me to
tell you a few things.
As always, open everything
except for computers.
You can wear a costume.
You can do anything you like as
long as it doesn't disturb
your neighbor, within reason.
Well, I guess if it doesn't
disturb neighbor,
it is within reason.
So maybe that's all
I need to say.
I'm not sure where we're
going to be.
But it's certainly the case
that, historically, there are
no visible clocks.
So, we soon run out of all of
our cellphones, wrist watches,
and other time pieces
as we hand them out.
So it pays to remember to bring
some kind of timepiece,
because we won't be able to
convey the time very well.
And finally, I see a little
calculator there.
I don't recall any exam where
you actually needed a
calculator.
But it's sort of a security
blanket to have one.
People sometimes see a problem
and say, oh my God, what am I
going to do?
I left my calculator at home.
So as to avoid that anxiety,
you might want to bring one
even though you won't need it.
So that's the final.
I'm sure there are
no questions.
Are there?
It's obvious.
Everybody will do well.
Two shots, that whole thing.
Now, what to do next?
Suppose this subject
has turned you on.
There are a variety of things
that you should be thinking
about doing next semester.
And I wanted to review
just a few of those.
One of them is Marvin Minsky's
subject, Society of Mind.
It's very different
from this class.
There are no prepared
lectures.
Marvin doesn't rehearse.
He doesn't think about what he's
going to say in advance.
It's like this except
it's just a
conversation with Marvin.
So many people find themselves
bored stiff for two lectures
out of three.
But then in the third lecture,
Marvin will say something that
you'll think about for a year or
for the rest of your life.
That's what happens to me.
I'm bored stiff two
out three times.
And then the third lecture he
says something, and I think
about it for at least a year
and maybe permanently.
So it's an opportunity to see
one of MIT's true geniuses
think out loud.
So it's an experience that you
don't want to miss because
that's what you come here
for, is to see the
geniuses think out loud.
Speaking of geniuses, then
there's Bob Berwick.
And he heroically is doing two
subjects in the spring.
Both of which I'd
take if I could.
One is his subject on Language
Understanding.
And the reason I'd take that
is because I believe that
language is at the center of
any explanation of our
intelligence.
So that's the subject I would
be, I suppose, most inclined
to take if I were you.
Well, maybe Minsky's.
It's hard to say.
And incidentally, very
heroically, Bob is also
teaching a course on how
evolution works,
how it really works--
in so far as we know how
it really works--
as well.
So both of those will be
offered in the spring.
I don't know how he does it.
I don't know how he does two
all at the same time.
Of course there are lots of
places where you can go to
school, and the faculty will
be teaching five courses at
the same time.
I just think they're
crazy or something.
I don't know how that works.
Gerry Sussman will be teaching
his Large Scale Symbolic
System subject.
That's sometimes--
oh, I forgot what he wasn't
able to call it, something
that wasn't politically correct
about programming for
people who really, really
like to program.
It's a splendid course on how
to build really big systems.
And we use the ideas in that
subject in our research
system, because it's
the only way--
understanding how that works is
the only way that you can
build systems that are
too big to be built.
I may say a word about
that a little later.
So those are my favorite
three/four picks.
But there's lots of
other stuff, too
many things to cover--
the media lab, [INAUDIBLE]
psychology.
There's tons of stuff
out there.
And I would only mention that
courses that have those three
names on them are bound
to be good.
These are colleagues that I
think I have a important
perspective--
not necessarily one I agree
with, but an important
perspective that you
should understand--
Richards, Tenenbaum,
and Sinha.
And now we come to my spring
course, the Human Intelligence
Enterprise.
It's 6.XXX not because there's
anything pornographic about it
but because for a long time, I
couldn't remember it's number.
So I developed a habit of
referring to it as 6.XXX and
it seems to have stuck.
Here's what that's about.
Yeah, that might
be interesting.
It's taught like a humanities
course, though.
No lectures, I just talk.
And all the TAs are veterans of
that class so if you want
to know if you should do it,
you have several resources.
You can talk to them.
Or you could look at the sorts
of things we talk about.
Here are some things we talk
about by way of packaging.
Yeah, I can hear a little
tittering there because people
have discovered the
last element.
Some people take the whole
subject because they want to
be present for that unit.
And we talk about all those
kinds of things.
And we look to see what the
common elements are in all
those kinds of packaging
problems of the sort that you
will face over and over again
when you become an adult, no
matter what you do.
If you become a business person,
an entrepreneur, a
military officer, a scientist,
or an engineer, that packaging
stuff will often make the
difference between whether you
succeed or don't.
And then, that's the second way
you can figure out whether
you want to take the subject.
The content, the TAs, and then
of course you can always
appeal to the Underground
Guide.
And that's why it's very rare
for someone that takes 6.XXX
who hasn't been at this final
lecture because they read the
Underground Guide.
Here is an element that appeared
in the Underground
Guide a few years back.
There are no exams.
But there is a tradition of
hacking the Underground Guide.
So this is another example of
something that appeared.
So it all came about because
early in the teaching of
6.XXX, I was whining to the
students about the fact that
I've been at MIT for a long
time, since I was freshman.
And I still have yet to have
any person I report to say
anything about my teaching--
good, bad, indifferent.
Nothing.
Not a word.
So the students decided that it
would be interesting to see
if they could say something
sufficiently outrageous to
force a conversation between
the department
chairman and me.
And so far they've been
totally unsuccessful.
And I've tried everything.
Winston shows up late if
he shows up at all.
Good instructor but constantly
sipping from
a brown paper bag.
All kinds of stuff.
But there it is.
It's a lot of fun.
It's a little oversubscribed so
we have to have a lottery
and there's about 50% chance
and so on and so forth.
But many of you will find
it a good thing to do.
Oh, yeah.
And now I also want to remind
myself that there is a IP
event that's become kind
of MIT tradition.
It's the How to Speak
lecture that I give.
This year it'll be on
January 28 in 6120.
6120 holds about 120 people
and about 250 show up.
So if you want to go to that
lecture you should show up 15
minutes early.
That's a little secret just
between me and 6034 students.
It's about packaging, too.
It's a one-lecture
version of 6.XXX.
But it's very nonlinear because
one thing that you
pick up from that one hour may
make the difference between
you getting the job and some
other slug getting the job.
So it's one of those sorts of
things that can make a big
difference in a short
period of time.
You may sleep through 50 minutes
of the 55 minutes in
that lecture and stay awake
for that one magical five
minutes when you learn something
about when to tell a
joke or how to open a lecture
or how to conclude one or a
job talk or a sales presentation
or anything.
And that will make it
worthwhile for you.
And then of course there's
the possibility of your--
anybody who's doing [? UROP ?]
with me is likely to be
interested in what I've
recently come to call a strong
story hypothesis, something
that we've talked about
from time to time.
That's what makes us humans, and
that's not what they are.
They're orangutans or
chimpanzees, even though the
DNA that we share is--
it goes up and down.
At one point it was 96%,
then it went to 98%.
Now I think it's back
down to 97%.
But whatever we are, it's not
because of a huge, massive
difference in DNA between
us and our cousins.
So in my group we
build a Genesis
system, modestly called.
And that's what it looks like.
And it has in it all the sorts
of things that we've talked
about from time to
time in 6034.
And it's about to move into
areas that are especially
interesting, like can you detect
the onset of a possible
disaster before it happens
and intervene?
Can you retrieve presses based
on higher-level concepts?
Things of that sort.
Would you like to see
a demonstration?
OK.
So you've see in a little
bit of this before.
In fact, I'm not even sure
what exactly I've
already shown you.
Let me just get over there
to see what goes on here.
What I'm going to do right now
is I'm just going to read
about Macbeth.
A short precis of the plot.
Not the whole thing,
of course.
Just a few sentences
about the plot.
Right now what it's doing is
absorbing the English and
translating into a sort
of internal language.
A sort of universal, internal
language that's all about
trajectories and transitions
and social
relationships and all that.
So it's being read there by
two different persona.
Each of those personas has
a different educational
background, you might say.
They might represent
different cultures.
So eventually they build
up graphs like that.
And everything in white
is stuff that's
explicit in the story.
And everything that's in grey is
stuff that's been inferred
by the system.
So there are several layers
of understanding.
One is what's there
explicitly.
And the other thing that's
there is stuff
that's readily inferred.
And because these persona have
different educational
backgrounds, you might say
they see the killing of
Macbeth at the end of the
play in a different
light from one another.
One sees it as an act of insane
violence and the other
sees it as a consequence
of a revenge.
So once you've got that
capability, you can do all
sorts of things.
For example, you can
ask questions.
So let me arrange it to ask--
by the way, this is a live
demonstration.
I'm thrilled to pieces that
it actually works so far.
Let's see.
Why did Macbeth kill Duncan?
You all know why, right?
Yeah, you're right.
He didn't actually do that.
On a common sense level, neither
Dr. Jeckll nor Mr.
Hyde have an opinion.
On a reflective level neither
Dr. Jeckll more Mr. Hyde have
an opinion.
That's because it
didn't happen.
So we call these two persona
Doctor Jeckll and Mr. Hyde.
And you're ready to complain
right away about the spelling
of Jeckll, aren't you?
Well, that's because with this
spelling the speech generator
makes it sound a little bit more
like when we say Jekyll.
But what really happened is that
Macduff killed Macbeth.
On a common sense level, it
looks like Dr. Jeckll thinks
Macduff kills Macbeth because
Macduff is insane.
It looks like Mr. Hyde thinks
Macduff kills Macbeth because
Macbeth angers Macduff.
On a reflective level, it looks
like Dr. Jeckll thinks
Macduff kills Macbeth as part of
an act of insane violence.
It looks like Mr. Hyde thinks
Macduff kills Macbeth as part
of acts of mistake, Pyrrhic
victory, and revenge.
Isn't that cool?
I bet you'd get an A if you had
said that in eighth grade.
But once you've got this ability
to understand the
story from multiple points of
view, you begin to think of
all kinds of wonderful
things you can do.
For example, you can have Dr.
Jeckll negotiate with Mr.
Hyde, because Dr. Jeckll will
be able to understand Mr.
Hyde's point of view and
demonstrate to Mr. Hyde that
he thinks that point of
view is legitimate.
Or, Dr. Jeckll can teach Mr.
Hyde the subject matter of a
new domain.
Or Dr. Jeckll can watch what's
happening in Mr. Hyde's mind
and avert disaster before
it happens.
So let me just show you another
situation here.
I want to turn on the onset
detector and read another
little snippet.
This one is about--
what should I do?
We'll do the Russia and
Estonia cyber war.
It's reading background
knowledge right now.
But pretty soon, in the upper
left-hand corner, as it begins
to read the story, you'll see
it spotting the onset of
potential revenge operations
or potential Pyrrhic
victories, and show their
foundations begin to emerge
and giving the system an
opportunity to intervene.
So there you can see all
the things that it
thinks might happen.
Not all of them do happen.
But some of them do.
And you'll note, incidentally,
that this is another case of
Dr. Jeckll and Mr. Hyde having
different cultural
perspectives.
One's an ally of Russia and
one's an ally of Estonia.
One sees it as unwarranted
revenge and the other sees it
as teaching a lesson.
So, I don't know.
What else have we got here?
Oh yeah, president recall.
A long time ago, we talked
about doing information
retrieval based on vectors
of keyword counts.
That's cool but not this cool.
This is doing it on vectors of
concepts that appear in the
stories, such as revenge, even
though the word revenge
doesn't appear anywhere.
So because we're able to
understand the story on
multiple levels, we can use
those higher levels that don't
involve the words in the
story at all to drive
the retrieval process.
So all that is a consequence of
a variety of things, one of
which is the specialists that
translate the English into an
internal language.
And it's also, incidentally--
I mentioned it a
little before--
it's also a consequence of our
use of Gerry Sussman's
propagator architecture.
So a student comes into
our group and says
he wants to do something.
And we say OK, here's how
the system is organized.
It's like a bunch of boxes
that are wired together.
So you get a box and we'll tell
you what the inputs are
going to look like.
And we'll tell you what we
want on the outputs.
And if you don't like the
inputs, just ignore them.
And if we don't like
your outputs,
we'll just ignore that.
So nobody can screw up anything
because they have a
very circumscribed piece of
the system to work with.
So I can say, for example, the
president wanted Iraq to move
toward democracy.
And bingo.
That starts a propagation
through that network.
All this would be unconvincing,
in my view, if
it weren't eventually connected
to perception.
Because if it's not eventually
connected with perception,
it's yet another system that
demonstrates how smart it can
seem to be without actually
knowing anything.
So another half of what we do
is an early stage attempt to
connect the language stuff with
things that are going on
in the world.
So we say, imagine
a jumping action.
And there is a jumping action
that's part of a test sweep
developed by the Defense
Research Projects Agency to
drive what they call the Mind's
Eye program, which was
developed largely as a
consequence of work done here
at MIT, focused on the idea
that if we're going to
understand the nature of
intelligence, we have to
understand how language is
coupled into our perceptual
systems and how those perceptual
systems can answer
questions posed to them by
the language system.
That's a little demo of
the Genesis system.
Here are the issues that we're
trying to explore.
Nothing too serious, just
the nature of what is
extraordinarily fundamental
to any
explanation of human thinking.
Now, all of this might
turn you on.
And you say to me, well, you're
sick and tired of MIT.
You'd like to go somewhere
else for graduate school.
So now that I've demonstrated
what we do here, one of the
many things we do here, I'll
talk a little bit about other
places you can go.
This is a sort of MIT-centric
view of the world.
It represents all of the
places you could go
when I was a kid.
But while I've got
this particular
diagram on here, I just--
sort of testing my
MIT arrogance--
I remember a story
often told by my
colleague Peter Szolovits.
He says that when he came to a
job interview from Caltech to
MIT, he was sitting here
for three days and
nobody spoke to him.
So eventually he said, I've
got to do something.
He walked up to a graduate
student and said, hi my name
is Peter Szolovits.
I'm from Caltech.
And the graduate student
said, Caltech
sucks, and walked away.
Anyway, we've populated all
these places now that you see
here, and more.
This is just a list that I
scratched up this morning.
I'm sure I've forgotten many
that have equal right
to be on this list.
But in the end, which one you go
to depends on who you want
to apprentice yourself to.
Because a graduate school
is an apprenticeship.
And that means if you go to a
place with just one person,
it's OK if that's the person
you want to apprentice
yourself to.
Each of these places has a
different focus because they
have different people.
So you need to find out if
there's somebody at any of
these places.
It doesn't matter if it's
AI or some other field.
Theoretical physic-- you've
got to find out if there's
somebody at that place
you want to
apprentice yourself to.
So those site visits are
really important.
And I would like to also stress
that when you make your
application to graduate school,
it's very different
from applying to undergraduate
school.
Because they don't care whether
their school is good
for you at all.
They only care about
one thing--
whether you're good
for their school.
So don't get confused and
talk about how it's a
wonderful fit for you.
Because what they're interested
in is whether
you're going to contribute to
their research program.
Oh, I should say that if you're
applying to artificial
intelligence that means you
don't say, I'm interested in
all aspects of thinking.
You need to be focused.
There's another reason why
you don't say that you're
interested in all aspects of
thinking and that is the
defect theory of AI
career selection.
It seems to be the case, strange
though it may seem,
that people in artificial
intelligence often specialize
their research on the things
that they don't do very well
themselves.
So people who study language,
with the exception of Bob
Berwick, often have
trouble getting
out a coherent sentence.
And people who do hand-eye
coordination are the sorts who
spill their coffee.
So don't say you want to study
all thinking because--
The most extreme case of
this, though, is--
if you don't mind, I'll tell you
a story about an extreme
case in this.
We had a visitor from Japan
in the old artificial
intelligence lab
many years ago.
He came for a year.
Let's call him Yoshiaki,
just to pick a name.
Yoshiaki spent a year at the
artificial intelligence lab,
and he left his wife in Japan.
And the reason was,
she was pregnant.
And at that time, you could not
get a visa to the United
States unless you had a
smallpox vaccination.
And because she was pregnant,
she didn't want to get a
smallpox vaccination because
there's a small danger to the
fetus if you get a smallpox
vaccination
while you're pregnant.
So she stayed back there.
So Yoshiaki, let us call him--
it was a day before he
was to get on the
airplane to go home.
I walked into his office and
his desk was covered with
pictures of his wife.
By the way, Yoshiaki, I should
tell you, is a computer vision
guy, interested in object
recognition.
So you might suspect he
has some problem.
So he's looking at
these pictures.
I thought, oh my God, this
is a tender moment.
He's anticipating his
return to Japan and
reunion with his wife.
So I muttered something
to that effect.
And then he looked at me like
I was the king of the fools.
And he said, it's not a
question tenderness.
I'm afraid I won't recognize
her at the Tokyo airport.
So I said, Yoshiaki.
How can this be?
You study computer vision.
You study object recognition.
This is your wife.
How can you think you wouldn't
recognize her
at the Tokyo airport?
And then he looks at me, and--
God is my witness-- he says,
they all look alike.
Well now as we come
close to the end,
what are the big questions?
Is it useful?
Of course it's useful.
It's part of the toolkit, now,
of everybody who claims to be
a computer scientist.
What are the powerful ideas
and these things?
Well, here's the most powerful,
powerful idea is the
idea of powerful idea.
And here are a few
of my favorites.
No surprises there.
That's just Winston's picks.
But there's one more I
would like to add.
And that is all great
ideas are simple.
A lot of times we at MIT confuse
value with complexity.
And many of the things that
were the simplest in this
subject are actually
the most powerful.
So be careful about confusing
simplicity with triviality and
thinking that something can't
be important unless it's
complicated and deeply
mathematical.
It's usually the intuition
that's powerful, and the
mathematics is the [INAUDIBLE]
element.
Sometimes people argue that real
intelligence is possible.
One of the most common arguments
is, well what if we
had a room?
And you're in the room and
you're asked to translate some
Chinese documents.
You've got a bunch of books.
And in the end you could
do the translation.
But you cannot be said to
understand Chinese.
This is the argument
of Berkeley
philosopher named Sorel.
So the trouble is,
it's also true--
well, the argument is the books
aren't intelligent.
They're just ink on a page.
And the person is just a
computer, just a processor.
It doesn't actually
know anything.
So since it can't be in either
the person or the
books, it can't be.
And that just forgets that
there's a magic that comes
about when a running program,
when a process, executes in
time over knowledge that it
continually contributes to.
So the reductionist arguments
are among the many that have
been ineffectually posed to
argue that artificial
intelligence is impossible.
But that bears longer
discussion.
Let me just bring up the biggest
issue in my mind,
which is, it's not the question
of whether we humans
are too smart to have our
intelligence duplicated or
excelled in a computer.
It's a question whether we're
smart enough to pull it off.
I once had a pet raccoon.
Now, it's illegal to
have a pet raccoon.
But this one was an orphan.
Its mother had been hit
by a car or something.
A friend of mine brought the
raccoon to me knowing I kind
of like animals.
And I have to say, I kept
the raccoon for a year.
At that point, she wanted to
go out and be on her own.
So I had this raccoon.
And this raccoon is smarter than
any dog I've ever had.
Within a day, she learned
how to pry the
refrigerator door open.
So I spent that whole year
taping the refrigerator door
shut every time.
And then, we'd play jokes
on each other.
She wouldn't eat hot dogs.
And I wanted her to eat hot
dogs desperately because
they're cheap and
easy to serve.
All she would eat was cooked
chicken wings,
wouldn't eat hot dogs.
So one day I said, well, I'm
going to play a trick on her.
I took a chicken bone.
I stuck it in the middle of
a hot dog and put it in a
garbage can.
And she went for it.
Her genes took over and
she went for it.
And she was happy with
hot dogs ever after.
She wouldn't let me read.
She would crawl up underneath
the book and
interfere and make me--
she always wanted to suck on
my thumb, which turned blue
eventually.
You'd be amazed at how much
a raccoon can suck.
It's just extremely powerful.
The best parts were when she
would go bike riding with me.
I put on a heavy sweater
because they have
a pretty good grip.
I'd put on a heavy sweater and
she'd kind of mount herself on
my back and look out over my
shoulder, stopped traffic for
miles around.
So she was really smart.
But the interesting thing is
that at no point did I ever
presume that that raccoon was
smart enough to build a
machine that was as smart
as a raccoon.
So when we think that we can, it
involves a certain element
of hubris that may or may
not be justified.
Well, there it is.
Just a couple more
things to do.
One of which is, you should
understand that Kendra and
Kenny and Yuan and Martin and
Gleb are doing a lot of stuff
that's outside their
job description.
All of these quiz review deals
that they've arranged are not
in their job description.
I didn't ask them to do it.
That's all just plain
old professionalism.
So they've been wonderful
to work with and
I'd just like to--
[APPLAUSE]
--offer them a round
of applause.
And of course, Bob and Randy
and Mark have done fabulous
stuff as well.
And we of the staff, the TAs,
Mark, Bob, and Randy, have
nothing else to do except wish
you good hunting on the final
and a good long winter solstice
hibernation period
after that.
And that is the end
of the story.
And we hope you live
happily ever after.