Advertisement
English transcript
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 y
…
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 y
…