YouSub
Untitled

en English

Untitled English subtitles

0 downloads · 795 lines · added Nov 18, 2015

Report a problem
What's wrong?
Advertisement

English transcript

PATRICK WINSTON: We might wonder a little bit about the
nature of human intelligence, and we might reflect a little
bit on the kind of intelligence we've been
talking about in the past few weeks.
It's been an intelligence of sorts.
Those programs, support vector machines, boosting, they can
do really smart things.
But the peculiar thing about systems that use those methods
is that those systems don't have any idea about what
they're doing.
They don't know anything.
So they don't give us very much of an insight into the
nature of human intelligence.
And, after all, I'd like to have a model of human
intelligence because, let's face it, we're the smartest
things around.
So there are lots of ways to approach that question.
And we'll approach that question first from a
evolutionary point of view.
Some scientists believe, me for instance, that we have a
family tree that looks about like this.
Too small to see much of that, but the main point is that we
haven't been around very long.
We humans have been around maybe 200,000 years, and the
dinosaurs died out 60 million years ago.
So in the blink of an eye, we seem to have, more or less,
taken over.
When you look at that family tree on a scale where you can
see it, one of the characteristics is increasing
brain size.
There we are on the left, chimpanzee on the right.
Clearly, mostly mouth, not to much brain in there.
That one down below is a reconstruction of one of those
pithecus type bipedal apes from about 4 million
years ago or so.
So we became bipedal a long time before we
had much of a brain.
So you might think, well, maybe brain size has a lot to
do with it, and I suppose it does.
So we can plot brain volume of our ancestors versus time.
So the picture I just showed you was from about 3 million
years ago, I guess.
And then on the upper right hand corner, oh, that's not
just us, that's also the Neanderthals.
Their brains might have been slightly bigger than ours.
So it isn't just brain size.
Here's what that guy looks like.
That's a Neanderthal, of course, on the left.
And that's one of us on the right.
Some conspicuous differences, big heads, the rib cage is
kind of conical in shape, a large pelvis.
People like to make a lot of speculations about how they
must have moved around.
But one thing is plain, they didn't amount to much.
They could make stone tools, but those stone tools didn't
change much over tens of thousands of years.
And that was pretty much the story with us too, until
something happened, probably in southern Africa, probably
in a group of individuals, maybe less than 1,000.
And what's the evidence for that?
The evidence for that is mostly--
comes from DNA studies with a lot of probabilistic
assumptions, and Monte Carlo simulations, but it seems
among the competing hypothesis for how we came to populate
the world, it seems that there was a group of us, homo
sapiens, in Southern Africa that got something that nobody
else had and the highest probability scenario is that
we quickly took over.
That population of homo sapiens dominated the rest,
went out of Africa, and within the blink of an eye did that
sort of stuff.
What's that sort of stuff?
Those two paintings are from Lascaux about
25,000 years ago.
Paleoanthropologists, like Tattersall, take that as plain
evidence that there was symbolic thought on the people
who were around at that time, us homo sapiens.
The head is a carving of a mastodon tusk of a woman
25,000 years ago.
Also, plainly symbolic, people are making a lot of jewelry
and doing self adornment.
The Neanderthals never seemed to do that.
That jewelry making seems to have gone back to Southern
Africa, maybe 70,000 years ago.
People were puncturing seashells and using them as
necklaces, apparently.
So something happened, and the paleoanthropologists, who
write fascinating stuff, don't quite know how to talk about
it other than to say it looks like we
became somehow symbolic.
And somehow that has something to do with language.
So if you talk to Noam Chomsky he will say--
let me get this precise, this is near his
quotation as I can get.
He thinks it was the ability to take two concepts and put
them together to form a third concept without disturbing the
original concepts and without limit, and each
part of that's important.
The without limit part is what separates us from species that
might be able to do that a little bit, but we can do it
without any apparent limit.
So that's a linguist speaking.
He talks a lot about the merge operation, and
combinators and language--
using terms foreign to us.
Better not use the term "combinator," it's kind of a
computer science term.
But whatever it is, it seemed to happen about that time.
It didn't happen slowly and proportion to brain size, it
seemed to happen all of a sudden in consequence of a
brain that had grown big enough to be an enablement but
the capability was not what pulled
evolution in that direction.
So, I believe, that whatever that was, that capability,
enabled humans, us humans, to tell and understand stories,
and that's what separates us from the other primates.
That ability to--
The symbolic ability, whatever it is, enabled storytelling
and understanding.
And that's what all education is about and that's why our
species is special.
So what we're going to talk about today is something you
might think of as an instantiation of that
hypothesis, one way of thinking about it.
And it's a way of thinking about what the linguists would
call the inner language.
It's not the language with which we communicate, it's the
language with which we think, which is closely related to
the language with which we communicate, but may not be
quite the same thing.
So many of you are bilingual.
Chris is bilingual.
Chris, have you ever had the experience of remembering that
someone said something to you, but not remembering what
language they used?
STUDENT: [INAUDIBLE].
PATRICK WINSTON: How about [? you, Sian, ?] have you had
that experience?
STUDENT: [INAUDIBLE].
PATRICK WINSTON: What?
STUDENT: I always [INAUDIBLE].
PATRICK WINSTON: [INAUDIBLE] experience of having--
David, have you ever had that experience of remembering some
conversation but not remembering the language in
which it was cast?
STUDENT: Well, you remember something you usually don't
know which language it was in anyway.
PATRICK WINSTON: You usually don't remember.
That's a common view.
You remember something was said, that there was a
conversation that had some content, but if it's with a
speaker of your own language and your embedded in another
place, you often don't remember what language the
conversation was in.
Is that right, [? Wana, ?] you remember things like that?
Not sure?
STUDENT: People who [INAUDIBLE] speak in a certain
language, so--
PATRICK WINSTON: Sometimes you don't have that confusion, she
says, because you always speak to particular people in a
particular language.
But many people report that they have that experience of
not remembering which language something was said in.
Well, OK, so what are we going to do?
We need an inner language, and maybe we can start just by
saying, let's have something that looks sort
of familiar to us.
We have an object and it's supported by some other
objects, so those are support relations.
That's one example of what we might call a semantic net.
It's a network that's got nodes and links, it's got--
it has some meaning.
That's where the word "semantic" comes from.
Well, we might have another example that looks like this.
There's Macbeth, there's Duncan, and Macbeth murders
Duncan, then we also know, somehow, there's a kill
involved as a consequence, and then ultimately,
Duncan has a property.
And that property is the property of being dead.
So there's another semantic net recording something that
happens in Shakespeare's Macbeth plot.
Now, we can decorate that a little bit.
So as to get a couple of other concepts in play.
First of all, the thing we've got already is we've got
combinators.
Then combinators.
Well, a fancy name for those links that connect the nodes.
Another thing we've got is an opportunity for connecting the
links themselves.
So the murder sort of implies the kill, and the kill leads
us to conclude that the victim is dead.
So that is treating the links themselves as objects that can
be the subject or object of other links.
So we call that process "reification."
Now, in artificial intelligence, semantic nets
we're all over the early work, but if you have a big network
that covers the wall you need some way of putting a
spotlight on some pieces of it.
So Marvin Minsky put a lot of technical content into that
idea and created the notion of--
a notion that deserves another color here--
he suggested that we need a localization process, so we
have frames, so-called frames or templates.
And a frame for this murder action might be that there's a
murder frame that has an agent and has a victim, and the
agent is Macbeth, and the victim is Duncan.
So that's a way of putting a localization layer on top of
what we've got so far.
Later on, I'll add sequence to that list.
So this is where it rested for a long time, and in some
sense, still rests there because as soon as you've got
combinators, you've got something
that's pretty much universal.
You can do anything with it.
The trouble is it's sort of down at the bit level.
It's like assembly code.
It doesn't have, as a concept, enough organization to help
you go to the next level of achievement.
There's also a little problem here that deserves also, some
mention, and that is that we have, over this whole thing,
the problem of parasitic semantics.
Parasitic semantics.
A kind of ugliness that surrounds this whole concept
because when we look at a diagram like that, and we say,
oh, Macbeth murdered Duncan, that means
Duncan's the victim.
We know there must have been a motive, maybe Macbeth wanted
to be king.
Well, we know all that stuff, and there's a tendency to
project that knowing into the machine.
If you're going to play with your telephone, please leave.
So if we project meaning into that that's our understanding
that's not the machine's understanding.
So much of the meaning can be said to be parasitic.
We're the parasite, and we're projecting
the into that thing.
Putting that diagram into some machine form doesn't mean the
machine knows anything.
It might be able to conclude some things, but it's
understanding is not grounded in any kind of contact with
the physical world.
So we have to worry a lot about that and where
philosophers would stop there and go off and write a few
books on the subject.
But we're not philosophers so we're going to just mention
the problem and go barreling ahead.
So we need to use this notion of semantic net, and we have
to ask ourselves some questions about what elements
of the inner language are most useful and yet it might be
very complicated, but here's usefulness number one.
The notion of classification.
So we know about stuff, and we know about,
for example, pianos.
And we know about tools.
And we know about maps.
But we know about those things on different levels.
So say I'm thinking about a tool, do you have a very good
image of what I'm talking about?
The answer has to be no because the notion of a tool
is very vague, so it's hard to form a picture of
what that's all about.
On the other hand, if I said I'm thinking about a mac.
Well, this is interesting because there's lexical
ambiguity there.
You don't know if I'm talking about the Apple type Mac or
the apple type Mac, or should I say the
fruit or the computer?
So there's lexical ambiguity there at two levels or more.
But let's fill this in a little bit.
If I know I'm talking about a piano, you can form a picture
of that so that seems to be on a more detailed level where
you can do hallucination.
At a higher level you have just a musical instrument.
And I can give you a tool to think about by writing hammer.
And if I'm going to have a mac, it's
going to be an apple.
And in this case, I want you to think about a fruit.
And down here, I can be more specific about
these things too.
I can add a slight refinement of detail and say, I'm
thinking about one of these.
Do you know what this is?
STUDENT: [INAUDIBLE].
PATRICK WINSTON: No, it's not a mere hammer, it's a
ball-peen hammer.
In some circles, it's called a ladies hammer.
I don't know why.
But it's--
What's it for?
Most people buy it mostly because it's
small and light weight.
But, in fact, it's for metal working.
It's for taking a piece of sheet metal and pounding it
out into an ashtray or something
or for seating rivets.
It's a metal worker's hammer.
So you might not have known about that before but now, at
least, you have a word to hang that knowledge on.
It's a ball peen hammer.
So we have various levels here, going from very specific
to very general.
And we can even go to a level of specificity for pianos by
saying we've got a Bosendorfer.
Why is a Bosendorfer special?
I mean, is it like a Baldwin?
Something's [INAUDIBLE], Yoka--
Lots of piano types, what's special?
You see, you don't know because unless you play the
piano, and probably unless you're a serious piano player
you don't know that a Bosendorfer--
Ariel, you know.
STUDENT: I think its supposed to have an extra octave at the
bottom, black keys.
Pretty cool.
PATRICK WINSTON: It's got some extra keys at the bottom.
And most people don't know that unless they're serious
about the piano.
Some professional piano players, when they're
confronted with a Bosendorfer have to have someone cover
those keys because it screws up their peripheral vision,
and hit the wrong key.
Because they're not used to having those
extra keys at the bottom.
So that's a little detail but the Bosendorfer.
So you can make a kind of graph, and you can say, let's
go from low, very general, to a basic level,
to a specific level.
So it is the case in human knowledge that that graph has
a tendency to look sort of like this.
So here's tool, here's hammer, here's ball peen.
So that level, where you have a big jump, that's the general
to basic level of transition.
So that basic level is probably there because that's
the level on which we hang a huge amount of our knowledge.
We know a lot about pianos, and it all seems to be hanging
on that word piano, which gives us
power with the concept.
So that's example number one of an element
of our inner language.
The ability to assemble things into hierarchies like that,
and hang knowledge about those objects on the elements in
that hierarchy.
Well, given that you have elements in the hierarchy, how
do you talk about them?
Well, I like to consider the possibility, just for the sake
of illustration, that you're thinking about a car crashing
into a wall.
So you've got things to think about like the speed of the
car, the distance to the wall, and the condition of the car.
And you've got the period before the crash, during the
crash, and after the crash.
So you might want to think about how to talk about those
objects in those three time periods.
So we can do that with a vocabulary of change and we do
that because we believe that most of human thinking is
thinking about change causing change.
And that flies in the face of what we learned as engineers.
Because in engineering, we learn about state, and once
you know the state of the system, you know everything
you need to know in order to predict the future.
The trouble is, in our heads, thinking about everything
there is in the world, including the current phase of
the moon, is too much stuff.
So mostly, our thinking, we think, is hinging on the idea
that change leads to change.
So that's why we have a vocabulary of change.
So in the period before the crash, the speed of the car is
not changing.
There's a little notation for not change, no delta.
The distance to the wall, that's decreasing.
The condition of the car, that's not changing.
Then, the car hits the wall.
So the speed of the car disappears, the distance to
the wall disappears, and the condition of the car will
change dramatically.
Finally, after the crash is over, the speed of the car
does not appear.
The distance to the wall does not change, and the condition
of the car also does not change.
So that's hinting at a vocabulary of change, and its
use, which will be the second element in our development of
a vocabulary of ways in which we might have constructed are
inner language.
So this a particular idea, that's classification.
This is transition, and a system that purports to
understand stories with a heavy emphasis on this notion
of transition and we believe, that is to say, I think, that
that vocabulary has to have decrease, increase, change,
appear, and disappear.
So there are 10 things you can have in such diagrams.
I've done five.
That's because, for each of those, there's a not
variation on that.
So with a vocabulary of 10 things can go a long way
toward helping to describe things that are in process of
change and making transition.
And we have a lot of those words in our vocabulary.
We use those words a lot in our vocabulary.
They seem heavily connected with vision.
Our friend appeared.
The cat disappeared.
The speed increased.
So this is a description of a crash.
In terms of those kinds of elements, now, I say to you,
how does a camera work?
Well, I could say the camera works because a photon crashes
into a photo receptor.
So when I say a photon crashes into a photo receptors, why am
I saying that, and how does it help?
I'm saying that and it helps because it's the same pattern
of change you already know about when you talk about a
car crashing into a wall.
How does that work?
The speed of the photon, the distance of the photon to the
receptor, and the condition of the receptor.
So analogies like that are very much of the core of what
we think about all the time.
Really then, there is representation number two.
Number one is class, number two is transition, and now,
you're ready for number three, which is trajectory.
Linguists, who study sentences, often talk in terms
of fundamental patterns that seemed to be in a lot of what
we say, and a lot of what we say is about objects moving
along trajectories.
So we can talk about a trajectory frame.
And a trajectory frame will have elements like this.
It has an object moving on a trajectory that ends up at a
destination.
You might start out at a source.
It's probably been arranged by some kind of agent, and the
agent may assist himself making the motion happen with
some kind of instrument.
There might be somebody helping out
over here, a co-agent.
Well, what else can we have?
A beneficiary, someone is helped out by the action.
Sometimes, the motion is arranged by a conveyance.
So these are a lot of slots, finite slots, and descriptions
of actions, many of which involve motion on a
trajectory.
We have a tendency in language to decorate these things in
one way or another, depending on the language.
And so in many languages, the decoration is by way of
position in the sentence.
In English, it's often by way of a preposition.
It's used to help zero in on a particular role of an object
in the trajectory scenario.
So if I say, I baked a cake with a friend, there's a with
preposition.
If I baked a cake for a friend, the friend is the
beneficiary.
If I baked a cake with an oven, that's an instrument.
The object may be moving to a destination from a source, and
if I'm going to New York by train, I put a
by on top of that.
If the agent isn't in subject position, I would say
something like, oh, I see all the work
was done by a student.
So those prepositions have a tendency to help us zoom in on
the actual role of particular objects in this whole package,
the whole frame.
So this is number three.
There's a variation on this in which there's no actual
trajectory in which case we'll just call that a role frame.
Because if there's no trajectory, we can still have
things such as an instrument, a co-agent, and a beneficiary.
So now, we've got three representations.
You might say, well, what good are the
…

Need subtitles from another video?