Artificial Intelligence

Prof. Deepak Khemani

Department of Computer Science and Engineering
Indian Institute of Technology, Madras

Lecture - 5
Introduction Philosophy

So, let us get back to this question on the philosophical side of our introduction, so what
are mines essentially. So, if you remember the view from Aristotle, it said that our
thoughts are our reflection of what is out there. If I see an apple or if I think of an apple, it
is because there is an apple out there and my thought is in the image of that apple that is out
there. The diametrically opposite view was given to us by Canter much much later; he said
that what is out there is the reflection of our thoughts that what we think is out there is what
we think is out there.

Now you see there is a gap that you can never cross, because what you think is out there is
only what you are thinking is out there, what is really out there. So, you see a person you
know person sitting here. So, I think there is a person sitting here especially. So, I have this
concept of people; I have concepts of you know boys, girls, men, women, all kind of
concepts, trees, chairs. So, that is how I think about the world and it is out there. So, I also
have concepts about clouds and you see an image of something in the clouds; you see a
dog in the clouds or you know something else.

(Refer Slide Time: 01:52)

What is out there really is a question we want to start with and then we want to we have
seen already that the notion of the mind has come about the people understood that there is
something called the mind; gradually over a period of time when they realize that what
they are thinking about is not in direct correspondence with what is out there. But now
what we want to discuss is that it is a necessity that if you have to have an intelligent agent,
we have to have a mind which can represent the world in a way that it can manage the
representation effectively.

So, that is the question I am driving at today. So, let us start with the basic question as that
what is reality; of course, we live in a world. We are also part of this world, but what is the
objective reality? If you wanted to create a model of the world, how would I do it? No what
is really out there, any suggestions? We are talking about reality. So, you are saying that
reality is as we see a collection of situations which we can manipulate. No, but simpler
things like people take this young man sitting over here; is he real or what is the young
man; what is the human being?

(Refer Slide Time: 03:31)

Okay, now let us take a scientific enforce point of view, the physics point of view;
everything in the physical world is made up of a small number of fundamental particles.
So, whereas a small number of fundamental particles I mean there is a number of different
particles are small; it is not that the number of particles is small. The different type is small
actually. So, depending on what theory you are following. So, the simplest one could
follow the bore atom module, for example; you know an atom is made up of protons and
neutrons and electrons and everything else in the world is made of these three kind of
things essentially.

So, physics is, of course, I have been struggling to figure out what the world is made up of
but let us take it for granted at it is some fundamental particle; it could be something
smaller that proton and neutron, but everything is there, and these particles they obey some
laws of physics. So, the laws of physics are sufficient to explain what is happening in the
world out there. So, that is a first assumption that we will work with and it is not something
that we can dispute because then you are saying the physics is wrong essentially. So, like
Penrose was saying that the physics of the brain we do not understand, then you would be
saying that we do not understand the physics of this world essentially.

So, the physics is believe that that you know we do not know what the. So, there is a sling
theory, then they are neutrinos and gluons and that kind of stuff, but there is something out
there. And they all behave according to the laws of physics which we are trying to
discover; by the way they behave according to these laws, and, therefore, everything can
be exchanged according to these laws. So, that is enough to understand the world in some
sense. So, we are talking about people, what is the person. So, even adult, it is made up of
about 10 raise to 27 atoms. I do not know whether you can visualize the number 10 raise to
27, but you should try; it is a huge number.

So, we can think of a person or person not in the social sense, but in the physical sense; a
human body is made up of about 10 raise to 27 atoms. So, this young man sitting over there
is just a collection of a 10 rise to 27 atoms which for some reason decide to stick together,
and the reason is given by the laws of physics. It can be explained by the laws of physics,
and of course, we may call it in different terms. We may talk about biology or chemistry
and so on, but deep down they obey these laws of physical is fundamental particles obeyed
essentially. So, we have mention earlier that there are two views of the world. One is
materialist view which says that there is only matter and the other is the idealism view
which says that there is no matter essentially.

So, which is why in the last line I had said that if there are fundamental particles; you know
some people believe that it is all energy or something out there essentially, but that let us
not get into that kind of a thing. So, we are trying to understand the world; we are trying to
model the world; we are trying to why do you want to do that? Because an intelligent agent
and we discussed this in the early lectures should be able to operate in a meaningful
manner in the world. Do something useful for itself and achieve its goals and you know
learn and that kind of stuff. So, it needs to represent the world out there, but what is the
world out there is the question we are asking.

(Refer Slide Time: 07:35)

So, human beings is about 10 raise to 27 atoms, and these 10 raise to 27 atoms are
continuously interacting with zillions of other atoms out there; you know we breathe, we
eat, we have sound wave impinging upon us which are also made up of oscillating particles
and so on and so forth essentially. So, if we try to talk about reality in these terms like a
physicist, can we ever make sense of this world out there? And when I say it makes sense
of this world as I mean in the practical sense as an intelligent agent in some environment
trying to do useful things especially.

So, can we ever hope to write down the equations for them and solve them even if we
know the equations and what would we get if we solve them especially. The trajectories of
their location how does that help us? You know that atom number 259 is moving in this
direction or something, does not help. So, if you want to interact with the world, we need
to have our own level of presentation, and this is what I am going to be driving at today.
The world that we interact with like handset, it is a world in our minds especially.
(Refer Slide Time: 08:58)

So, we have already said that the world around us and including us operates according to
and can in principle be explained by the fundamental laws of physics; nothing else is really
needed, but it is too big for us to work with, because the number of amount of information
we would have to use would be too much essentially. What we do is that we create our own
worlds in our minds, and it is only our creation that is meaningful to us essentially. So,
coming back to the example of a person, I do not think of a person as an example of 10
raise to 27 atoms behaving in some concerted action.

I think of it as a single entity a person who is sitting on a single entity called a chair or
eating a single entity which is I call a dosa some masala dosa; for example, I think of it as
one thing essentially. Of course, it is not one thing; it is made up of so many things and so
many processes and I do not want to get into that at all. So, we create levels of abstraction
at which we represent things and reason at those levels of abstractions. So, we should also
keep in mind Newell and Simon physical symbol system hypothesis is that we can create
symbol systems and manipulate that symbol; that is enough for us to reason about the
world in an intelligent fashion essentially.

So, Douglas Osheroff who I will talk about again in a moment has a different notion of a
symbol which we will not peruse very much here. He talks about how the mind creates
symbol, how the brain creates these same symbols that we are talking about. We stand for
a person and so on and so forth. So, when I was talking about a symbol, I said something
which was perceptible. So, for example, I write the name of a person; I can read it
essentially or I can you know type it in a word processor and things like that, but Osheroff
also talks about symbol processing in the human brain that the human brain leads symbols
and what are the symbols? These are kind of concerted patterns of activity in thousands of
neurons.

So, we will not go into more detail than that, but somehow neurons act in concert with each
other in a manner which we are inspecting ourselves, think of a symbols that we are
reasoning with symbols. So, this kind of an idea has been exploited in movies; of course,
movies do not necessarily depict reality. So, again coming back to the matrix, so if those of
you have seen matrix would remember that its main character called Neil Neo; in Neo’s
mind, Neo was a software engineer working in New York city, and that is how the movie
begins, but in reality, whatever, that means, neo is in some cell in some human battery
which the machines have constructed to extract the energy out of him.

So, which is the complicated sequence which we have to sought off watch the movie
carefully to understand that when eventually that something is pulled out of his brain, he
really finds himself in some very unknown like place you know inside some cell where he
is just unit of a large battery or something like that. The important thing that I am trying to
say is that it uses this idea that the world we live in is actually in our minds out there
essentially which is why, of course, sometimes people can hallucinate. They imagine
something what is not out there, because their minds are sought of not in sync with again
let me use the word reality out there.

But there is a big case in that you know we do not know what is reality; we only know what
we know essentially. Inception is another film in which you would not know whether you
are dreaming or whether you are in some real world. And this is not the physical system
hypothesis says that if you can create a level of representation and is it not that level that
should be enough for creating intelligence systems.
(Refer Slide Time: 13:22)

We have dense slides that I have put in; we are not going to go through this slide. It is just
to illustrate the levels of scale at which our concepts exist. Remember that in the end
everything is made up of examples of this fundamental particles, but then we talk of
people, we will talk of football field, we talk of a planet. All these are at different levels of
scale. There is a very nice movie; I do not know how many of you have seen it called,
‘powers of ten’ which is available on the web, and there is a link I have given here which
essentially is rooms out of a level in which a couple is seen in a park and goes to the very
top most level here which you can see a 10 rise to 26 meters and then zooms down back
and goes to the bottom most level which is 10 raise to minus 17 meters and things like that.

So, at different levels you see the world with a different perspective. Suggest some
examples, for example, mustard seed is about one millimeter thick whereas the distance
from Chennai to Pune is about 1100 kilometers. So, in these powers of ten, you keep
magnifying the image all reducing the image by powers of ten every time and then you sort
of keep diving in deeper and so on and so forth. So, you can see at the top most level, there
are very things that we cannot I mean our mind bowels to talk about something like ten
billion light years.

First of all, you have to imagine what is a light year or remember what is a light year. It is
a distance covered by light in one year and we do not even think of light having speed. I
mean I just see you and you raise your hand and I see instantaneously; where is the
question of speed? We do not even have a notion of speed; of course nowadays we have
thanks to Einstein and all these people but light travels at a finite speed. And we do not
realize it because we know it is so fast for all of us here that everything happens
instantaneously.

We do not suffer from these effects of relativity and things like that. But ten billion light
years, how much would the light travel in ten billion years? I mean we cannot even think
of these kinds of things especially, alright, the very extreme end 10 raise to minus 17
meters where you have quacks and gluons and so on and so forth. So, the world as we think
of nowadays exists at these very different scales of things; that are at one level. At another
level, it is all just collections of fundamental particles which we have already agreed that
we cannot deal with at that level; that is why we think of these different scales especially,
okay.

(Refer Slide Time: 16:30)

So, our perceptible universe is a small subset of the scales essentially. So, the largest thing
we can see is may be about a kilometer across. So, like the golden gate bridge and San
Fransisco or something like that, and the smallest thing is may be a mustard seed or may be
some people can see pollen you know there is rays of light and you can see dust particles.
Some of them may be 0.1 millimeter across. So, you can perhaps think about things at this
level of scale essentially. So, the human mind as far I am trying to get at; the human mind
has evolved to create concepts at these scales essentially.
So, we tend to think of objects at these scales; that is why we are comfortable with this. We
are not even comfortable talking about how far the planet Jupiter is from here, because we
cannot even imagine that kind of a thing. And at least we have not evolved to imagine that
kind of a thing, and to reasons at different levels, we have created this different discipline.
So, we become a specialist in biology or geography or aestrophysics or anything.

(Refer Slide Time: 17:41)

But each discipline operates in its own level of scale; social science is operated some level
where we are talking about collections of human beings. Remember human being, each
human being is collection of 10 raise to 27 particles. So, social science in some sense is
talking about collections of 10 raise to 27 particles, but we do not, obviously, think in
terms of fundamental particles any longer, and we have these different disciplines
especially.
(Refer Slide Time: 18:06)

So, Hofstadter, so remember this mind-body problem which Decant was blipping with that
if there is this world of the mind which is reasoning about the real physical world out there
the body; how do the two things interact essentially? Nor the physics we say we have laws
of physics of the fundamental particle level, and if you know that, you will know how the
rest of the system is having essentially. But Hofstadter says that we have to introduce a
notion of what he calls as downward causality which means the causality is from a higher
level to a lower level.

Even though the laws of physics can explain going from particle level to ensemble particle
level, he says that is not useful for us, we have to think about how. So, for example, if I
want to drink a cup of tea, then I am thinking at a level about cup of tea and so on and so
forth. And this level of thinking which is happening with its concepts at this level of
abstraction is eventually driving at one level; you might say my muscles or my nerve cells
or something at even lower level you might say the very fundamental particles which make
up my hand, for example; in such a manner that my hand eventually reaches out for that
cup of tea and pick it up and you know take a sip from it essentially.

So, the causality is from our level of reasoning to the lower level where things are actually
happening. Now physics, of course, does not have a notion of causality; that is why
Canter’s even when he was talking about human categories, he was saying that is face and
causality are given to us that we accept we have to start working with those things
essentially.

(Refer Slide Time: 20:09)

So, these things are calling epiphenomenon. So, things like pressure, for example, in a
balloon, we talk of pressure, but what is really happening that lower level activities you
know molecules of different kinds of molecules in air, nitrogen, hydrogen, carbon dioxide,
everything; they are moving around randomly and you know impinging upon the inner
surface of the balloon. And this cumulative activity or the epiphenomenon phenomenal of
pressure is dealt essentially. Likewise, in our human brains, there are these billions of
neurons which are firing away in some fashion.

We do not tend to think of our brain in that fashion. I tend to think of my brain and say, oh,
I want to have this cup of tea which is operating at a very higher level essentially. So, can
a machine operate at epiphenomenon level like this, and that we feel is necessary for
machines to be intelligent. So, we have run through this. So, computers are manmade
objects; we know how they operate. So, it is easy for us to explain; for example, if you type
something in a word processor, in principle somebody can say that at the lowest level,
these are the kind of micro level operations which were taking place. But we do not do
that, of course; we as human beings tend to think of machines as doing higher level things.

So, how do we see a machine computer? As a music player or a web browser or a game or
any of these many things that a machine can do; so the important thing is that many people
have called starting with tuning is there is a universal machine. The computer just like us is
a universal machine. So, if you want to call yourself as a universal machine and the
universal machine is a machine which not only, of course, a simple machine can do only
what it is designed to do, but a universal machine can imitate other machines and do what
they were doing.

So, they are flexible in that nature. Scans as a machine be intelligent; Hofstadter, so I come
back to Hofstadter. He says that if this machine can introspect and examine its own
behavior, then it is possible for it to become intelligent essentially. So, he is going one
more step from Newell and Simon. Simon and Newell they said that if you can create
symbolic representations and create algorithm, you still work on this representation; that is
sufficient and necessary to create intelligent behavior. Now we can see that that is at one
level of.

So, there have been layers and layers that one has to talk about. In computers, we have bit
level representation; they have machine code, assembly language, higher level languages,
higher level data structures, representations, objects, all kinds of things and you keep
going higher, likewise in the real world out there. So, Newell and Simon said that one level
of representation which he calls as the symbol level is enough, but now Hofstadter is going
one step further. He is saying that in addition to that, you need this capability to introspect
essentially. So, if you read his book, it is quite an interesting book to read; I must change
the loop.

He sought of goes through a long detailed argument of how Godel discovered this idea of
self reference in Rusell and Whitehead Principia Mathematica, and this in spite of the fact
that Rusell and Whitehead went out to formalize everything, and they wanted to keep
away self referential structure. So, they had layered logical representation or typed logical
representation where self reference would not be possible. The same type of an argument
element could not be an argument to a sentence in that same language essentially, but
Godel constructed this very elaborate; he gave us very elaborate mechanism of how to
construct a sentence.

So, there is this two levels at which things are operating upon. One is this level of number
theory which is principle of mathematics is all about, but there is also this level of
encoding things into this number theory and then encoding sentences like I am lying or
something like that essentially or this sentence is not true and things like that. So,
Hofstadter is saying that if a machine can have this capacity to introspect and reason about
its own actions which means also reason about other people’s actions, then it can in
principle be intelligent.

So, let us talk about intelligence agents for a moment; it is a very popular term nowadays.
So, these are programs; we will talk of intelligence recent programs which are persistent
which means like the operating systems, for example; if we leave a machine on that is exist
all the times essentially. They are autonomous which means that nobody is saying that run
this program, run this routine or call this subroutine or something like that; they are
proactive. If the see an opportunity in the environment, they sense an environment; they
will go after it essentially.

And they are goal directed which means that you know they have goals; of course, these
goals may not be self generated; they could be given by the creator essentially. It is just
like we have that secret agents and the governments are supposed to have who have all
these properties they are persistent, autonomous and proactive, but they carry out the
billing of the government essentially.

(Refer Slide Time: 26:20)

This is a rough diagram of what an intelligent agent should be like. So, that white figure at
it is supposed to be the agent, and the thing inside that is what is in the head of the agent.
And what is in the head of the agent is the module of the world out there, and the module of
the world should contain itself which means it can introspect on itself. And, obviously, you
might ask the question as to it that if the model of the agent has a world in which the agent
is there, then in that agents head also I should create the model of the world. So, there is an
infinite level of next thing which is possible in principle. So, they are these kinds of very
curious loops which can form, okay.

(Refer Slide Time: 27:13)

So, we are slowly coming towards moving away from history and coming to what we want
to do. If you want to build an intelligent agent which interacts with the real world, then you
have to have at least these layers of different kinds of reasoning. One is the outermost layer
is what we can call a signal processing. It means you are receiving signals, sound waves,
light waves or whatever from the world, and the innermost layer is symbolic reasoning
which is what is classically AI is what all about that you can create symbol systems and
reason with them.

And you may have and this is my interpretation of this whole thing that an intermediate
layer of neuro-fuzzy systems which serve the purpose of converting signals into symbols
essentially. So, for example, if I am speaking and what I am creating is a signal which is
you know sound waves of a particular pattern, but your brain is converting these sound
waves into linguistic entities essentially. So, you are recognizing words out of this sound
wave; from these signals you are extracting symbols essentially.

So, if I say the word apple, it may be certain sound wave which is meaningless in itself just
like neural activity in our head is meaningless in itself, but you can process it to understand
at a higher level to stand for a symbol apple, and neural networks are particularly good at
doing this kind of things. So, you must have heard about fact that character recognition if I
were to draw the letter a on a hand written characters.

(Refer Slide Time: 28:57)

So, if I write A like this, if I write A like this, if I write A like this; we have no difficulty in
recognizing that these are A and neural network is also very good at this sort of a thing. But
at some point, you may start getting a doubt about whether I am writing an A or whether I
am writing an H. And it is very difficult to describe rules to say that this sequence of
segment forms A, this sequence of segments forms H and so on. Whether the learning
system which will learn these characters in context of other letters around them will
eventually learn to recognize the character A for example,

So, neural networks are very good at this sort of a thing, but if you want to do give an
explanation of let us say the Pythagoras theorem, what is the Pythagoras theorem and how
do we prove it. Then neural networks are not really very good at that kind of a thing; for
that we need this symbol manipulation ability which everybody is from Simon Hostettler
is saying it is necessary for intelligent behaviors essentially.
(Refer Slide Time: 30:06)

So, these are the topics that one can identify if you look at AI in general. This is not the
topics of this course. If you look at the enterprise of AI, then we have all kinds of topics
here, knowledge representations, semantics, anthology, models, search, memory, machine
running, problem solving, planning, adversarial reasoning, qualitative reasoning, natural
language understanding and all kinds of topics. So, on the left what I have drawn in this
figure are the sensing kind of activities. Signal to symbol kind of activities, speech
processing, image processing, video processing, computer vision, neural networks, pattern
recognition, studs sensors and that kind of thing.

On the right hand side, it is the opposite from symbol to signal. So, you have motor control
of. If you want to build robots, you have to eventually make the robot do what the robot is
thinking about doing. If the robot is thinking about going from place A to place B, it must
do something to make their physical movement possible. So, we need activators and things
like that at that time essentially. So, these are the topics of AI. In the circle, basically this
figure is taken from my book the circle roughly it is a kind of describe what is there in the
book.

So, of course, we are doing this course on AI which we are doing here, but in our
department, there are a whole lot of courses which cover these areas. And I just want to
give you some idea of the kind of courses that we offer. The first four courses that will
come are courses which I am personally involved with, but the rest of the courses you
know mostly my colleagues are handling.

(Refer Slide Time: 31:56)

We will start with this course which is Ai which kind of covers some of this stuff inside
this thing here, then planning and constraint satisfaction. These are names of courses. So,
this will be offered next semester; for example, both these courses will be offered next
semester knowledge representation reasoning as well. Then there are other courses which
my colleagues teach machine learning which is being offered now. Pattern recognition is
also being offered now I think. Natural language processing as well is being offered at this
moment.

Probabilistic reasoning is not being offered at this moment; very often I think Dr.
Ravindran offers it as a self study course. Then we have computer vision; I am not quite
sure which semester, may be this semester it is being offered. Speech technology, Kernel
methods, this is I think next semester, visual video processing, computer graphics. We do
not have so much on the output side. So, you can see our department is not very strong in
things like robotics. So, we do not really offer courses in that. Imagine mechanical
department may be offering some courses. So, in terms of assignment, I might have
mentioned this earlier; one assignment is going to be on game playing.

So, I will try to do game playing not in this order but a little bit earlier than this order; may
be after heuristic search or something like that. So, that you can get going, and we will
decide which game and we have to implement an algorithm for it. And your programs will
play against each other that kind of stuff. And another assignment would be
implementation of some of these algorithms.

(Refer Slide Time: 33:44)

So, we will sort of assign some algorithm and you should implement in that way; we will
go into the details as we go along course, okay.

(Refer Slide Time: 33:55)

The text books as I said it is a text book which I have just written and everything that I am
teaching is from there and vice versa in the sense what is there is what I teach. So, we will
use that as a text book, and then there is a host of others reference books that I will point to
as and when needed essentially. So, already from these reference books, we have in some
sense finished with two of them which is Pamela McCorduck machines who think and
John Haugeland AI the very idea artificial intelligence is the very idea, but some of the
other books we will refer to as and when the time comes essentially.

So, we will stop here and next on Friday when we meet, we will have quality type shift and
start devising algorithms for simple search that just we mentioned essentially.