AI Intro
AI Intro
GENESIS
• An old dream
– To produce an intelligent replica of one’s own
– From Merry Shelley’s Frankestine to Stephen
Spilberg’s summer movie AI the dream has cast
its shadow
Some Important Observations
• Intelligence is the key to progress of
mankind.
• Power of intelligence lies in the knowledge
• Knowledge is a product of human
cognition/thought processes
Philosophers’ Endeavour
• For more than 2000 years, philosophers are
trying to reveal the mysteries of human
cognition
A Valuable Outcome
• The Hypothesis
– Human mind is principally governed by
reasoning with knowledge contained in it.
The Idea
• Intelligence can be capture by doing the
following.
– Defining Knowledge
– Mastering the mechanism of reasoning
analytically
Materialization of the Idea
• Past efforts on how to capture intelligence
have led to the development of the subject
of “Artificial Intelligence”
Three Popular Examples
• Deep Blue
• Humanoid Robots
Three Popular Examples
• Deep Blue
– It is an IBM made chess playing system, which defeated
the world champion Gary Kasparov in 1997
• Humanoid Robots
– They are often found dancing, singing and even playing
soccer.
• Every year, a soccer tournament for robots are organized in
Japan under the auspices of RoboCup
• Robocup’s ultimate goal is to beat humans in soccer by 2050
• Many AI professionals consider RoboCup small league
competition as an ideal platform for testing distributed AI
techniques
Three Popular Examples
contd..
• Autonomous Vehicles
– A system called Autonomous Land Vehicles in a
Neural Network (ALVINN) was developed in CMU in
early 1990’s
– It is capable of steering automobiles autonomously
– It successfully drove a van from Washington DC to San
Diego.
– During this journey, it has covered a distance of all but
52 of 2,849 miles at an average speed of 63 mph over
day and night, under all weather conditions.
Artificial Intelligence
• It is a subject that attempts to understand the
mechanism of intelligence and strives for
computational modeling of intelligent behaviour
• Intelligent behaviour involves the following
– Perception
– Reasoning
– Learning
– Communicating
– Acting in a complex environment
A Grand Question
• Whether artificial consciousness can really emerge in
advanced computers?
– The belief that all important aspects of human cognition
may in principle be captured computationally forms the
very basis of AI
– The belief has made it possible for AI people to focus on
the abstract causal structure that the mind embodies (at an
appropriate level) instead of the biological substrate of
the mind, i.e., neurons and biological wetware in general
– According to this belief, formation of best possible minds,
i.e., human minds, on a carbon based substrate is possible
due to constraints of evolution and possibly historical
accidents, rather than to absolute metaphysical necessity
A Grand Question
• The view is made in two steps
Step 1: Church-Turing Thesis supports the idea that any
abstract causal structure can be modeled computationally
Step 2: Since publication of an article by Alan Turing,
supporting the idea for making the computer intelligent, in
1950, AI people started to believe the possibility of viewing
human minds through abstract causal structures
• Viewing human minds through abstract causal
structures has become known as functationalism
– After Turing (1950), it was explicitly articulated in various
forms by Putnam (1960), Armstrong (1970) and Lewis
(1970)
A Grand Question
• Viewing computational modeling as essential for
capturing essence of cognition is called
computationalism
• The Church-Turing Thesis has bridged the gap
between functionalism and computationalism
• Some believe in fnctionalism but not in
computationalism, but there are many people who
believe that two are synonymous
– Penrose (1989), who seems to be a functionalist without
being a computationalist, believes that certain causal
properties of the mind cannot be captured algorithmically
A Grand Question
• Philosopher John Searle has termed the
doctrine of computationalism as ‘strong AI’
– In other words, the belief that a computer
system equipped with sufficient processing
power and sufficiently powerful artificial
intelligence would actually be capable of
having mental states in the humans is called
strong AI.
A Grand Question
• In contrast to ‘weak AI’, holding that the computer
can be used as a powerful tool for understanding
the mind, strong AI goes far beyond holding that
the appropriately programmed computer really has
(is) a mind
– Weak methods in AI attempt for a general purpose
search mechanism trying to string together elementary
reasoning steps to find complete solutions.
– Such methods that had arisen in the first decade of AI
research use weak information about the domain
A Grand Question
• The concept of strong AI is severely criticized by Searle in
“Minds, Brains, and Programs” (1980)
• He argues through his now famous Chinese Room argument
that a computer can never recreate such vital properties of
human mentality as internationality, subjectivity, and
understanding
• In his view, strong AI is thoroughly implausible.
• Later David J. Chalmers in “Subsymbolic Computation and
Chinese Room” (1992) observes that Searle’s arguments
provide at least a plausible criticism of the Symbolic
approach to AI, but connectionist models are less vulnerable
to his arguments
– He advised Searle to desist from his blanket criticism of the entire
AI research endeavor and to think over the potential of
Connectionist approach to AI.
Major Goals of AI
• The Long Term Goal
– To develop a general information processing theory of
intelligence
• The Short Term Goal
– To realize the intelligent abilities, such as the
following, through computers
• Theorem proving
• Program synthesis
• Expert decision making
• Game playing
• Language comprehension
• Commonsense reasoning
Dartmouth College Workshop
• It was an international workshop held in 1956
• How every aspect of learning or any other feature
of intelligence can be realized with the computer
was the theme of the workshop.
• To express the theme, John McCarthy coined the
term “Artificial Intelligence”
• The event was a land mark for emergence of AI as
a subject of study first time in an international
forum.
Dartmouth College Workshop
• Some earlier developments that necessitate
holding of one such workshop are the
following
– Commercial use of stored program computers
– Alan Turing’s arguments supporting the idea of
programming a computer to exhibit intelligent
behaviour
– The work of Claude Shannon on the possibility
of computer chess
Dartmouth College Workshop
• Development of the first neural network
computer by Marvin Minsky and Dean
Edmonds
• The progress of the work by Alan Newell
and Herbert Simon on automatic theorem
proving
• Noam Chomsky’s work on the theory of
generative grammars
Turing’s Article
• The hope that machines would eventually
compete with human was expressed in the
article in 1950
• Turing argued for the idea of programming
a computer to exhibit intelligence
• He proposed a test to judge the intelligence
of a machine
Turing Test
• A human interrogator
sitting in a room asks
a series of questions
• The questions are A computer
randomly directed to
either a computer or a
person in a separate A human
interrogator
room.
A person
Turing Test
• If the interrogator is
not able to distinguish
which of the two has
responded to his A computer
queries then the
computer will be
called intelligent A human
interrogator
A person
Deep Blue and Turing Test
• Deep Blue seems to have passed Turing
Test
– Kasparov could never wholeheartedly accept
that he played with a computer itself
Secrets of Deep Blue’s Feats
‘ABCD is a rectangle’
– All real world objects
and concepts are RECTANGLE(ABCD)
represented with
symbols Concept Object
‘Socrates is a man’
MAN(SOCRATES)
Concept Object
Symbol Processing Approach
• Symbols may be
combined to form
complex structured ‘All rectangles are parallelograms’
expressions
(x) RECTANGLE(x) PARALLELOGRAM (x)
• Complex expressions
refer to more complex ‘Man is mortal’
entities or knowledge
(x) MAN(X) => MORTAL(x)
• All kinds of data that can
improve efficiency and
effectiveness of a
problem solver,such as
facts, beliefs, and rules of
thumb(heuristics), are
considered as knowledge
Symbol Processing Approach
• Under this approach, computation is performed directly upon
symbols to replicate intelligent behaviour, i.e., to generate
inferences
• Inferencing is realized with a class of specially designed programs
– The programs can compute new symbols from the existing ones,
e.g.,
MAN(SOCRATES)
(x)MAN(x) => MORTAL(x)
MORTAL(SOCRATES) {SOCRATES/x}
RECTANGLE(ABCD)
(x) RECTANGLE(x) PARALLELOGRAM (x)
3. CLASS-MATE(RAM,RAHIM)
4. CLASS-MATE(RAHIM,ALI)
A Complex Chain of Inferences: An Example
contd..
• Proof:
CLASS-MATE(RAHIM,ALI)
(u) (v)CLASS-MATE(u,v)FRIEND(u,v)
FRIEND(RAHIM,ALI) {RAHIM/u, ALI/v}
CLASS-MATE(RAM,RAHIM)
(x) (y) (z){[CLASS-MATE(x,y)ΛFRIEND(y,z)]FRIEND(x,z)}
FRIEND(RAM,ALI) {RAM/x,RAHIM/y,ALI/z}
The Physical Symbol System Hypothesis
Training
data
Knowledge
Engineer
Domain Expert
Avoid
Module
Final output
1 if >0
y=
0 Otherwise
– Positive weights correspond to excitatory synapse
– Negative weights model inhibitory ones
– Threshold function approximates the activity in soma
– Wires and interconnections model axons and dendrites
– The simplifying assumptions for the MCP model do not
reflect the true behavior of biological neurons
x1
w1
1 y
x2 w2
.
. u
.
wn
xn
X2
x2=(-w1/w2)x1+(-w0/w2)
(0,0) X1
Hyper plane
A straight line dividing the input plane
Generation of a Hyper Plane by the MCP Neuron
f2 f2
44444444
444444 444
4 44 4 44 44
44444444
44444 4
4 4
9999999999
9 9 9 99 9 9 9 9 9 9
99 99 9999 99
99 99 9 9999
9 9 9 99
9
f1
f1
f1
Two linearly separable
Two nonlinearly separable
pattern classes
pattern classes
Recognition of AND Inputs using an MCP Neuron
0 1 0
Separating line
1 0 0 x2=(-w1/w2)*x1+
(-w0/w2)
1 1 1
(0,0) 1 1.5 x1
0 1.5 Σwixi
1 2
The threshold function
Recognition of AND Inputs using an MCP Neuron
x1 x2 w1x1+w2x2 Σwixi + w0 y
0 0 0 -1.5 0
0 1 1 -0.5 0
1 0 1 -0.5 0
1 1 2 0.5 1
Recognition of EXOR Inputs using an MCP Neuron
x2
A two input EXOR function
c0 c1
X1 ^ X2 (0,1) (1,1)
x1 x2
0 0 0
0 1 1
1 0 1
1 1 0 (0,0) (1,0)
x1
X2
L2
(0,0) 1 X1
L1
Two straight lines are required for proper class separation
The architecture of a neural network that can implement the
EXOR logic
X
2
1
L2
0 X
L
x0 , 1
0 1 X
2
1
L
x1 0 1 2X
, L 1
0
1
X
2 1
x2
0 X Recognition of EXOR
L1 1
,
0
1
inputs using MCP
neurons
Pattern Recognition as a Problem of Machine
Learning