What Is Artificial Intelligence

1
What Is Artificial Intelligence
Overview 1. Objectives of AI 2.
Definitions of AI 3. A brief history of AI
4. Major challenges 5. Schools of thought
2
Objectives of Artificial Intelligence
The general goal of AI is to study intelligent
behavior in humans, as well as in animals
and machines. For example, 1. Perception
(vision, auditory, olfactory, tactile, )
2. Reasoning (think, plan, query, )
3. Acting (navigation, maneuver,
) 4. Learning (adaptation,
discovery, ) 5. Consciousness (sense of
self)
3
Objectives (cont.)

• 1. Scientific goal
• To understand the functions of biological
  systems and fundamentals
• of intelligence.
• (for example, how does the brain work?
• Limits and
  bounds what tasks are achievable and what are
  impossible?
• what are the
  optimal way to perform a task? )
• Engineering goal
• To design intelligent machines (artifacts,
  programs, autonomous robots) that behave
  intelligently in real environment.

AI is mostly concerned with the engineering
aspect.
4
What Does It Mean By Intelligence ?
Intelligence is an ill-defined, subjective, and
time-variant concept
For example, commercial products intelligent
keys, intelligent washer, are not
part of the AI study.
AI is concerned with programming computers
(robots) to perform tasks that are presently
done better by humans --- Minsky
It seems true that once computers outperform
humans in an area, then research on that problem
will gradually spin off from AI.
5
Three Levels of Study
Intelligence can be studied at many levels.
For example, David Marr proposed a
three level concept in studying vision

• Representation --- mathematical formulations of
  the problem
• Computation --- algorithms for solving the
  problem
• 3. Implementation --- machines for carrying
  out the algorithm

For example, stereo vision (computing depth from
2 images) can be formulated as Finding the
maximum value of a function. Then there could
be many algorithms that can solve for this
problem. Given an algorithm, it can be
implemented in a PC, a parallel computer, a
VLSI chip, or other biologically motivated
devices.
6
The Physical Symbol System Hypothesis
There was heated debates on whether human-level
intelligence can exist independent of the
biologic system. The physical symbol system
hypothesis by (Newell and Simon 1976) argued that
it can.
A physical system that can manipulate
symbols, like a computer, then It has the
necessary and sufficient means for general
intelligent behavior. It does not matter what
the physical symbol system is made of, proteins
or silicon.
This also implies that human intelligence may not
be the optimum. For example, a computer can
outperform humans in playing chess. Obviously
computers use totally different algorithms and
implementations. Computers dont have the
emotional problems (fear, panic) that human
players have.
7
The Symbol Grounding Problem
One argument against the physical symbol system
hypothesis is the symbol grounding problem
How are the symbols (or mental statuses)
grounded in its sensory experience?
For example, it will be extremely hard to explain
to a blind person what a red color is, or to
explain what a Wendy Hamburg to someone who
never ate a sandwich before. Thus, How can we
build a machine that can exhibit human level
intelligence if it cannot share the sensory
experience?
8
The Subsymbolic Processing Approaches
People taking this view argue that much of the
human intelligence and behavior is the result of
the subsymbolic processing --- the processing
of sensory signals, not symbols. For example,
vision, speech, in contrast to playing chess.
9
Behavior and Environment
Other people argued that what important for
intelligence is not reasoning but behavior of a
system. Intelligence is reflected in the
emergent behavior between an agent and its
dynamic environment. Thinking may not be
necessary, or thinking is our fiction.
For example, when we touch a hot object, we take
back our hands immediately to avoid burning. This
process does not involve the brain function.
10
Views of AI
There are four different views for what an AI
system should be
Human vs. Rationality
Thinking vs. Behavior
11
Views of AI
Weak AI Machine can be made to act as if they
are intelligent.
Strong AI Machines that act intelligently
have real conscious mind (can think).
12
AI Successful Applications

• Playing chess, (defeated world champion)
• Expert system, (trouble shooting, )
• Voice recognition, (booking tickets)
• Vision system, (face recognition, 3D
  reconstruction, monitoring)
• Autonomous Vehicle

13
AI Challenges
The vodka is good but the meat is rotten
--- machine
translation from Russian The spirit is willing
but the fresh is weak
Difficulties include 1. Hard to represent
domain knowledge and common sense knowledge 2.
Complexity (intractability) of the problems 3.
The embodiment problem.
14
The Embodiment Problem
Full scale human level intelligence may be too
complex, at least too dependent on the precise
physiology of humans, to exist apart from its
embodiment in humans situated in their
environment.
This problem is similar to simulating the climate
conditions on earth surface or simulating the
stock market. There are simply Too many details
which often have catastrophic influence.
15
Summary of History

• Gestation of AI (1943---1956)
• Early enthusiasm and great expectation
  (1952---1969)
• A dose of reality (1966---1974)
• Expert systems (1969---1979)
• AI becomes an industry (1980---1988)
• Neural network becomes hot (1986---1995)
• 7. Statistical approaches (1987 --- present)

16
AI Three Schools of Thought

• Logics representation and reasoning
• --- propositional and predicate calculus
• --- search algorithms
• --- knowledge reasoning systems
• Statistical modeling and inference
• --- calculus of probability
• --- Stochastic inference
• --- Uncertainty reasoning
• Connectionism
• --- Biologically motivated systems, such
  as various neural nets
• --- Learning and adaptation.