€ Artificial intelligence: devoted to producing systems that perform tasks
that would require intelligence if done by humans.

Common point of view‹using human performance as a benchmark.

€ Begun in a 1950 paper by Alan Turing, in which he proposed a test for an
intelligent program.

Turing claimed that if a program acted like a human, we could claim it
was intelligent.

Metaphor: The HAL 9000

We will use Arthur C. Clarke's fictional computer as the theme of this
module.

€ Unquestionably, HAL is an intelligent machine. "His" abilities include

Human-quality natural-language processing

Sophisticated visual processing

A large and richly interconnected real-world knowledge base

The ability to plan and solve problems

The ability to learn by assimilating new information

People and Machines

We have only the most rudimentary knowledge of what human intelligence is.

Thinking effortlessly:

€ The hardest things to simulate on a computer are precisely the things we
do without conscious thought.

Thinking deeply:

€ One thing humans are very good at is dealing with analogy and metaphor.

Thinking hard:

€ Compared to other areas, it is relatively easy to write a program that
exhibits goal-directed behavior and the ability to form logical inferences.

€ Somewhat ironically, this supposedly "highest" level of human behavior is
easiest to mimic.

Brains versus Computers

€ Consider the hardware differences between brains and computers.

(1) Storage: Estimates are that human brains can store the equivalent
of 50 trillion bits of information. The largest computers today can
store about a trillion bits.

Advantage: 50-1 for people

(2) Complexity: Brains are massively interconnected. Parallel
computers today have no more than 1,000 independent processors, while
brains have many more.

Advantage: 1,000-1 for people

(3) Speed: Neural signals are transmitted by chemical changes, at about
a thousand feet per second. Electrons move through circuits at nearly
a billion feet per second.

Advantage: 1,000,000-1 for computers

€ Moral: Brains have the edge . . . for now.

Artificial Skills

Language Processing

€ HAL is as skilled a language user as the humans he speaks to. AI
research has a way to go here.

€ Language-processing programs have grown to include:

(1) Parsing

Example: "Colorless green ideas sleep furiously."

(2) Semantic analysis

Example: "Ron lies asleep in his bed." Is he telling the truth?

(3) Use of context to determine the sense of an utterance.

Example: "The clams are ready to eat." Are they hungry or are we?

(4) Models of informal rules of conversation.

Example: "Do you know the time?" We expect more than a yes or no
answer.

(5) An extensive knowledge of the real world.

Example: "Sally was fed up. She got up angrily from her table at the
restaurant and left just enough to cover the check. The waitress
sneered at her as she walked out." Why?

Knowledge Processing

€ HAL "knows" a great deal about the world.

€ How do we represent real-world information in the memory of a computer?

€ Early game-playing programs maintained a "game tree" of the possible
positions and their relations:

€ We might represent linguistic data in a semantic net:

€ An expert system contains representations of real-world knowledge, often
in the form of inference productions such as:

IF the engine won't turn over and the lights won't turn on, THEN check
the battery.

IF checking the battery and the connectors are corroded, THEN first
clean the connectors and then try to start the car.

IF the engine won't turn over and the lights work, THEN check the
starter solenoid.

Visual Processing

€ HAL also possesses an uncanny skill at visual processing, including the
ability to read lips.

€ Visual processing is a difficult task to program: we don't even have a
good idea about how people process visual information.

€ We seem to rely on many visual cues (about distances, sizes of familiar
objects, shadows and filling in hidden edges) to help us interpret visual
images.

OCR: A Simple Visual Processing Task

€ Optical character recognition (OCR) is the problem of designing a machine
and a program that can scan a printed page and recognize the letters on the
page.

Two main approaches:

(1) Matrix matching, in which the scanned character is matched
against a stored image

(2) Pattern extraction, in which the program recognizes, say, the
essential "a-ness" of a letter.

€ Matrix matching is fairly easy, but it requires that we know the typeface
the program is reading.

€ Pattern extraction is more difficult: What geometric features commonly
define the letter g below?

age age

Helvetica Palatino

age age

Courier Zapf Chancery

€ Context would be a big help here (the "Wheel of Fortune" model):

Wn nan rnnd tnns, nven wntn mnnt on thn lentnns nnssing.

Learning

€ HAL shows a humanlike ability to adapt to new situations and modify his
knowledge and behavior accordingly.

€ A program that can learn relieves us of the burden of coding in all
necessary knowledge at the beginning.

Spiders have many built-in abilities, humans have to learn. Shall we build
spiders or babies?

Artificial Attributes

€ HAL exhibits a range of human attributes: pride, artistic appreciation,
frustration, embarrassment, self-awareness, and fear.

€ It is not inconceivable that such behavior might arise in the future.

€ We should be concerned.

Pamela McCorduck poses this problem in provocative form in Machines
Who Think:

Can a made-up mind be moral?