Testing our assumptions about Intelligence: Introduction to AI - March 3, 2003
Overview of lecture:

1. What is “intelligence”?
2. What are the different approaches to AI?
3. How can we know if any agent (human or computer) is intelligent?

Web resource: American Association of Artificial Intelligence’s pages
http://www.aaai.org/AITopics/html/overview.html

www.aaai.org/AITopics/html/sitemap.html
(sitemap)

www.aaai.org/AITopics/html/ethics.html
(required reading for March 24: article by Bill Joy + comments on his views)

www.aaai.org/AITopics/newstopics/ethics.html#dave
(for reports on stories in news about AI and ethical issues)

1. What is "intelligence"?

“an open collection of attributes.” We expect an intelligent agent to be able to:

--> understand and use language and related symbolic tools

[lecture on Mar.10 on natural language processing and the inherent difficulties getting a machine to understand what we mean when we say/write something.]

use language to refer to concrete things == like a “rose”
use language to refer to abstract things == like “love”
we know the difference between concrete and abstract things

--> be original, synthesize new concepts and ideas, and acquire and employ analogies

e.g., “my love is like a red, red rose”

by using a simile we compare two dissimilar things to come up with a new idea
poet, Robert Burns came up with an original idea—juxtaposing two disparate things.

-- > Draw distinctions between situations/things despite similarities

e.g, two pens/ two stories

--> generalize (find a common underlying pattern in superficially distinct situations)

e.g., cake, candles, presents = birthday party

--> understand, including the ability to make sense out of ambiguous or contradictory information

e.g., “I saw the man with the binoculars”

--> plan and predict the consequences of contemplated actions

e.g.,
IF I don’t follow instructions in my second term assignments,
THEN I won ’t get a good mark.

--> know the limits of its knowledge and abilities

e.g., “I know I can’t draw.”

--> learn (have ability to acquire new knowledge)

e.g., learn that the word “femme” is the French word for “woman”

--> solve problems, including the ability to break complex problems into simpler parts

[see lecture on Mar.12 – the classic example is the Tower of Hanoi – see kit]

--> have mental attitudes (beliefs, desires, and intentions)

e.g., “I believe that men and women are equal.”

--> perceive a model of the external world

e.g., see the interrelatedness of all things

[see lecture Mar. 17 -- an issue with expert systems which have knowledge of only one segment of knowledge]

N.B.: not all humans exhibit all these attributes, and some other mammals exhibit some of these attributes (to a limited degree)…

As well, there are some attributes, related to, but distinct from, intelligence which we need to consider:

--> awareness (consciousness)
--> aesthetic appreciation
--> emotion
--> sensory acuteness
-- > muscular coordination

(from Fischler and Firschein, Intelligence: The eye, the brain and the computer, 1987)

2. Different Approaches to AI:

Early AI researchers took one aspect of “intelligence” --problem solving and then one type of problem--games (like checkers, and chess)--to get the computer to exhibit intelligence.

Then other researchers started working on expert systems...

BUT both of these areas focus on a specific / limited domain of knowledge.

We could take a more functional/holistic approach to define intelligence: "Intelligence is precisely our ability to cope with the world." (Moody, 1993, p. 127)

--> coping means solving the thousands of "problems" we encounter every day. Such as?

- brushing our teeth
- getting on the right bus
- talking to friends
- (knowing who one's friends are!)
- understanding a joke
- writing an exam
- buying a coffee
- playing a game (say chess)
- arguing with someone about whether there should be a war against Iraq

--> these all represent certain goals--the problem is to figure out how to reach the goal (what kinds of knowledge does it take to solve the various kinds of problems?):

Of these 2 problems which is it proving hardest for the computer to manage?
playing chess, or getting on the right bus?

3. How can we know if any agent (human or computer) is intelligent?

Q: What is the Turing Test?

When a computer “passes” as a human when communicating with a human through a keyboard.

Turing (1950) set out these original questions with what he hoped the computer could reply:

Q: Please write me a sonnet on the subject of the forth bridge
A: Count me out on this one. I never could write poetry.
Q: Add 34957+ 70764
A: (after 30 seconds) 105621 (which is wrong)

see: http://www.macrovu.com/CCTMap2.html

Q: Why adopt the imitation principle as a criterion of a computer's intelligence?

A: Since there is no way of telling what other people are 'thinking' except by a comparison with oneself, why treat computers any differently?

Q: What are the implications of this measurement of intelligence?

A: That intelligence is assessed by "product" not process.
AND
That the ability to successfully communicate with a person (or a machine) is a better indication of intelligence than any other attribute accessible to measurement.

Q: But is "imitation" enough?

A: yes, according to Turing: a perfect simulation of thinking IS thinking.
no, says Searle. "simulation is not duplication. Symbols don't mean anything to the computer." See his example of the Chinese room)

(see http://psych.utoonto.ca/%7Ereingold/courses/ai/turing.html for analysis of the Chinese Room)

Q: Why does the computer HAVE TO pass the Turing Test?

A1: this shouldn't be the goal of AI community: the goal should be AI systems that can complement--not mimic--human mental activity.

A2: ....one might well contend that machines can't think, FOR THEY DO MUCH BETTER THAN THAT...We could forever deny that a machine could "think through" a math problem and still claim that in many respects the achievement of machines was on a higher level than that attained by humans since machines can almost simultaneously and infallibly produce accurate and sometimes original answers to many complex math problems...they do not NEED TO THINK OUT the answers." 9K Gunderson, "The Imitation Game")