Transcript PPT - 서울대 : Biointelligence lab

Level Headed
Drew Mcdermott
Yale University, USA
Artificial Intelligence 171(2007) 1183~1186
Sangyoon Yi
[email protected]
Bi. Lab.
Drew McDermott


B.S., M.S., Ph.D., Massachusetts Institute of Technology
Artificial Intelligence
Professor of Computer Science, Yale University, USA

planning algorithms

 calculate structures of actions for autonomous agents of various sorts

knowledge representation
 attempt to formalize what people know in a form usable by a computer

www.cs.yale.edu/~dvm/
2
© 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/
Overview

“I don’t believe that human-level intelligence is a well
defined goal.”

“Computer programs will eventually have many such
skills, but there will never be a time where their total
“equals” those of the average human”
3
© 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/
Human-level AI

How we will attain human-level artificial intelligence is so
ill-posed.
 Frankenstein’s monster (1820) --- Turing
 This sort of loss in exactly what we would hope for.


Ex) phrenology – superficial, unverifiable observations…
Intelligence is the ability to imagine
 There are as many different kinds of intelligences as there are
kinds of imagination

“Computer programs will eventually have many such
skills, but there will never be a time where their total
“equals” those of the average human”
© 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/
4
Behaviorism

Superhuman intelligence - Kurzweil
 We don’t know what we’re measuring.

Behaviorism didn’t even ask the right question.
 How do the rats compute the relevant properties of the current
situation?
5
© 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/
Two little wooden cubed problem

Display today’s date
 Digits (0, 1, 2, 3, 4, 5, 6, 7, 8, 9)
 Single-digit days are displayed as 0d.
0
1
0
2
1
2
6
© 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/
Two little wooden cubed problem

Display today’s date
 Digits (0, 1, 2, 3, 4, 5, 6, 7, 8, 9)
 Single-digit days are displayed as 0d.
0
1


0
2
1
2
6=9
{0, 1, 2, 3, 4, 5}, {0, 1, 2, 6(9), 7, 8}
7
© 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/
Two little wooden cubed problem

What would it take for a computer to solve this problem?
 Simplifying assumptions away from the full physical reality of
characters stamped on wood, and yet to make good guesses about
which of those assumptions to revoke when trouble arises.
0
1
0
2
1
2
 Digits as arbitrary tokens whose only property is to represent a
number between 0 and 9.
 Neglect the orientation of the visible face of a cube.
© 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/
8
Superhuman human


Seymour Papert
“Success of AI depends on creating programs that do as
well as the best humans at various tasks”
  So to achieve human-level intelligence, we might have to build
a person.
9
© 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/
Conversation

“Is it okay to use one face to represent both the 6 and the 9?!”

The biggest problem the cognitive-science community will
face is language.

If you listen to two or three people talking, and ask, what
exactly is each of them trying to do and to what degree are
they succeeding?

There is a lot going on besides exchanges of formally defined
Qs and As.
 Date-display puzzle.
© 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/
10
Conclusion

Machines will have goals and abilities very different from
ours, and talking with them will probably not feel like
talking to people.

Will the intelligent ones be able to explain what’s going on
and justify themselves to us?

Will they apologize and say that they can’t introspect about
the computational processes squeezing and tugging at our
lives any more deeply than we can introspect about
digestion?
11
© 2008, SNU Biointelligence Lab, http://bi.snu.ac.kr/