Intelligent Systems for Intelligent Learners

1
Intelligent Systems for Intelligent Learners

• Henry Tirri
• Complex Systems Computation Group
• Department of Computer Science
• University of Helsinki
• http//www.cs.Helsinki.fi/u/tirri/

2
Objective

• What can learning in machines teach us about
  learning in humans and in societies?

3
Intelligent systems
4
Convergence? ...
First Color TV Broadcast, 1953
HBO Launched, 1972
Interactive TV, 1990
Telephone, 1876
Early Wireless Phones, 1978
Handheld Portable Phones, 1990
WinTel
Pentium PC, 1993
Computer Modem 1957
First PC Altair, 1974
IBM PC, 1981
Apple Mac, 1984
Apple Powerbook, 1990
IBM Thinkpad, 1992
Apple Newton, 1993
Eniac, 1947
HP Palmtop, 1991
Red Herring, 10/99
5
Divergence and Competition
Atari Home Pong, 1972
Game Consoles Personal Digital Assistants Communic
ators Smart Telephones E-Toys (Furby, Aibo)
Pentium PC, 1993
Network Computer, 1996
Free PC, 1999
Sega Dreamcast, 1999
Internet-enabled Smart Phones, 1999
Pentium II PC, 1997
Apple iMac, 1998
Palm VII PDA, 1999
Red Herring, 10/99
6
(No Transcript)
7
Internet
8
What is the Internet?Its the TCP/IP Protocol
Stack
Applications

• Applications
• Web
• Email
• Video/Audio
• TCP/IP
• Access Technologies
• Ethernet (LAN)
• Wireless (LMDS, WLAN, Cellular)
• Cable
• ADSL
• Satellite

Middleware Services
Transport Services and Representation Standards
Narrow Waist
TCP/IP
Open Data Network Bearer Service
Network Technology Substrate
Access Technologies
9
Smart Dust (Kahn)

• Autonomous node incorporating sensing, computing,
  communications power source in 1 mm3 volume
  (current prototype 8 cm3)
• Dispersed through (outdoor) environment

10
Observations

• Many computing devices and access networks
• Currently a need to scale the computing to
  magnitudes reaching 109 units (more later)
• Robustness
• Computing with incomplete and uncertain
  information

11
What does not work?

• Learning by being told (memorizing facts)
• Learning crisp concepts (comprehensive
  definitions)
• Execution under a centralized control (master
  machine)
• Learning absolutes
• Deterministic approaches

What does work then?
12
Emergence

• Complex computing patterns arise from local
  interactions of the computing units

13
(No Transcript)
14
Advantages

• Any individual can be replaced
• Efficient communication between units
• Learning simpler models
• faster
• more reliable
• Universality

15
Local interaction
16
Intelligent Learners
17
Decentralized Mind

• Society of Mind (Minsky)
• Parallel Distributed Processing (McClelland,
  Rumelhart)
• Vehicles (Braitenberg)
• Subsumption (Brooks)

18
December 7, 1991

• Decentralization is a means to cope with
  complexity where centralized control is not
  feasible, for example
• learn to drive a car vs. traffic
• buying and selling stocks using thresholds vs.
  macroeconomy

19
Virtual Fish Tank (Boston)
20
Machine-learner interaction
21
Machine-learner interaction
22
Observations

• Due to emergence it is an inherently difficult
  task to identify what has been learned (in
  functional sense)
• Should everybody be taught (or learn) the same
  local interaction patterns?
• Testing learners in isolation does not reveal
  much about their emergent behavior
• How does one design curriculums for emergent
  behavior?
• If the emergent system is not working properly,
  how do you modify what needs to be learned?

23
Summary

• Decentralization and emergence are necessary
  elements for complex systems
• Intelligent machines force learners to be
  intelligent
• Evaluation of learning achievements for emergent
  behavior is problematic

24
A Vision

• If we want everything to stay as it is, it will
  be necessary for everything to change.
• Giuseppe Tomasi Di Lampedusa (1896-1957)