Toward HumanLevel Machine Intelligence

1
Toward Human-Level Machine Intelligence Lotfi
A. Zadeh Computer Science Division Department
of EECSUC Berkeley ICTAI06 Washington
D.C. November 14, 2006 URL http//www-bisc.cs.be
rkeley.edu URLhttp//www.cs.berkeley.edu/zadeh/
Email Zadeh_at_eecs.berkeley.edu
2
BACKDROP
3
PREAMBLE

• We are in the midst of what is popularly called
  the information revolutiona revolution which was
  born shortly after the end of World War II.
• As a student at MIT and later as an instructor at
  Columbia University, I witnessed the birth of
  this revolution and observed at close distance
  its progression and impact

4
THE BEGINNING OF THE AGE OF INFORMATION AND
CONTROL

• Three major events (ca.1946)heralded the
  beginning of the age of information and control
• Invention of the transistor
• Debut of cybernetics (Wiener)
• Debut of information theory (Shannon)
• I heard the first presentation by Shannon of his
  work at a meeting in New York, in 1946, and was
  deeply fascinated by his ideas. His lecture
  opened a new world

5
THE NEW WORLD

• The new world was the world of machine
  intelligence and automated reasoning
• It was widely believed that there were no limits
  to what machines could do
• The era of thinking machines has arrived
• Inspired by what I saw, heard and read, I wrote
  an article about thinking machines which was
  published in a student magazine

6
THINKING MACHINESA NEW FIELD IN ELECTRICAL
ENGINEERING
Lotfi A. Zadeh

• Psychologists Report Memory is Electrical,
  Electric Brain Able to Translate Foreign
  Languages is Being Built, Electronic Brain Does
  Research, Scientists Confer on Electronic
  Brain,these are some of the headlines that were
  carried in newspapers throughout the nation
  during the past year. What is behind these
  headlines? How will electronic brains or
  thinking machines affect our way of living?
  What is the role played by electrical engineers
  in the design of these devices? These are some of
  the questions that we shall try to answer in this
  article.

Columbia Engineering Quarterly, January 1950
7
CONTINUED

• Through their association with mathematicians,
  electrical engineers working on thinking machines
  have become familiar with such hitherto remote
  subjects as Boolean algebra, multivalued logic,
  and so forth. And it seems that the time is not
  far distant when taking a course in mathematical
  logic will be just as essential to a graduate
  student in electrical engineering as taking a
  course in complex variable is at the present
  time. Time marches on.

Columbia Engineering Quarterly, January 1950
8
EXAGGERATED EXPECTATIONS

• One of the headlines read "Electric Brain
  Capable of Translating Foreign Languages is Being
  Built." Considering that the only computers that
  were in existence at that time were relay
  computers, gives an idea of the depth of
  underestimation of the difficulty of building
  machines that can come close to human-level
  intelligence. Today, close to sixty years later,
  we have machine translation programs but their
  performance leaves much to be desired.

9
EXAGGERATED EXPECTATIONS

• On the occasion of inauguration of IBMs Mark 1
  relay computer in 1948, Howard Aiken, Director of
  Harvards Computation Laboratory, had this to
  say
• There is no problem in applied mathematics that
  this computer cannot solve
• In 1953, Burroghs Corporation started a project
  to design, manufacture and market a phonetic
  typewriter

10
EXAGGERATED EXPECTATIONS

• Like others, I had exaggerated expectations. Here
  is an example drawn from my 1950 paper.

11
A GLIMPSE INTO THE FUTURE (LAZ 1950)

• It is 1965. Three years ago for reasons of
  economy and efficiency the trustees of Columbia
  University have decided to disband the Office of
  University Admissions and to install in its place
  a thinking machine to be called the Electronic
  Director of Admissions.
• Installation was completed in the spring of 1964,
  and since then the Director has been functioning
  perfectly and has won unanimous acclaim from
  administration, faculty and student body alike

Columbia Engineering Quarterly, January 1950
12
ELECTRONIC DIRECTOR OF ADMISSIONS (1950)
probabilistic if-then rules record (a1, ,
an) accept if Prob Event (a1, , an) ? ? and
Condition D Event survive first
year Condition registration ? N If X is A and
Prob (Y is BX is A) is C and Condition is
D then Action is E
encoding
13
BRILLIANT SUCCESSES AND CONSPICUOUS FAILURES

• successes
• landing men on the moon
• GPS systems
• search engines
• bioinformatics
• failures
• summarization
• simultaneous translation
• automation of driving in city traffic
• tennis-playing robot

14
INFORMATION SYSTEMS / INTELLIGENT SYSTEMS
INFORMATION REVOLUTION
INTELLIGENT SYSTEMS REVOLUTION
INTERNET SMART CAMERAS WORLD WIDE WEB SMART
APPLIANCES WIRELESS TELEPHONY SMART
CARS FAX SMART ELEVATORS DIGITAL
LIBRARIES SMART ROBOTS DATA MINING INTELLIGEN
T MANUFACTURING INFORMATION RETRIEVAL EXPERT
SYSTEMS SMART SEARCH ENGINES SMART
QUALITY CONTROL Measure of intelligence
MIQ (Machine Intelligence Quotient)
15
MACHINE INTELLIGENT QUOTIENT (MIQ) (ZADEH 1993)

• Dimension of MIQ
• handwriting recognition
• speech recognition
• natural language understanding
• summarization
• disambiguation
• image understanding and pattern recognition
• diagnostics
• unstructured storage and retrieval of information
• execution of high level instructions (expressed
  in NL)
• learning
• reasoning
• planning
• problem solving
• decision making

16
INFORMATION /INTELLIGENT SYSTEMS (I/IS)
intelligent systems
intelligent information systems
information systems
Information/intelligent systems information
systems intelligent systems
intelligent/information systems

• information/intelligent systems are emerging as
  the primary component of the infrastructure of
  modern societies
• conception, design, construction and utilization
  of information/intelligent systems constitute the
  core of modern science and technology

17
ULTIMATE GOAL
Intelligent Decision Systems
SUBGOAL
Intelligent Information Systems
18
INFORMATION SYSTEM vs. INTELLIGENT INFORMATION
SYSTEM
SIEMENS FUZZY PARKING CONTROL (1996)
Parking garage
Parking Garage Marienplatz Parking Garage Stachus
FULL
FREE
19
INFORMATION/INTELLIGENT SYSTEMS (I/IST)

• Information/intelligent systems are becoming a
  reality
• But why did it take so long?
• The necessary technologies and methodologies were
  not in place
• Key technologies advanced computer
  hardware and software
• advanced sensor hardware and
  software
• Key methodology soft computing

20
TIMELINE OF GROWTH OF MIQ (LAZ)
MIQ
Human-level intelligence
1960
1980
2000
perception-based AI
logic-based AI (symbolic AI)
New AI symbolic logic probability theory
21
ACHIEVEMENT OF HUMAN-LEVEL MACHINE INTELLIGENCE

• Humans have a remarkable capability to perform a
  wide variety of physical and mental tasks without
  any measurements and any computations. Familiar
  examples are driving in city traffic
  summarizing a story and playing tennis.
• In performing such tasks humans employ
  perceptionsperceptions of distance, speed,
  direction, intent and other attributes of
  physical and mental objects.

22
CONTINUED

• Limitations of todays AI reflect the
  incapability of existing AI techniques to deal
  with perception-based information.
• What is widely unrecognized is that to achieve
  human-level machine intelligence it is necessary
  to endow AI with the capability to deal with
  perception-based information.

23
INFORMATION
INFORMATION
measurement-based numerical
perception-based linguistic

• it is 35 C
• Eva is 28
• probability is 0.8

• It is very warm
• Eva is young
• probability is high
• it is cloudy
• traffic is heavy
• it is hard to find parking near the campus

• measurement-based information may be viewed as
  special case of perception-based information

24
BIRTH OF AI

• Officially, AI was born in l956. At its birth
  there was widespread expectation that within a
  few years it will be possible to build machines
  that could think like humans. The AI pioneers,
  notably John McCarthy, Herbert Simon, Allen
  Newell and Neils Nilsson, but not Marvin Minsky,
  were firm believers in the ability of classical
  symbolic logic to lead to human-level machine
  intelligence.

25
CONTINUED

• I did not share that belief because the world of
  symbolic logic is an unreal world in which there
  is no imprecision, no uncertainty and no
  partiality of knowledge, truth and class
  membership. The world of symbolic logic is an
  idealized model of the real world. 

26
NEW AI

• For the AI establishment, anything that involved
  numerical computations was unwelcome. It took
  close to thirty years for probability theory to
  gain grudging acceptance. In large measure, it
  was the work of Judea Pearl that made probability
  theory respectable. Today, so-called "New AI" is
  probability-based. Indeed, Bayesianism has become
  as fashionable as symbolic logic was in its time.

27
CONTINUED

• Clearly, adding probability theory to the
  armamentarium of AI is a step in the right
  direction. But is it sufficient? In my view, the
  answer is No.

28
CONTINUED

• This view was advanced in my paper A New
  Direction in AIToward a Computational Theory of
  Perceptions, which was published in the AI
  Magazine, Vol. 22, No. 1, 73-84, 2001. The
  initial reviews of my paper were critical.
  Eventually, my paper was accepted for publication
  with its provocative title.
• My principal conclusion was that to achieve
  human-level machine intelligence it is necessary
  to have a capability to deal with
  perception-based information.

29
THE PROBLEM OF IMPRECISION

• Perceptions are intrinsically imprecise,
  reflecting the bounded ability of human sensory
  organs and ultimately the brain, to resolve
  detail and store information. It is this
  imprecision that places computation and reasoning
  with perceptions beyond the reach of symbolic
  logic and probability theory.

30
A NEW DIRECTION

• My AI Magazine paper outlined what may be called
  the computational theory of perceptions. The key
  idea in this theory is that of dealing with
  perceptions through their descriptions in a
  natural language. In other words, a perception is
  equated to its description, and computation with
  perceptions is reduced to computation with
  information which is described in natural
  language.

31
CONTINUED

• For this purpose, what is employed is fuzzy
  logica logic which mirrors the remarkable
  cability of human mind to reason with information
  which is imprecise, uncertain and partially true.
  

32
CONTINUED

• The distinguishing features of fuzzy logic are
  graduation and granulation. More specifically, in
  fuzzy logic everything is or is allowed to be
  graduated, that is, be a matter of degree, or
  equivalently, fuzzy. Furthermore, in fuzzy logic
  every variable is or is allowed to be granulated,
  with a granule being a clump of values which are
  drawn together by indistinguishability,
  similarity or proximity.

33
CONTINUED

• A granule may be interpreted as a representation
  of ones state of knowledge regarding the true
  value of the variable. As a simple example, Age
  is granulated when its granular values are
  assumed to be young, middle-aged and old.
  Graduation and granulation play essential roles
  in human cognition.

34
THE CONCEPT OF GRANULAR VALUE
A
granular value of X
a
singular value of X
universe of discourse

• singular X is a singleton
• granular X isr A granule
• a granule is defined by a generalized constraint
• example
• X unemployment
• a 7.3
• A high

35
GRANULATION OF A VARIABLE

• continuous quantized granulated
• Example Age

middle-aged
µ
µ
old
young
1
1
0
0
Age
quantized
Age
granulated
36
GRANULATION OF A FUNCTION
Y
f
0
Y
medium x large
f (fuzzy graph)
perception
f f
summary
if X is small then Y is small if X is
medium then Y is large if X is large then Y
is small
0
X
37
ANALOGY

• In bivalent logic, one writes and draws with a
  ballpoint pen
• In fuzzy logic, one writes and draws with a spray
  pen which has an adjustable and precisely defined
  spray pattern
• This simple analogy suggests many mathematical
  problems
• What is the maximum value of f?
• Precisiation/imprecisiation principle

Y
X
38
A BIT OF HISTORY

• A precursor of fuzzy logic was the theory of
  fuzzy sets. Informally, a fuzzy set is a class
  without unsharp boundaries, e.g., the class of
  beautiful women, the class of honest men, and a
  class of historical monument. My first paper on
  fuzzy sets appeared in 1965.

39
CONTINUED

• Its reception was mixed and mostly critical. My
  best friend, Richard Bellman, the father of
  dynamic programming was one of the few who
  welcomed the idea. This is what he wrote when I
  sent him the manuscript of my paper, Fuzzy Sets.

40
CONTINUED

• Journal of Mathematical Analysis and
  Applications
• Dear Lotfi
• I think that the paper is extremely interesting
  and I would like to publish it in JMAA, if
  agreeable to you. When I return, or while in
  Paris, I will write a companion paper on optimal
  decomposition of a set into subsets along the
  lines of our discussion.
• Cordially,
• Richard Bellman

41
CONTINUED

• Many others were not so kind. Here is a sample.
  Following the presentation of my first paper on
  the concept of a linguistic variable, Professor
  Rudolf Kalman, a brilliant scientist and a good
  friend of mine, had this to say

42
CONTINUED

• I would like to comment briefly on Professor
  Zadehs presentation. His proposals could be
  severely, ferociously, even brutally criticized
  from a technical point of view. This would be out
  of place here. But a blunt question remains Is
  Professor Zadeh presenting important ideas or is
  he indulging in wishful thinking?

43
CONTINUED

• No doubt Professor Zadehs enthusiasm for
  fuzziness has been reinforced by the prevailing
  climate in the U.S.one of unprecedented
  permissiveness. Fuzzification is a kind of
  scientific permissiveness it tends to result in
  socially appealing slogans unaccompanied by the
  discipline of hard scientific work and patient
  observation.

44
CONTINUED

• In a similar vein, my esteemed colleague
  Professor William Kahana man with a brilliant
  mindoffered this assessment in 1975.
• Fuzzy theory is wrong, wrong, and pernicious.
  says William Kahan, a professor of computer
  sciences and mathematics at Cal whose Evans Hall
  Office is a few doors from Zadehs. I can not
  think of any problem that could not be solved
  better by ordinary logic.

45
CONTINUED

• What Zadeh is saying is the same sort of things
  Technology got us into this mess and now it
  cant get us out. Kahan says. Well, technology
  did not get us into this mess. Greed and weakness
  and ambivalence got us into this mess. What we
  need is more logical thinking, not less. The
  danger of fuzzy theory is that it will encourage
  the sort of imprecise thinking that has brought
  us so much trouble.

46
FUZZY LOGICWHERE ARE WE TODAY? METRICS

•   PATENTS
• Number of fuzzy-logic-related patents applied for
  in Japan 17,740
• Number of fuzzy-logic-related patents issued in
  Japan  4,801
• Number of fuzzy-logic-related patents issued in
  the US around 1,700

47

• PUBLICATIONS
•  
• Count of papers containing the word fuzzy in
  title, as cited in INSPEC and MATH.SCI.NET
  databases. Compiled by Camille Wanat, Head,
  Engineering Library, UC Berkeley, August 25,
  2006.
•  
• Number of papers in INSPEC and MathSciNet which
  have "fuzzy" in title

MathSciNet - "fuzzy" in title 1970-1979
443 1980-1989 2,465 1990-1999
5,487 2000-present 5,714 Total 14,109
INSPEC - "fuzzy" in title 1970-1979
569 1980-1989 2,403 1990-1999
23,210 2000-present 21,919 Total 48,101
48

• JOURNALS (fuzzy or soft computing in
  title)
•  
• Fuzzy Sets and Systems
• IEEE Transactions on Fuzzy Systems
• Fuzzy Optimization and Decision Making
• Journal of Intelligent Fuzzy Systems
• Fuzzy Economic Review
• International Journal of Uncertainty, Fuzziness
  and Knowledge-Based Systems
• Journal of Japan Society for Fuzzy Theory and
  Systems
• International Journal of Fuzzy Systems
• International Review of Fuzzy Mathematics
• Soft Computing
• International Journal of Approximate
  Reasoning--Soft Computing in Recognition and
  Search
• Intelligent Automation and Soft Computing
• Journal of Multiple-Valued Logic and Soft
  Computing
• Mathware and Soft Computing
• Biomedical Soft Computing and Human Sciences
• Applied Soft Computing

49
SUMMATION

• Humans have a remarkable capability to reason and
  make decisions in an environment of imprecision,
  uncertainty and partiality of knowledge, truth
  and class membership. It is this capability that
  is needed to achieve human-level machine
  intelligence.
• Achievement of human-level machine intelligence
  is beyond the reach of existing AI techniques.
  New direction is needed. Computational theory of
  perceptions is a step in this direction.