Title: What is Artificial Intelligence Today? Zenon Pylyshyn, Rutgers Center for Cognitive Science http://ruccs.rutgers.edu/faculty/pylyshyn.html
1What is Artificial Intelligence Today?Zenon
Pylyshyn, Rutgers Center for Cognitive
Sciencehttp//ruccs.rutgers.edu/faculty/pylyshyn.
html
CSCSI/SCEIO 15th Annual Meeting, Calgary, May
27, 2002
- Is it an engineering profession, like EE?
- Is it a branch of Computer Science, like software
engineering? - Is it a nascent science, like Genomics?
- Is it a methodology, like Statistics?
- Is it the home of lost causes, like
Parapsychology?
2What is Artificial Intelligence Today?Zenon
Pylyshyn, Rutgers Center for Cognitive
Sciencehttp//ruccs.rutgers.edu/faculty/pylyshyn.
html
CSCSI/SCEIO 15th Annual Meeting, Calgary, May
27, 2002
- Is it an engineering profession, like EE?
- Is it a branch of Computer Science, like software
engineering? - Is it a nascent science, like Genomics?
- Is it a methodology, like Statistics?
- Is it the home of lost causes, like
Parapsychology? - Short answer All of the above.
3Where is Artificial Intelligence Today?a highly
personal view
CSCSI/SCEIO 15th Annual Meeting, Calgary, May
27, 2002
Zenon Pylyshyn, Rutgers Center for Cognitive
Sciencehttp//ruccs.rutgers.edu/faculty/pylyshyn.
html
4Where has Artificial Intelligence been and where
is it going?a highly personal view
Outline of talk
- A lighthearted look at what has A.I. been doing
these past 45 years in order to survive. - A more serious look at what I believe we have
learned from work in A.I. and Cognitive Science. - Some speculations on whether current trends
provide any grounds for optimism for the future
of A.I.
5Where has A.I. been these past 45 years?
- It has been battling the forces of darkness and
has survived
In the meantime it has split into many partsand
has assumed many different disguises.
It has been 45 years since the 1957 Dartmouth
conference.
6A.I. Has Survived Many Threats
- It has survived
- The illusions of innocence
- The foolishness of futurism
- The perils of popularity
- The dazzle of demos
- The shameless pursuit of short-term goals
- The corruptions of capitalism
7It has survived the illusions of innocence
- The age of innocence Cybernetics, Weiner,
Turing, Newell Simon and the promise of
mechanized intelligence (Thats what drew me and
many others to it)
Photographs courtesy of Lord of the Rings
8It has survived foolhardy futurism
- Its hard to predict, especially the future
(Japanese saying)
9Examples of foolhardy futurism
- A computer will be chess champion and will prove
a significant mathematical theorem within ten
years (Simon, 1957) - Within a generation the problem of creating
artificial intelligence will be substantially
solved (Minsky, 1967). - I believe that robots with human intelligence
will be common within 50 years (Hans Moravec,
1988) - by the year 2020, neural net computers will
have doubled about 23 times ... resulting in a
speed of about 20 million billion neural
connection calculations per second, which is
equal to the human brain. The emergence of
machines that exceed human intelligence in all
its broad diversity is inevitable. (Ray
Kurzweil, 1999) - "Guys like Kurzweil and Moravec ... somehow think
that they can take Moore's law and project things
into the future and say that, by 2020, you'll
have human-level intelligence. I don't believe
that at all. I think new ideas are required."
(John McCarthy, 2001).
10It has survived the perils of popularity
- For many years (in the late 1970s and 1980s)
A.I. was the sexiest topic in computer science
and psychology (which may be why many of us are
here!)
11And that leads to certain distractionsthe
perils of popularity
12It has survived the dazzle of demos
- The Gee Whiz School of Science
- Toy systems (GPS, Eliza, Shrdlu, Shaky,
Planner/Strips) - Performance systems (Dendral, XCON, Prospector,
DART)
Lothlorien
- But how do you know when you should be impressed
by a demo?
13- Its not easy to tellwhether you should be
impressed by a demo. - How impressive the demo is depends on what else
you assume the system can do. - How impressive the demo is may depend on what
tools are available.
14It has survived the shameless pursuit of short
term goals
- Artificial intelligence used to mean robots that
think like people now it means software for
rejecting junk e-mail. Low expectations could
yield better applications, sooner." Doug Lenat.
MIT Tech Review, 2001
- So, whats wrong with that?
15- Short-term solutions rarely scale up.
- It is possible to attain spectacular short-term
achievements without any fundamental new
understanding of A.I. problems (e.g., Deep Blue). - Short term criteria dont tell you when your
assumptions are hopelessly wrong and you should
start over. - Short term problem-solving detracts from the
bigger AI goal.
16A.I. has even survived the corruption of
capitalism
- Teknowledge, Tekmoney, CAIP, Cognicom, Gomi AI, ..
- What we tell venture capitalists vs. what we
really believe
17Some of the commercial claims make one wonder who
our colleagues are working for!
Sometimes I wonder if we havet carried
ecumenism a bit too far
18AIs dilemma
- In the end, we in AI get high marks on the
applications side, to which we have been pushed
by sponsor pressure and seduced the lure of
billions. And certainly, it is hard to quarrel
with the good done. But from the perspective of
answering big questions and realizing big dreams,
the disappearance of AI into computer science
retards progress by shifting the way AI people
are judged and rewarded. - Patrick Winston, keynote address to the American
Association for Artificial Intelligence, July 20,
1999
19Is it time to return to the original goal?Can we
keep our eye on the prize
- Is it time for looking back on what weve learned
and for refocusing on the scientific puzzles
From Nils Nilssons 1995 AAAI paper title
20What have we learned in the past 45 years? Some
current criticisms of AI not my own views
Lessons I have learned from Cognitive Science
- Representing and manipulating knowledge is not
everything KR has been overemphasized in AI - Logical formalisms are hopeless for representing
knowledge - Human cognition has a lot in common with that of
other organisms, so AI should start by simulating
the simple ones first (e.g., insects), and
letting complex ones evolve - Intelligence is distributed among minds, bodies,
and environments, and A.I. has not recognized
these enough in its pursuit of KR.
21There is more to intelligence than processing
knowledge.
- No matter how little of the system deals with
knowledge, it will always remain the core of
intelligence because Intelligent behavior is
essentially knowledge-dependent or cognitively
penetrable. This follows from the following
important facts about (human) cognition - The equivalence classes of events under which
intelligent agents behavior is predictable are
semantically defined inputs with the same
meaning have the same effect. - Its how the world is represented, rather than
how it actually is, that determines behaviour. - There is no rule or principal of intelligent
behavior that cannot be overturned by changing a
persons goals and beliefs. In other words, some
part of every intelligent behaviour is cognitive
penetrable.
For details, see my Computation and Cognition,
MIT Press, 1984
22What about the Intelligence in the I/O?
- Its true that a lot of intelligence occurs at
the I/O, where the processes are encapsulated in
modules. - Its also true that the most progress (in AI and
CogSci) has been made in understanding the
modular systems of cognition, such as language,
vision, and motor control. - Exploring I/O systems and new computational
architectures is important, but the basic engines
of intelligence are semantically-driven processes
like inference. The intentionality of the mental
is what distinguishes human-level intelligence
from merely clever or impressive behavior. - For more on this, see Pylyshyn, Z. (1984).
Computation and Cognition. MIT Press.
23(2) Logical formalisms are hopelessly inadequate
for representing knowledge
- As Winston Churchill said about democracy It is
the worst form of government, except for all the
rest. - So also logic is the worst form of knowledge
representation, except for all the rest.
24- Much has been said about why logic is
inadequate e.g., it has trouble representing
knowledge that is procedural, quantitative,
pictorial (or sensory), higher-order (meta
knowledge) and indexical although excellent
work is being done on all of these problems. A
reasonable assumption is that there may have to
be more than one form of representation.
Unfortunately, none of the non-symbolic ones
proposed so far (e.g., neural nets, images) are
adequate for reasoning. - Logic or Logical Calculus is just another name
for symbol system except that logic usually
provides a formal semantics (i.e., a theory of
how the meaning of expressions is composed from
the meanings of its parts) and a system of
inference (at least for deductive inference, if
not for inductive, abductive and practical
reasoning). Nothing like that is available for
any of the other forms of representation.
25So logic is a special form of symbol system But
why do we need symbol systems at all?
- What makes it possible to represent and process
knowledge in a physical system, and to connect
this knowledge with both perception and action,
is what Newell Simon called a Physical Symbol
System. This is perhaps the single most important
idea of the 20th century because it made possible
computing.
- But in order to be adequate for encoding
human-level knowledge, physical symbol systems
must meet some stringent conditions on format.
26ASIDE on symbol systems
- To be adequate for encoding human-level
knowledge, physical symbol systems must meet some
stringent conditions on format.
27Conditions on the format of representations
- The format must have the capacity to distinguish
an unlimited number of distinct representations.
This is called the criterion of productivity
which von Humboldt characterized as the Infinite
use of finite means. This means it must be
combinatorial. - The capacity for representation and inference
must be systematic In intelligent agents the
capacity to represent or infer a certain
situation is always accompanied by the capacity
to represent or infer other related situations.
This means it must be compositional.
28Conditions on the format of representations 2
- The capacity for representation and inference is
systematic In intelligent agents the capacity to
represent or infer a certain situation is always
accompanied by the capacity to represent or infer
other related situations. - If an intelligent system can represent some
situations, say Q51 and Q27, then it will also
have the capacity to represent some other
situations related to Q51 and Q27. - If an intelligent system can infer, say, Q36 from
Q75 then it will also have the capacity to infer
other things related to Q36 and Q75. - Systematicity follows naturally from the use of
an encoding system that has constituents (or
parts) and builds up complexes by recursively
embedding parts in wholes, as is done with all
numerals, algebraic expressions, language or
sentences in a logical calculus.
29Conditions on the format of representations 2a
- If you encode representations in terms of
expressions that have a constituent structure,
then systematicity holds in virtue of the format
alone. - E.g., if you can represent John gave the book to
Mary and Mary gave the pen to Fred then you
already have the capacity to represent many other
situations e.g., John gave the pen to Mary,
Mary gave the book to John, Fred gave the pen
to Fred, and all other combinations. - E.g., if you can infer from It is cloudy and
rainy and cold that it is cold then you must
also be able to infer from It is rainy and cold
that it is cold. Such a principle (called
conjunction elimination) is an inference schema
that can be applied to representations that have
constituents (or parts), and once the rule is
part of the system it will apply to arbitrarily
many other representations, resulting in
inferential systematicity.
30Conditions on the format of an adequate system of
representation
- The conditions of productivity and systematicity
entail the compositionality of representations
complex representations are built out of simpler
representations by rules of composition. - Any system of representation meeting these
constraints is a logical calculus or language of
thought (lingua mentis). - The invention of systems of logic that can encode
wide range of beliefs are among the greatest
achievements of the 20th century. - For a more detailed argument see
- Fodor, J. A., Pylyshyn, Z. W. (1988).
Connectionism and cognitive architecture A
critical analysis. Cognition, 28, 3-71.
31Not every smart system is a physical symbol
system
- There are many very complex and intriguing
systems in the mind/body that are not PSSs, and
intelligent action depends on them as well. Much
of early vision, motor control and even
visual-motor coordination is like that. But no
matter how much complexity is built into such
systems they will sooner or later run up against
the inference wall they will have to interface
seamlessly with knowledge-based processes in a
way that does not require some hidden
intelligence in the interface (e.g., without a
homunculus). - Thats why layered intelligences (e.g., Brooks)
will never scale up. Intelligent systems are
not layers of architecture all the way up.
Sooner or later the process depends on
representations rather than on fixed structures.
32End of ASIDE and summary
- The conditions of productivity and systematicity
entail the compositionality of representations
complex representations are built out of simpler
representations by rules of composition. - Any system of representation meeting the above
constraints is a logical calculus or language of
thought (lingua mentis). - For a more detailed argument see
- Fodor, J. A., Pylyshyn, Z. W. (1988).
Connectionism and cognitive architecture A
critical analysis. Cognition, 28, 3-71.
33(3) It has been suggested that since human
intelligence is continuous with animal
intelligence, we should study lower organisms
(e.g., insects) in which cognition is simpler,
and then move up to human intelligence later.
- Comparative psychology does indeed show that many
aspects of vision, visual-motor coordination, as
well as ontological categories and conceptual
systems (e.g., concepts such as physical object,
animate, cause, conspecific) and even parts of
arithmetic are shared by most organisms,
including human infants. - But the research strategy of working up from
lower organisms will not work in general. Even
though we may share a lot with lower organisms,
what we do not share is critical it is
constitutive of intelligence.
34- Although many human cognitive capacities are
found in animals, many other capacities appear
suddenly as we go up the philogenetic scale
(evolution and punctate equilibrium Steven
Jay Gould). - All humans, in contrast with members of other
species, possess the capacity for language, the
instinctive attribution of beliefs and desires to
others, the capacity for counterfactual
reasoning, and the need to provide a theoretical
explanation for events around them (i.e., to
practice inductive and abductive reasoning). - Human vision, language, and action are at the
service of goals and beliefs in a way that makes
human behavior, unlike other animal behavior,
largely stimulus-free - Because all intelligent behaviour is cognitive
penetrable by goals and beliefs.
35(4) AI has traditionally downplayed the role of
the environment in intelligent action But
intelligence must be situated and embodied.
- The first recognition of the importance of the
environment-agent relation in shaping behavior
may be in Simons ant example, but it now makes
an appearance in many areas of contemporary AI. - There is considerable truth in the observation
that intelligence is situated and embodied, but
the deep intellectual puzzles still concern how
the organism represents the world, because (as
discussed earlier) - Its how the world is represented, rather than
how it actually is, that determines intelligent
behavior.
36But also recent A.I. trends have been in the
direction of taking the environment into account.
Recent examples that recognize the importance of
taking the agents environment into account
include
- Emphases on reactive planning, where plans allow
for unexpected sensory input, - Emphasis on active vision, in which vision
interrogates the environment for relevant clues, - Allowing for nonsymbolic aspects of reasoning
through closer links with perception (e.g.,
visual inference Jon Barwise), - The use of indexicals in representations.(See
Lespérance Levesque, 1995 Pylyshyn, 2000,
2001)
37Forms of representation for a robot using
indexicals
From Pylyshyn, Z. W. (2000). Situating vision in
the world. Trends in Cognitive Sciences, 4(5),
197-207.
38Finally, after all this, are there grounds for
optimism about the future of Artificial
Intelligence?
- The simple answer is Yes, of course, otherwise
we would not be here! - But what are some promising trends that offset
the enormous problems still to be solved?
39Some grounds for optimism
- There are a very many accomplishments for which
AI can take credit, despite the AI winter we
have come through. The military, the airlines,
traditional Operations Research domains, robotics
and the games world have all seen major AI
achievements (see Nilsson, 1995). There are also
many small accomplishments, like Google, seen
daily on the web and in nearly every walk of
life, that have widespread effects. Many have
not used the term artificial intelligence but
they nonetheless owe their technology to AI which
has infiltrated much of conventional CS. - Unlike the early days of AI, many more approaches
are being entertained. While most of the grand
polemical claims made by their advocates are
almost certainly wrong, these approaches are
leading to different lines of inquiry, which is
healthy. The diversity is also producing
powerful niche results in areas within
computational vision, robotics, computational
linguistics, and speech understanding.
40There is strength in diversity(also distractions)
Neats
Scruffies
Brooks
Minsky
Newell
Logic
Schank
McCarthy
Analogues
Fuzzy sets
Planning
Robots
Neural nets
Expert ystems
41Some grounds for optimism
- 3. One of the consequences of de-emphasizing pure
reasoning has been more progress in areas where
human cognitive skill is modular (and may not
involve any reasoning). - Vision (especially early vision).
- Computational Linguistics (a large part of
grammatical analysis is modular), - Speech recognition (much of phonetic analysis is
modular), - Visual-motor coordination (much of that is
modular in humans and animals). Lends itself to
Rod Brooks approach. - Hybrid systems, in which the developments of
modular smart subsystems (especially vision and
control) are combined with knowledge-based
systems (e.g., UBCs hybrid controllers and U of
Ts cognitive robotics).
42Some grounds for optimism
- (4) Finally, there has been a large (and
unexpected) resurgence of interest in the very
broad questions of the nature of intelligence,
and it relation to consciousness, to biology, to
evolution and to technology. Books and articles
by Kurzweil, Moravec, Joy, Wolfram, Dennett and
the critical writings of Searle, Fodor, Lanier
and others have once again returned the bigger
questions of human and machine intelligence to
centre stage. - While these may get people thinking about where
we stand in history, they probably only move AI
ahead by focusing public debate and perhaps
awakening funding agencies. But some problems
are just not ready for scientific scrutiny they
are the mysteries as opposed to the puzzles. The
trouble with mysteries, such as the question What
is consciousness? and What are the limits of AI?
is that they are inherently ill-posed they are
not stated in a way that could connect to a
recognizable answer.
43Yet despite the possibility that this new
interest will let in misguided views and give us
bad press, a multitude of approaches have to be
tolerated because nobody knows where the key
discoveries will come from. So we shouldnt put
them down, even when they are wrong (and even
when they have larger grants than we do)!
44.because then we could all lose ..
45References cited
- Pylyshyn, Z. W. (2001). Visual indexes,
preconceptual objects, and situated vision.
Cognition, 80(1/2), 127-158. - Pylyshyn, Z. W. (1999). Is vision continuous with
cognition? The case for cognitive impenetrability
of visual perception. Behavioral and Brain
Sciences, 22(3), 341-423. - Reiter, R. (2001). Knowledge in Action Logical
Foundations for Specifying and Implementing
Dynamical Systems. Cambridge, MA MIT Press. - Winston, P. H. (1999, July, 1999). Why I am
optimistic. Paper presented at the AAAI Annual
Conference.
- Brooks, R. A. (1991). Intelligence without
representation. Artificial Intelligence, 47,
139-159. - Fodor, J. A., Pylyshyn, Z. W. (1988).
Connectionism and cognitive architecture A
critical analysis. Cognition, 28, 3-71. - Lespérance, Y., Levesque, H. J. (1995).
Indexical knowledge and robot action - a logical
account. Artificial Intelligence, 73, 69-115. - Levesque, H. J., Lakemeyer, G. (2001). The
Logic of Knowledge Bases. Cambridge, MA MIT
Press. - Nilsson, N. J. (1995). Eye on the prize. AI
Magazine (summer), 9-17. - Pylyshyn, Z. W. (2000). Situating vision in the
world. Trends in Cognitive Sciences, 4(5),
197-207.