Disciplines in Distress:

imminentpoppedAI and Robotics

Feb 23, 2014 (7 years and 2 months ago)

234 views

Disciplines in Distress:
Artificial Intelligence and
Connectionism

Ath. Kehagias

School of Engineering

Aristotle Univ.

Some Remarks


Not a philosophical talk (since I
am not a philosopher but a
philo
-
philosopher).


Not a thesis, just an interesting
(?) story and some questions.


Twenty
-
five slides, roughly one
quote per slide.


Feel free to interrupt at any
point.

First Definitions


Artificial Intelligence:
The science of making
machines do things that would require
intelligence if done by people.


Also known as
Symbolicism, Symbolic Artif. Int.
(SAI), Good Old
-
Fashioned Artif. Int. (GOFAI)
etc.


(Because of the extensive use of symbol
-
manipulating approaches by the practitioners.)


Connectionism
:
A computational approach to
modeling the brain which relies on the
interconnection of many simple units to produce
complex behavior


Also known as
Neural Networks


(from the
Dictionary of Philosophy of Mind
, www.artsci.wustl.edu/~philos/MindDict)

01

An Example of AI (GPS)

02a

(define *school
-
ops*


(list



;; If your son is at home and your car works, it is


;; possible to drive him to school. (Then he'll be at


;; school and will no longer be at home.)



(make
-
op "drive son to school"


'(son
-
at
-
home car
-
works)


'(son
-
at
-
school)


'(son
-
at
-
home))




;; If your car needs a new battery, and the mechanic


;; knows the problem



;; and has been paid, it is possible him to install the


;; new battery. Then the car will work.



(make
-
op "have the mechanic install a new battery"


'(car
-
needs
-
battery mechanic
-
knows
-
problem


mechanic
-
has
-
money)


'(car
-
works)


'(car
-
needs
-
battery))




Here, then, are a couple of problems that
GPS can solve, using these operations:




> (GPS '(son
-
at
-
home car
-
works) '(son
-
at
-
school) *school
-
ops*)

drive son to school


> (GPS '(son
-
at
-
home car
-
needs
-
battery have
-
phone
-
book
have
-
money) '(son
-
at
-
school) *school
-
ops*)

look up the telephone number

telephone the mechanic

tell the mechanic what the problem is

pay the mechanic

have the mechanic install a new battery

drive son to school


02b

An Example of AI (GPS)

An Example of NN

03

The training procedure for TD
-
Gammon is as
follows: the network observes a sequence
of board positions starting at the opening
position and ending in a terminal position
characterized by one side having removed
all its checkers. The board positions are fed
as input vectors x[1], x[2], . . . , x[f] to the
neural network. Each time step in the
sequence corresponds to a move made by
one side. For each input pattern x[t] there
is a neural network output vector Y[t]
indicating the neural network's estimate of
expected outcome for pattern x[t]. At each
time step, the TD(lambda) algorithm is
applied to change the network's weights.
The formula for the weight change is as
follows:


Second “Definitions”

There is a community of people who do (
S)AI
.
They attempt to create intelligent entities (usually
software). To this end they use computer programs
which manipulate symbolic data structures (lists,
trees, graphs etc.).


There is
another

community of people who do
NN
. They attempt to create intelligent entities
and/or model human intelligence. To this end they
use computer programs which manipulate
numbers.








(from
Thanasis’ Own Dictionary
)

04

Timeline

Year
Connectionism Events
Symbolic Artificial Intelligence
Events
1943
McCulloch+Pitts: “A Logical Calculus of the Ideas Immanent in
Nervous Activity”
1950
Turing: “Computing Machinery
and Intelligence”
1954
Hebbian Learning
1956
Dartmouth conference, 1
st
use of
the term “Artificial Intelligence”
1957
Newell,
Shaw+Simon: “General
problem solver”
1958
Rosenblatt: The Original
Perceptron
Samuel:
Chekers playing
program
The LISP language introduced
1960
Widrow:
Adaline
1962
Rosenblatt: The Classic
Perceptron,
Principles of
Neurodynamics and the Theory
of Brain Mechanisms
Thomas: Analogy
1963
Dept. of
Defence: beginning of ARPA projects
Dendral (1
st
expert system)
1968
Minsky+Papert:
Preceptrons
1971
ETAOIN SHRDLU
1972
The PROLOG language
introduced
Dreyfus:
What Computers
Cannot Do
1979
Neural network research
drastically reduced
INTERNIST, MYCIN expert
systems
1982
Hopfield networks
1986
PDP book
1987
1
st
Int. Conference on Neural
Networks
1990
Dept. of
Defence: major funding
of neural network projects
Lots of stuff going on
1991
TD-GAMMON (Neural network
backgammon playing program)
1993
Dreyfus:
What Computers Still Cannot Do
05

Early AI Goals

“It is not my aim to surprise or shock you …
But the simplest way I can summarize is to
say that there are now in the world
machines that think, that learn and that
create. Moreover, their ability to do these
things is going to increase rapidly, until
--
in a visible future
--

the range of problems
they can handle will be coextensive with
the range to which the human mind has
been applied.”




H. Simon and A. Newell, “Heuristic problem solving: the next advance in operations
research”,
Op. Res.
, vol.6, p.6, 1958.

06

Early Critique of AI

What Computers Can’t Do
, H.L. Dreyfus,
1972.


Four Assumptions


Biological Assumption:


Psychological Assumption


Epistemological Assumption


Ontological Assumption


Two Important (and Missing) Factors:


The Body (Embodied Intelligence)


The Situation

07

Current AI Goals

“Douglas Lenat has a ten
-
year program of
building a huge semantic memory (CYC).
Then we will see … When people start to
build programs at that magnitude and they
still cannot do what they are supposed to,
then

we will start worrying.”

.

(H.A. Simon, “Technology is not the problem” In P.Baumgartner and S.Payr,
Speaking
Mind
, Princeton UP, 1995.)


“AI no longer does Coginitive Modeling. It is
a bunch of techniques in search of practical
problems.”


(J. Feldman cited in H.L. Dreyfus,
Artif. Intelligence
, vol.80, p.171
-
191, 1996)

08

Current Critique of AI

What Computers Still Can’t Do
, H.L.
Dreyfus, 1993.

09

Early NN Goals

"Perceptrons are not intended to serve as
detailed copies of any actual nervous
system. They're simplified networks,
designed to permit the study of lawful
relationships between the organization of a
nerve net, the organization of its
environment, and the 'psychological'
performances of which it is capable.
Perceptrons might actually correspond to
parts of more extended networks and
biological systems; in this case, the results
obtained will be directly applicable.”


(F. Rosenblatt,
Principles of Neurodynamics: Perceptrons and the Theory of Brain
Mechanisms, Spartan Books, 1962)

10

Perceptrons

(the book)

“The final episode in this era was a campaign
led by Marvin Minsky and Seymour Papert
to discredit neural network research and
divert neural network research funding to
the field of “artificial intelligence” … The
campaign was waged by means of personal
persuasion by Minsky and Papert and their
allies, as well as by limited circulation of a
technical manuscript (which was later de
-
venomized and, after further refinement
and expansion, published in 1969 by
Minsky and Papert as the book
Perceptorns.



(R. Hecht
-
Nielsen,
Neurocomputation,
1990)

11

NN Resurgence

(AI Stagnation)

“PDP models...hold out the hope of offering
computationally sufficient and
psychologically accurate mechanistic
accounts of the phenomena of human
cognition which have eluded successful
explication in conventional computational
formalisms…”











(D.E. Rumelhart, J.L. McClelland, and the PDP Research Group
,Parallel Distributed
Processing: Explorations in the Microstructure of Cognition,

MIT Press, 1987)

12

The BandWagon Effect

“Undoubtedly, the emergence of 'new'
connectionism was accompanied by a
certain amount of jumping on the
proverbial connectionist bandwagon.”










Istvan S. N. Berkeley, “A Revisionist History of Connectionism”, 1997,
http://www.ucs.louisiana.edu/~isb9112/dept/phil341/histconn.html

13

NN Stagnation 1

“I think and have thought for the last twenty
years that the future consist of, just in a
few days time, discovering efficient
unsupervised learning algorithms that find
a suitable representation. And I still believe
that. I think this can be a huge
technological payoff to making this work
well. And I think the talk I gave this
morning is a small amount of progress in
that direction and on the technological
front I think that is one of the major things
that can happen in the next five years. For
the last twenty years I’ve been saying "it’s
gonna happen in the next five years" and I
keep believing that.”







G. Hinton,
The Future and Prospects of Neural Networks: The Workshop in Edinburgh
(Sep 8, 1999)

14

NN Stagnation 2

“ The maturing of neural networks presents
an interesting study in hyperbole and
substance. Those of us who jumped on the
bandwagon early (in the heady days of
"connectionism") foretold of a revolution
that has not materialized as yet; looking
back, I think we thought neural nets would
have the same scale of impact as the World
Wide Web has had.”









From the Editor,
Control Systems Magazine
, October 2000


15

NN Stagnation?

“Given the multiple relationships and
interdependencies that now exist between
financial markets, neural nets have a
natural role to play. Their capacity to
process and detect relationships and
patterns in huge quantities of data goes far
beyond that of a human trader. `Neural
networks can find patterns in what would
otherwise be disparate data that a human
being would not visually be able to
discern,’ says Mendelsohn. `From the
standpoint of performing intermarket
analysis, it is the right tool for the job’. ”








Vantage Point: Intermarket Analysis Software. At their WebSite
(http://www.profittaker.com/futures_options_new.asp) Andrew Webb reports on the
latest resurgence of interest in neural network technology (2000).

16

AI/NN: Similarities


They both have attempted to perform some
form of cognitive modeling.


They both have claimed that they can
produce intelligent behavior.

17

AI/NN: Differences


AI: works at the symbol manipulation
level.


NN works at the parallel distributed
computation (sub
-
symbolic) level.

18

AI/NN: Sociohistorical
Comparison


In both cases grandiose claims were made
at the start.


In both cases the claims were not realized.


In both cases a sub
-
product was a toolbox
of (very) useful agorithms


In both cases we have a degenerating
research program (?)

19

AI/NN: Interactions

A story from the 1988 Connectionist Models
Summer School


“Hybrid” Systems (e.g. “Connectionist
Symbol Processing”).

20

Beyond AI/NN:
Computational Intelligence

“Flash forward to the late 1990s, and you'll
find a hauntingly familiar atmosphere
surrounding the evolutionary computation
field (also referred to in some circles as
genetic algorithms, after the technology
that has been successfully embodied into
software tools, or the loftier catch
-
all term
artificial life). … Neural networks and
evolutionary computing and fuzzy logic
can all be lumped under the general phrase
"computational intelligence," and for only
the second time in this decade an entire
conference was devoted to the research
efforts of all three groups.”




Intelligent Systems Report,
May 1998, Vol. 15, No. 5, www.lionhrtpub.com/ISR/isr
-
5
-
98/wcci98.html

21

Philosophical Issues


Which of the two (AI vs. NN) is more
succesful? Why?

22

Sociohistorical Issues


The ebb and flow of each field’s
popularity?


Is there some kind of vacuum which must
be filled by
one
theory of intelligence?

23

“Religious” Issues


Why did the debate between Symbolicists
and Connectionists become so emotional?


Why is there so strong resistance to the
idea of Artificial Intelligence? (Dreyfus,
Searle, Fodor)


Some Possible Explanations


Scientific Conflict


Financial Conflict


The Frankenstein Syndrome


24

Generalizations and
Extensions


Mathematics and Computer Science


Anthropology and Sociology


Philology and PostModern Studies




To what extend is

Academia / University changing

and how?

25