The Symbolic vs Subsymbolic Debate

Download Report

Transcript The Symbolic vs Subsymbolic Debate

The Symbolic vs Subsymbolic
Debate
H. Bowman
(CCNCS, Kent)
Disclaimer
• Serious simplification of a complex debate.
• Emphasize extreme positions to clarify
basic points of controversy.
• What I present is not necessarily what I
personally believe!!
The Mind-Body Problem
I am a
Cartesian
Dualist!
symbolic
• putative characteristics of
cognition
• information processing metaphor
subsymbolic
• inspired by neurobiology
• how cognition emerges from
neurobiology
The Symbolic Tradition
The Computer Metaphor
• Takes inspiration from,
– programming languages & computational logic
• data structures & knowledge representation
• link to programming langs. such as Lisp & Prolog
– computer architectures
• Von Neumann architectures: centralized processing
– computability theory
• the Church – Turing hypothesis
Key Assumptions
• Symbols are available to the cognitive
system
• Symbol processing engine, characteristics,
– Systematic, i.e. combinatorial symbol systems
and compositionality
– Recursive knowledge structures
A set of
“rules”
Syntax: Grammars 1
Sentence:
S ::= NP VP
Noun phrase:
NP ::= det AL N | N
Verb phrase: VP ::= V NP
Adjective list: AL ::= A AL | A
S, NP, VP, AL : molecules
det, N, V, A : atoms
recursion
Syntax: Grammars 2
S
NP
VP
det
AL
N
V
NP
the
A
boy
eats
N
happy
ice cream
A Tree Data Structure
Compositionality
• Plug constituents in according to rules
• Structure of expressions indicates how they should
be interpreted
• Semantic Compositionality,
“the semantic content of a (molecular) representation is a
function of the semantic contents of its syntactic parts,
together with its constituent structure”
[Fodor & Pylyshyn,88]
Compositionality in Semantics 1
• Meanings of items
plugged in as defined
by syntax
M[ X ] denotes
meaning of X
M[ John loves Jane ]
=
M[ John ] + M[ loves ] + M[ Jane ]
+
composed together appropriately
Compositionality in Semantics 2
• Same example in more detail
M[ John loves Jane ]
=
M[ John ] M[ loves ] M[ Jane ]
………….
..………..
Compositionality in Semantics 3
• Meanings of atoms constant across
different compositions
M[ Jane loves John ]
=
M[
Jane
]
M[
loves
]
M[
John
]
………….
..………..
Compositionality in Semantics 4
• Also, meanings of molecules constant
across different compositions
M[ Jane loves John and Jane hates James ]
=
M[ Jane hates James ]
M[
and
]
…..…..….…….
……….………..
M[ Jane loves John ]
Compositionality in Semantics 5
M[ Jane hates James and Jane loves John ]
=
M[ Jane hates James ]
…..…..….…….
M[ Jane loves John ]
M[ and ] ……….………..
Caveat
• Compositionality of course not absolute,
e.g. Idioms: “kicked the bucket”
Compositionality: non-linguistic
examples
• Not just an issue for language
• Reasoning / planning / deductive thought
• Representation of knowledge:
– hierarchical / superordinate structures
Representation of
Visual Objects
From Marr’s
theory of
object Recognition
Whole-part Hierarchies
S
S
S
S
S
S
SSSSSSSSS
S
S
S
S
S
S
Production System Architectures
of the Mind
• Most detailed and complete realisation of
symbolic tradition, e.g.
– SOAR (Unified Theories of Cognition) [Newell]
– ACT-R [Anderson]
– EPIC [Kieras]
• GOFAI (Good Old Fashioned Artificial
Intelligence)
– Based upon expert systems technology
Symbol Systems and Nature vs
Nurture
• Learning theories of symbolic architectures
are rather limited
– although, chunking-based theories do exist
• Where does the symbolic processing engine
come from?
THEREFORE
• Evolutionary explanations, e.g.
– Chomsky’s Universal Grammars
Symbol Systems and the Brain
• For symbolists, the algorithmic / specification
levels are critical, the implementation level is
insignificant (using Marr’s terminology)
• “… for a [Symbolist], neurons implement all
cognitive processes in precisely the same
way, viz., by supporting the basic operations
that are required for symbol-processing …
[i.e.] … all that is required is that you use
your [neural] network to implement a Turing
machine” [Fodor&Pylyshyn,88]
• A sort of compilation step.
• Computers used as an analogy, where software
is the interesting thing and the hardware
mapping is fixed and automatic.
• “… one should be deeply suspicious of the
heroic sort of brain modelling that purports to
address the problems of cognition. We
sympathize with the craving for biologically
respectable theories that many psychologists
seem to feel. But, given a choice, truth is more
important than respectability.”
[Fodor&Pylyshyn,88]
The Sub-symbolic Tradition
Connectionism
• Inspiration from neurobiology
• Long tradition [at least from 50’s], e.g.
Hebb, Rosenblatt, Grossberg, Rumelhart,
McClelland, O’Reilly.
• Nodes, links, activation, weights, learning
algorithms
Activation in Classic Artificial
Neural Network Model
output - yj
sigmoidal
activation
node j value - yj
integrate h   x w
(weighted sum)
net input - hj
w1j
x1
w2j
x2
inputs
y j  1h j
1 e
j
i
i
wnj
xn
ij
Sigmoidal Activation Function
Responsive
around net
input of 0
1
0.9
activation (y )
0.8
0.7
0.6
Unresponsive
at extreme net
inputs
0.5
0.4
0.3
y j  1h
1 e j
0.2
0.1
0
-4
-3
-2
-1
0
1
net input (h )
2
3
4
Threshold:
unresponsive
at low net
inputs
Connectionism and Nature vs
Nurture
• Powerful learning algorithm
– directed weight adjustment
– extracting regularities (Hebbian learning)
– supervised learning (Back-propagation)
• Typically ascribe more to learning than to
evolution
Example Connectionist Model
• Word reading as an example
– Orthography to Phonology
• Words of four letters or less
• Need to represent order of letters,
otherwise, e.g. slot and lots the same
• Slot coding
A (Rather Naïve) Reading Model
PHONOLOGY
/p/.1 /b/.1
/u/.1 /p/.2 /b/.2
/u/.2
/p/.3 /b/.3
/u/.3 /p/.4 /b/.4
/u/.4
A.1 B.1
Z.1 A.2 B.2
Z.2
A.3 B.3
Z.3 A.4 B.4
Z.4
SLOT 1
ORTHOGRAPHY
Connectionism & Compositionality
• Highly non-compositional, e.g.
– a in ant and cat completely unrelated representations
– no sense to which plug in constituent representations
– maximally affected by context
• Same argument would generalise to semantic
compositionality
• Alternative connectionist models do better, e.g.
activation gradient models, but not clear that any
model is truly systematic in sense of symbolic
processing
Systematic / symbolic
Centralised Production Systems
Architectures, e.g. SOAR [Newell]
Spectrum of
Approaches
Parallel Production Systems, e.g. EPIC [Kieras]
(Symbolic) Distributed Control, e.g. Actors [Hewett], Agents
[Kokinov], ICS [Barnard], Society of Minds [Minsky]
Hybrid Approaches, e.g. [Gabbay]
Localist Models with Serial Order, e.g. Solaris [Davis]
Localist / Competitive Learning, e.g. IA model [McClelland]
Distributed Representations / Back-prop., e.g. [Seidenberg]
Unsystematic / subsymbolic
Possible Topics 1
• Introduction to connectionism
– O’Reilly & Munakata, 2000
• Production system architectures
– ACT-R [Anderson,93]
• Connectionism: Strengths and Weaknesses
– Fodor & Pylyshyn, 88
– McClelland, 92 and 95
• Symbolic-like Connectionism
– Hinton, 90
Possible Topics 2
• Past tense debate
– Pinker et al, 2003
• Localist vs distributed debate
– Bowers, 2002 and Page, 2000.
• Dual process theory – system 1 (neural) –
system 2 (symbolic)
– Evans, 2003.
References
•
•
•
•
•
•
•
•
•
•
Anderson, J. R. (1993). Rules of the Mind. Hillsdale, NJ: Erlbaum.
Bowers, J. S. (2002). Challenging the widespread assumption that connectionism and distributed
representations go hand-in-hand. Cognitive Psychology., 45, 413-445.
Evans, J. S. B. T. (2003). In Two Minds: Dual Process Accounts of Reasoning. Trends in Cognitive
Sciences, 7(10), 454-459.
Fodor, J. A., & Pylyshyn, Z. W. (1988). Connectionism and Cognitive Architecture: A Critical Analysis.
Cognition, 28, 3-71.
Hinton, G. E. (1990). Special Issue of Journal Artificial Intelligence on Connectionist Symbol Processing
(edited by Hinton, G.E.). Artificial Intelligence, 46(1-4).
O'Reilly, R. C., & Munakata, Y. (2000). Computational Explorations in Cognitive Neuroscience:
Understanding the Mind by Simulating the Brain.: MIT Press.
McClelland, J. L. (1992). Can Connectionist Models Discover the Structure of Natural Language? In R.
Morelli, W. Miller Brown, D. Anselmi, K. Haberlandt & D. Lloyd (Eds.), Minds, Brains and Computers:
Perspectives in Cognitive Science and Artificial Intelligence (pp. 168-189). Norwood, NJ.: Ablex
Publishing Company.
McClelland, J. L. (1995). A Connectionist Perspective on Knowledge and Development. In J. J. Simon &
G. S. Halford (Eds.), Developing Cognitive Competence: New Approaches to Process Modelling (pp.
157-204). Mahwah, NJ: Lawrence Erlbaum.
Page, M. P. A. (2000). Connectionist Modelling in Psychology: A Localist Manifesto. Behavioral and
Brain Sciences, 23, 443-512.
Pinker, S., Ullman, M. T., McClelland, J. L., & Patterson, K. (2002). The Past-Tense Debate (Series of
Opinion Articles). Trends Cogn Sci, 6(11), 456-474.