Fodors and others Critique of Connectionism

1
Fodors (and others) Critique of Connectionism

• Elisa Lawler

2
Fodor and Pylyshyn 1988

• Necessary properties of any theory of cognition
• Systematicity
• Productivity

3
Systematicity

• John loves the girl
• The girl loves John
• you don't find people who can think the
  thought that John loves the girl, but can't think
  the thought that the girl loves John." (FP) "But
  now if the ability to think that John loves the
  girl is intrinsically connected to the ability to
  think that the girl loves John, that fact will
  somehow have to be explained." (FP)

4
Related thoughts have shared constituents

• Fodor (1998) writes "mental representations are
  compositional, and compositionality explains
  systematicity. The reason the capacity for 'John
  loves Mary' thoughts implies a capacity for 'Mary
  loves John' thoughts is that the two kinds of
  thoughts have the same constituents ..."

5
Problem with Systematicity

• Systematicity itself is questioned
• In non-linguistic cognitive architecture (which
  is shared by animals)
• Capacity to think I want to eat the bird does
  not imply that the same animal has the capacity
  to think The bird wants to eat me
• Possibly phenomenon of systematicity is actually
  a result of experiences

6
Systematicity and Connectionism

• FP (1988) claim that connectionist models are
  incapable of accounting for systematicity because
  they lack syntactic complexity
• Systematically related propositions share some
  syntactic parts, therefore some sort of syntactic
  representational structure must exist (they dont
  in traditional connectionist networks)

7
Productivity

• Representational capacity is unbounded given
  finite resources
• E.g. given rules of language and words, an
  infinite number of sentences can be produced
• E.g. given numbers and formulas, infinite number
  of computations can be executed
• Connectionist models can not account for this
• Requires infinite memory capacity because there
  would be an infinite number of node
  combinations/patterns

8
Classical View and LOT

• Fodor claims that Classical views naturally
  account for systematicity and productivity
  whereas Connectionist models do not
• Classical Model postulates a 'language of
  thought' (LOT) which explains these properties

9
LOT

• Uses combinatorial syntax and semantics in
  forming mental representations ( FP, 1988)
• Example in natural language
• John loves Mary
• John, Mary loves are syntactic components
  which combine (via syntax and grammar) into
  sentence each word also has semantic meaning

10
LOT explains Systematicity

• If concepts are words and phrases in the LOT,
  then accessing a concept is just the occurrence
  of words and phrases from the LOT -gtyou can find
  similarity in the constituents of 2 different
  concepts (systematicity follows)

11
LOT explains Productivity

• Separation of rules and semantics allows
  productivity
• Linguistic competence

12
Smolenskys PTC

• "As far as I can tell, the argument has gone like
  this Fodor and Pylyshyn claimed that you can't
  produce a connectionist theory of systematicity.
  Smolensky then replies by not producing a
  connectionist theory of systematicity. Who would
  have foreseen so cunning a rejoinder." (Fodor,
  1991, 279)

13
Smolensky in pursuit of connectionist model that
recognized systematistic cognitive capacities

• 1) decompose target knowledge into roles and
  fillers to represent lta,b,cgt ordered string,
• 2) roles are position1, position2, position3
• 3) fillers represented as letters a,b,c
• 4) each role and filler get a connectionist
  vectorial
• representation, and role and vector are
  bound
• together by vector multiplication
• 5) As a result, a system can differentially
  represent
• ltabcgt and ltbcagt

14
Fodor rejects PTC

• Fodor claims that Smolenskys PTC is really a
  disguised implementation of the Classical view,
  and can not provide a completely independent
  account for systematcity
• Fodor calls vector products derived constituents

15

• BROWN TREE
• Given a tree, the algorithm yields a
  corresponding vector given a tree, the algorithm
  yields the tree from which the vector is derived
• Vectors dont have constituents, but vectors can
  be derived from trees, which do have constituents.

16
Changes in Belief

• If there is no way to generalize elements in one
  network to another, then there is no way to embed
  the same belief in different situations (The Mind
  Doesnt Work That Way)
• Connectionist networks can only manipulate
  strength of the connections, can not change the
  underlying architecture of the system

17
Other Criticisms

• Not a correct neurological representation of
  brain not as complicated as real brain
• But bigger nets pose problem of overwriting and
  breaking down
• 1st wave of Connectionist models attacked for
  representing only limited/specific cognitive
  tasks
• Artificial input/output representations

18
Conclusions

• Fodor others have rejected Connectionist models
  because they lack any independent way of
  explaining systematicity and productivity
• Smolenskys attempt at a Connectionist model for
  systematicity was rejected by Fodor who claimed
  that they fundamentally relied on Classical view
  for explaiming systematicity and productivity
• Brain is more complicated than any Connectionist
  model
• Training is based on sterile/artificial
  situations which discredit validity of resulting
  performance as a true cognitive model