Summer 2011
Tuesday, 8/2
608. No supposition seems to me more
natural than that there is no process in
the brain correlated with associating or
with thinking; so that it would be
impossible to read off thought-processes
from brain-processes. I mean this: if I talk
or write there is, I assume, a system of
impulses going out from my brain and
correlated with my spoken or written
thoughts. But why should the system
continue further in the direction of the
centre? Why should this order not
proceed, so to speak, out of chaos? . . .
609. It is thus perfectly possible that
certain psychological phenomena cannot
be investigated physiologically, because
physiologically nothing corresponds to
them.
Connectionism and Folk-Psychology
• Stich, Ramsey and Garon argue that if (a certain
version of) Connectionism is true, then we must
be Eliminativists about propositional attitudes.
• They first argue that if Connectionism is true, we
must either be Eliminativists or Reductivists
about propositional attitudes. (The theory change
must either be “conservative” or “radical”)
• Given that there is nothing, in a connectionist
model, that we could plausibly identify
propositional attitudes (e.g. beliefs) with, the
truth of connectionism forces us to deny any sort
of reality to the attitudes (or to make a “radical”
theory change).
Connectionism and Folk-Psychology
• Propositional Modularity (Fodor’s “essence
of the attitudes”): the idea that our ordinary
use of propositional attitude talk involves a
commitment to discrete, semantically
interpretable states that play a causal role in
the production of other mental states and
behavior.
• Reasons to accept it: (1) our talk of gaining
and loosing beliefs one at a time, (2) our
explanations of people’s behavior involve
citing specific beliefs as causes.
Connectionism and Folk-Psychology
Properties of Connectionist Models:
1. Their encoding of information in the
connection weights is widely distrubuted.
2. Individual hidden units in the network have
no comfortable symbolic interpretation. The
networks are not semantically transparent,
but are sub-symbolic.
Implementational vs. Cognitive connectionism.
Connectionism and Folk-Psychology
Reasons to think that Connectionism is
incompatible with Propositional Modularity.
Consider a connectionist network that’s
trained to give “yes”/”no” answers to a set of
16 questions, e.g. “do dogs have fur?”, “do fish
have fur?”, etc.
Connectionism and Folk-Psychology
1.
“In the connectionist network, there is no distinct state or part of the
network that serves to represent any particular proposition. The
information encoded in the network is stored holistically and
distributed throughout the network. Whenever information is
extracted from the network, by giving it an input string and seeing
whether it computes [a yes or no answer] for the output unit, many
connection strengths…many hidden units play a role in the
computation. And any particular weight or unit will help to encode
information about many different propositions. It simply makes no
sense to ask whether or not the representation of a particular
proposition plays a causal role in the network’s computation. It is in
just this respect that our connectionist model seems radically
incongruent with the propositional modularity of common sense
psychology.” (Stich, et el)
Connectionism and Folk-Psychology
2. Consider what happens when we compare the original
network with one that involves an additional piece of
knowledge.
Intuitively, we want to say that the two networks share many
of the same beliefs.
Yet the two nets may have no subset of weights in common!
Their commonality is invisible at the level of units and
weights. But if there were some discrete states in the
connectionist model that played the role of beliefs, we would
expect there to be lots of commonalities.
[1] and [2] suggest that there are no discrete, semantically
evaluable and causally potent states in a connectionist
network that could plausibly be identified with beliefs. So if
connectionism is the correct model of the mind, we must
deny that the mind really contains entities like beliefs.
Connectionism and Folk-Psychology
Objection 1: Connectionist models do not really violate
propositional modularity, since the propositions the
system has learned are coded in functionally discrete
ways, though this may not be obvious.
(Think of the way propositions are stored in a
computer, at physically scattered memory addresses.
Still, if one knew enough about the system, one could
erase any one sentence by tampering with the contents
of the relevant memory addresses.)
Reply: This is a possibility, but at present, there is no
reason to take it seriously.
Connectionism and Folk-Psychology
Objection 2: The propositions are encoded in
the patterns of activation of the hidden units,
when a given question is presented to the
network.
Reply: Such patterns of activation are not
enduring states of the network. So it’s
implausible that they play the role of
knowledge or beliefs.
Connectionism and Folk-Psychology
Objection 3: Beliefs should not be identified with
activation patterns (or any transient states) but
with dispositions to produce activation patterns
given certain inputs.
Reply: Dispositions are enduring states, but they
are not the right sorts of enduring states to be
identified with beliefs. In particular, they do not
seem to be capable of playing a causal role.
Connectionism and Folk-Psychology
Other responses to Stich (et al) are:
1. Question the commitment to propositional
modularity (Dennett).
2. Accept that if (cognitive) connectionism is
true, then Folk Psychology is false, but argue
that this conditional is a reason to reject
connectionist models of the mind rather than
Folk Psychology.
Systematicity
Fodor would go with 2, but he also argues
against connectionist models of (certain parts
of) the mind like this:
1. Thought is systematic.
2. So internal representations are structured.
3. Connectionist models lack structured internal
representations;
C. Connectionist models are not good models of
human thought.
Systematicity
Responses:
1. Classical symbol systems are not the only way to
support systematically structured cognition.
(Problem with this response: neural networks
are not particularly good at the kind of rule
based processing that is thought to undergird
language, reasoning, and higher forms of
thought.)
2. Human thought may inherit its systematicity
from the grammatical structure of language.
(This reply downplays the extent and
importance of systematicity, something that
may be independently plausible).
Quiz
• What are artificial neural networks and how
do they differ from physical symbol systems?
• What reason is there for thinking that
(cognitive) connectionism implies
eliminativism about propositional attitude
psychology?