Of the various theories for psychological content naturalization put forward within the
past two decades, I believe that a Dretske-style approach that explains the content of a
mental state in terms of the causal history of past tokens of that state holds out the
most promise of giving us a workable theory describing the role that content plays in
learned behavior. While I favor this general approach, the particular theory laid out by
Dretske in Explaining Behavior has a shortcoming that must be addressed before his
theory can be applied to real systems: Dretske fails to provide an analysis of identity
conditions for being a token of an indicator state type. The shortcoming is serious
because of the critical role that past tokens of an indicator type play in fixing the
content of a current token of the indicator type — without identity conditions, there
is no way to specify which previously tokened states among the many that have been
instantiated during the learning period of the organism are of that indicator type.
I begin with a very brief review of Dretske's theory from Explaining Behavior.
Some organisms possess indicator states (i.e., internal states that indicate whether some
external conditions hold). For example, organism O may token an instance of I (the
internal indicator state type) whenever external conditions F obtain. Prior to learning, I
indicates F does not mean F. Let's suppose that external conditions F are relevant in some
manner to O's continued functioning, perhaps because environments in which F obtains are
environments that are relatively inhospitable for O. Let's also suppose that O is capable
of learning using reinforcement information (via operant conditioning), such that future
tokenings of I come to cause movements that are appropriate to conditions F. (My use the
evaluative term "appropriate" here rests on two assumptions: (1) that the
learned pattern of movements given I tends to maximize the pleasure and/or minimize the
pain experienced by O, and (2) that the prior evolutionary history of O's species has
resulted in O being hardwired such that there is a strong correlation between experienced
pain (by O) with inhospitableness and experienced pleasure with hospitableness of the
environment.) After this learning process, I no longer just indicates F, it also means F.
Note that the representational content of I is learned Type III (i.e., the
representational content has no conventional aspect).
One advantage of Dretske's version of content acquisition is that it explains how
content can be efficacious as a structuring cause. Even though, as triggering cause, the
content of a particular psychological state is explanatorily irrelevant, still it was what
the indicator state indicated about external conditions in the past that played a role in
the restructuring of an organism's nervous system, such that this (current) token of that
psychological state type produces this particular movement. It is generally assumed that a
necessary condition for causal efficacy is counterfactual relevance. We can see that, on
Dretske's account, this condition is satisfied, for the content does make a difference in
the movements of O. Had a token of I meant something else (by virtue of indicating some
other external conditions), then the reinforcement information O received would (likely)
have been different, thus resulting in O's having a different post-learning configuration
in accordance with which tokens of I would (likely) fail to cause the movements that Is
came to cause in the actual world.
As mentioned above, I believe that Dretske's theory for the naturalization of content
holds the most promise of providing an analysis of content acquisition resulting from
learning. As it stands, however, there is an important shortcoming in Dretske's theory:
the lack of clear identity conditions for the indicator state types. In the remainder of
this paper, I shall examine this shortcoming, making explicit its relevance for the
overall success of Dretske's program. I then examine several ways in which a proponent of
Dretske's theory may attempt to address the shortcoming, and argue that each of these is
inadequate.
Recall that, according to Dretske's theory, the content of a (just-tokened)
post-learning mental state depends upon the environmental conditions that obtained when
previous tokens of this same type occurred — these environmental conditions played a
role in the reinforcement received by the organism, resulting ultimately in physical
changes to the organism such that now tokens of that indicator state type produce bodily
movements appropriate to the sensed environment. Given the important place that past
instances of an indicator state type have in Dretske's theory of content, it is surprising
that he fails to mention how these state types are to be picked out. One possible
explanation for this is his blurring of the type/token distinction throughout Explaining
Behavior. A more likely explanation is that he is supposing something like this:
tokens of I are physical states, so the most straightforward identity condition for I-hood
is being a token of a particular physical state type. (Presumably, we do not need to go
all the way down to basic physics to specify the state type for all instances of I;
rather, one assumes that the state type can be described in the vocabulary of
neurophysiology.) The trouble with this identity condition is that there is no reason to
believe that exactly the same internal (neuro/)physical state type will be tokened
reliably whenever the external conditions F obtain. Dretske's theory might not demand
flawless tokening of I when F, but his theory requires at least many occasions of a
tokening of I when F during learning. However, neurological research into learning in real
central nervous systems suggests that even this weak requirement will not, in the general
case, be satisfied. If sharing the same (neuro/)physical state type is supposed to
"link together" the tokens of I during the learning period with post-learning
Is, such that the content of the latter is partially grounded in the causal etiologies of
the former, then Dretske's theory is in serious trouble.
Perhaps my dismissal of (neuro/)physical state type as identity condition was too
hasty, for there is a way of describing I that ensures at least that identical sensed
conditions result in the tokening of the same (neuro/)physical state type. If one
specifies that I occurs on the periphery of O's central nervous system (put crudely, I is
the unprocessed raw sense data being fed into the organism's brain), then one is
guaranteed: same sensed external conditions, same (neuro/) physical state type tokened.
There are, though, two important reasons to reject this version of the identity
conditions for I-hood. (I combine these two reasons under the banner the "too
specific" objection.) The first reason to reject it is that it makes representational
content too specific. If psychological generalizations are to be possible, then the
representational content that emerges after the learning process will have to be at least
moderately general. Consider how severely limited possible learned content would be under
this assumption: the sensed input (or, at least that part of the sensed input that
constitutes the indicator state) would have to be identical in order for the same
(neuro/)physical state to emerge from one exposure to the next. Any slight change would
cause a physically different state to be tokened — a state that, by definition, was
not the indicator state in question. The resulting content would likewise be very
specific, in order to circumscribe the conditions F that I indicates, for conditions F
would not just involve some perspective-free external state-of-affairs, but also lots of
other (very specific) conditions about the spatial relation of O to objects mentioned in
F, and (depending on how I is specified) perhaps many other irrelevant ambient conditions.
Dretske cannot non-circularly say "while the actual state- of-affairs that obtains
whenever an I is tokened is severely circumscribed, still the representational content of
I is the general state-of-affairs, G, because it is the function of I to indicate G".
He cannot do this because the function of I is cashed out solely in terms of the learning
history of O; in particular, in terms of the external state-of-affairs F that led to I's
being tokened and that played a role in the reinforcement signal received by O (and the
concomitant changes in O's internal structure). We, as external observers of O during the
learning process, may assign to I the (Type II) representational content of G; but, to O,
I comes to represent the very specific state-of-affairs F.
I want to reiterate that it is inappropriate for us, as outside observers of the
system, to stipulate the content of the indicator state, and, with a wave of our hand,
generalize the content by removing the context-specificity. If the end result of the
learning process is a Type III representational state, then there is a
non-observer-relative fact-of-the-matter with respect to the content of that state.
Perhaps an example from the literature on learning in connectionist systems will help to
bring this point home. Oftentimes failed experiments are more instructive than successful
ones. A case in point is the experiment during the mid-1960s involving the "optical
perceptron" reported in Christiansen and Chater(92). The purpose of the experiment
was to determine whether a connectionist net was capable of learning the concept
picture-with-tank. The training regimen consisted of providing encoded pictures as input,
some of which were scenes containing tanks, and some of which weren't. The net was trained
to output one value when the input encoded a tank-scene, and to output a different value
when the input encoded a scene without a tank. Training went well — the net not only
learned to correctly discriminate between tank-scenes and non-tank-scenes for all of the
scenes presented during the training period, but it could also correctly discriminate on a
few novel scenes taken at the same time as the pictures used during training. Then, the
researchers photographed some more tank-scenes and non-tank-scenes. They were surprised to
find that their trained net failed miserably at discriminating between the two scene
types. On closer examination of the features generalized by the network, they found that
it was really discriminating between high-intensity light scenes and low-intensity light
scenes. On re-checking the original photographs used in training, they discovered that the
tank pictures were all shot during the same time period of bright sunlight, and the
non-tank pictures were all shot during the same time period of less-bright sunlight. The
moral here is two-fold: (1) there is a matter-of-fact (independent of external observers)
with respect to the content of a Type III representational state, and (2) to assume that
one can stipulate away context-specificity is wholly unjustified.
There is another reason to reject the identification of an indicator state type with a
(neuro/)physical state type on the periphery of O's central nervous system. This is the
second part of the "too specific" objection. Even if one is willing to allow
highly specific contents as the end-product of learning, there are still problems for
Dretske's theory. His theory assumes that O will have tokened Is many times in the past,
and that (as a result) O has enough reinforcement information to place I in the causal
pathway leading to some movement appropriate to conditions F. But if I is only tokened a
few times (or maybe only one time) during O's learning period, O will not have enough
information to place I correctly, as required. The extreme specificity of I ensures that
it will rarely be tokened, because it is only rarely that an organism receives exactly the
same sensed data. This is a twist on the standard "poverty of the stimulus"
argument.
Thus, the identification of an indicator state type with an internal (i.e.,
non-peripheral) (neuro/)physical state type flies in the face of recent results from
neuroscience, while the identification of an indicator state type with a peripheral
(neuro/)physical state type is subject to the "too specific" objection.
What other options are open to a proponent of Dretske's theory for psychological
content naturalization? The first approach one might take is to bite the bullet on
specificity, but to make it less objectionable by construing specificity as a form of
context-relativity. Anyone who buys a Dretske-style explanation of content acquisition is
already willing to have contents be context-relative — maybe this specificity is just
another aspect of context-relativity. So, the extreme specificity of content is
transformed from a flaw of the causal-historical view of content into a feature. This tack
shows a misunderstanding of what context-relative means as an aspect of any
causal-historical account of content. If the contents of my mental states are in part a
function of my past learning history, then those contents will be relative to certain
contingent facts about what states-of-affairs I found myself in and my particular
reinforcement schedule. So much is undeniable and not in itself a drawback of the theory
of content according to which content is relative to the causal-historical contexts of the
agent. The "too specific" objection is not an objection against this aspect of
Dretske's theory, but rather an objection against over-specificity of the learning
contexts. On this account, the content of my mental states would be relative not only to
the particular objects with which I had interacted, but also to my exact location relative
to those objects, the ambient lighting conditions present at the time I saw the objects
(for cases in which the indicator state has a visual component), etc. This is not good old
context-relativity in the benign sense, and does not constitute a bullet that one should
be too willing to bite, for its acceptance implies the illegitimacy of the view of
psychology as the search for (at least weakly) universalizable generalizations.
A second approach that one might take to circumvent the "too specific"
objection is to change the identity conditions for I-hood to membership in a set of
(neuro/)physical state types. The problem with this approach is explaining how being a
member of a set can make a difference, over and above being a collection of individual
instantiations of (neuro/)physical state types. It will still be the case that each of the
member state types in the set will only rarely be instantiated. Without some way of
showing how being a member of this set makes a causal difference to the learned behavior
(such that the ultimate content acquired by tokens of I can legitimately be adverted to in
psychological explanations), it is unclear how this approach could avoid the "too
specific" objection.
Maybe, one might reply, there is some causally relevant property that binds together
the various (neuro/)physical state types in the set (and that could in turn serve to bind
together the tokens of those state types) — namely, being tokens that indicate F. The
main problem with using indicates F as the criterion for inclusion of a (neuro/)physical
state type in the set that constitutes the indicator is circularity. As this is also the
main problem with the third failed attempt discussed below, I postpone a closer
examination of it until after I offer the third attempt at thwarting the "too
specific" objection.
All of the individual tokens of I have one feature in common: they all indicate that
external conditions F obtain. Maybe this is the identity condition for I-hood that we
should adopt: a state is an instance of the indicator type if it indicates F. In its
favor, this does circumvent the "too specific" objection, but at a huge cost.
The theory of content that results is circular: it attempts to explain how a physical
state (picked out by virtue of the fact that that state indicates F) comes to means F. The
advantage of describing I-hood in terms of a state's (neuro/)physical state type is that
it does not presuppose what it is attempting to explain: it does not presuppose that there
is something (prior to learning) about the state that refers to some external
state-of-affairs. At least (neuro/)physical state type descriptions do not advert to
external conditions.
What we need is some way of picking out instances of I that is syntactic (i.e., some
way of picking out instances of I that can, in theory, be cashed out in a non-extrinsic
physicalist vocabulary), and that can be used to provide a causal explanation for why
current tokens of I produce the movements that they produce. Identifying I-hood by its
indicated external conditions fails to satisfy the first criterion above, whereas
identifying I-hood by membership in a set of (syntactically-described) (neuro/)physical
state types fails to satisfy the second. We, as observers of O, are also not allowed to
bridge the gap between set membership and causal relevance by a judicious choice of the
(neuro/)physical state types that make up the set. We are assuming that O will eventually
be a Type III representational system requiring no outside agents to ground the content of
its representational states.
So, where does this leave us (and Dretske)? The lack of an analysis of the indentity
conditions for indicator state types is a serious shortcoming for Dretske's theory. Of the
three attempted analyses I considered here, all are seen to be inadequate. Obviously, this
does not show that an adequate analysis is impossible. (Indeed, I have argued elsewhere
that, with the aid of the connectionist model of operant conditioning, an adequate
analysis can be offered). However, the above argument at least demonstrates that Dretske's
theory as it now stands is incomplete.
|
