<<

. 26
( 62 .)



>>

the objects of computational mechanisms, Marr™s references to the “real
world” do not commit him to an externalist view of the taxonomy of visual
states and processes.
The second is to take these comments to suggest merely a heuristic role
for the structure of the real world, not only in developing a computational
taxonomy but also in the computational theory of vision more generally.
Representation, Computation, and Cognitive Science 155

That is, turning to the beyond-the-head world is a useful shortcut for
understanding how vision works and the nature of visual states and com-
putations. This heuristic is effective either in providing important back-
ground information that allows us to understand the representational
primitives and thus the earliest stages of the visual computation, or by
serving as an interpretative lens that allows us to construct a model of
computational processes in terms that are meaningful. Again, as with the
previous option, the beyond-the-head world plays only a peripheral role
within computational vision, even if Marr at times refers to it prominently
in outlining his theory.
Beyond questions of how to interpret Marr™s own comments, individ-
ualists have objected to Burge™s argument in two principal ways. First,
Gabriel Segal and Robert Matthews have both in effect denied (2), with
Segal arguing that these intentional primitives (such as edges and gener-
alized cones) are better interpreted within the context of Marr™s theory
as individuated by their narrow content. Second, Frances Egan has more
strikingly denied (1), arguing that, qua computational theory, Marr™s the-
ory is not intentional at all. Both objections are worth exploring in de-
tail, particularly insofar as they highlight issues that remain contentious
in contemporary discussions. In fact, Marr™s theory raises more founda-
tional questions than it solves about the nature of the mind and how we
should investigate it.15


5 segal and egan on computation and representation
Segal points out that there are two general interpretations available when
one seeks to ascribe intentional contents to the visual states of two indi-
viduals. First, one could follow Burge and interpret the content of a given
visual state in terms of what normally causes it. Thus, if it is a crack in a
surface that plays this role, then the content of the corresponding visual
state is “crack”; if it is a shadow in the environment that does so, then
the content of the visual state is “shadow.” This could be so even in the
case of doppelg¨ ngers, and so the visual states so individuated are not indi-
a
vidualistic. But second and alternatively, one could offer a more liberal
interpretation of the content of the visual states in the two cases, one
which was neutral as to the cause of the state, and to which we might give
the name “crackdow” to indicate this neutrality. This content would be
shared by doppel¨ ngers, and so would be individualistic.
a
The crucial part of Segal™s argument is his case for preferring the sec-
ond of these interpretations, and it is here that one would expect to ¬nd
an appeal to the speci¬cs of Marr™s theory of vision. While some of Segal™s
Individualism and Externalism in Cognitive Sciences
156

arguments here do so appeal, Segal also introduces a number of quite
general considerations that have little to do with Marr™s theory in partic-
ular. For example, he points to the second interpretation as having econ-
omy on its side, thus appealing to considerations of simplicity, and says,

The best theoretical description will always be one in which the representations
fail to specify their extensions at a level that distinguishes the two sorts of distal
cause. It will always be better to suppose that the extension includes both sorts of
thing.16

Why always? Segal talks generally of the basic canons of good explanation
in support of his case against externalism, but as with the appeals to the na-
ture of scienti¬c explanation that turned on the idea that scienti¬c taxon-
omy individuates by “causal powers” that we discussed in Chapter 4, here
we should be suspicious of the level of generality (and corresponding lack
of substantive detail) at which scienti¬c practice is depicted. Like Burge™s
own appeal to the objectivity of perceptual representation in formu-
lating a general argument for externalism, these sorts of a priori appeals
seem to me to represent gestural lapses entwined with the more interest-
ing, substantive, empirical arguments over individualism in psychology.
When Segal does draw more explicitly on features of Marr™s theory, he
extracts three general points that are relevant for his argument that the
theory is individualistic: Each attribution of a representation requires a
bottom-up account, a top-down motivation, and is checked against be-
havioral evidence. Together these three points imply that positing repre-
sentations in Marr™s theory does not come cheaply, and indeed is tightly
constrained by overall task demands and methods. The ¬rst suggests that
any higher-level representations posited by the theory must be derived
from lower-level input representations; the second that all posited repre-
sentations derive their motivation from their role in the overall percep-
tual process; and the third that “intentional contents are inferred from
discriminative behavior.”17
Segal uses the ¬rst assumption to argue that since the content of
the earliest representations (“up to and including zero-crossings”) in
doppelg¨ ngers are the same, there is a prima facie case that downstream,
a
higher-level representations must be the same, unless a top-down motiva-
tion can be given for positing a difference. But because we are considering
doppelg¨ ngers, there is no behavioral evidence that could be used to diag-
a
nose a representational difference between the two (Segal™s third point),
and so no top-down motivation is available. As he says, “[t]here would
just be no theoretical point in invoking the two contents [of the twins],
Representation, Computation, and Cognitive Science 157

where one would do. For there would be no theoretical purpose served
by distinguishing between the contents.”18
How might an externalist resist this challenging argument? Three dif-
ferent tacks suggest themselves, each of which grants less to Segal than
that which precedes it.
First, one could concede the three points that Segal extracts from his
reading of Marr, together with his claim that the lowest levels of represen-
tation are individualistic, but question the signi¬cance of this. Here one
could agree that the gray arrays with which Marr™s theory begins do, in a
sense, represent light intensity values, and that zero-crossings do, in that
same sense, represent a sudden change in the light intensity. But these
are both merely representations of some state of the retina, not of the
world, and it should be no surprise that such intraorganismic representa-
tions have narrow content. Moreover, the depth of the intentionality or
“aboutness” of such representations might be called into question pre-
cisely because they don™t involve any causal relation that extends beyond
the head. They might be thought to be representational in much the way
that my growling stomach represents my current state of hunger. How-
ever, once we move to downstream processes, processes that are later on in
the temporal dimension to visual processing, genuinely robust represen-
tational primitives come in to play, primitives like “edge” and “generalized
cone.” And the contents of states deploying these primitives, one might
claim, as representations of a state of the world, metaphysically depend
on what they correspond to in the world, and so are not individualistic.
The plausibility of this response to Segal turns on both the strength of
the distinction between a weaker and a stronger sense of “representation”
in Marr™s theory, and the claim that we need the stronger sense to have
states that are representational in some philosophically interesting sense.
Second, and more radically, one could allow that all of the represen-
tational primitives posited in the theory represent in the same sense, but
challenge the claim that the content of any of the corresponding states is
narrow: it is wide content all the way out, if you like. The idea that the rep-
resentational content of states deploying gray arrays and zero-crossings is
in fact wide might itself take its cue from Segal™s second point “ that rep-
resentations require a top-down motivation “ for it is by re¬‚ecting on the
point of the overall process of constructing reliable, three-dimensional
images of a three-dimensional visual world that we can see that even early
retinal representations must be representations of states and conditions
in the world. This view would of necessity go beyond Marr™s theory itself,
which is explicitly concerned only with the computational problem of
Individualism and Externalism in Cognitive Sciences
158

how we infer three-dimensional images from impoverished retinal infor-
mation. But it would be very much in the spirit of what we can think of
as a Gibsonian aspect to Marr™s theory.
Third, and least compromisingly, one could reject one or more of
Segal™s three points about Marr™s theory or, rather, the signi¬cance that
Segal attaches to these points. Temporally later representations are de-
rived from earlier representations, but this itself doesn™t tell us anything
about how to individuate the contents of either. Likewise, that Marr him-
self begins with low-level representations of the retinal image tells us little
about whether such representations are narrow or wide. Top-down moti-
vations are needed to justify the postulation of representations, but since
there are a range of motivations within Marr™s theory concerning the
overall point of the process of three-dimensional vision, this also gives
us little guidance about whether the content of such representations is
narrow or wide. Behavioral evidence does play a role in diagnosing the
content of particular representations, but since Marr is not a behaviorist,
behavioral discrimination does not provide a litmus test for representa-
tional difference.19
This third response seems the most plausible to develop in detail, but
it also seems to me the one that also implies that there is likely to be
no de¬nitive answer to the question of whether Marr™s theory employs
either a narrow or a wide notion of content, or both, or neither. Marr
was not concerned at all himself with the issue of the intentional nature
of the primitives of this theory. But the depth of his methodological
comments and asides has left us with an embarrassment of riches when
it comes to possible interpretations of his theory. This is not simply an
indeterminacy about what Marr meant or intended, but re¬‚ects a similar
indeterminacy within the computational approach to vision itself, and, I
think, in computational psychology more generally.
With that in mind I shall turn now to Egan™s claim that the theory is not
intentional at all. This is a minority view of Marr™s theory that has brought
some incredulous stares in light of the constant references within Marr™s
work to notions of representation. How could Egan™s claim be squared
with such appeals? Of interest in this connection is Chomsky™s endorse-
ment of Egan™s interpretation of Marr. Speaking of Ullman™s studies of
the determination of structure from motion within a broadly Marrian
framework, Chomsky says

The account is completely internalist. There is no meaningful question about
the ˜content™ of the internal representations of a person seeing a cube under the
Representation, Computation, and Cognitive Science 159

conditions of the experiments, or if the retina is stimulated by a rotating cube,
or by a video of a rotating cube; or about the content of a frog™s ˜representation
of™ a ¬‚y or of a moving dot in the standard experimental studies of frog vision.
No notion like ˜content™ or ˜representation of™ ¬gures within the theory, so there
are no answers to be given as to their nature.20

On Chomsky™s view, the nonintentional interpretation of such theories of
vision is made plausible by the claim that “content” and “representation
of” are terms that are neither de¬ned within such theories nor with a clear
meaning within them. The theories are nonintentional in the relevant
sense, and thus it is a moot question whether they operate with a narrow
or a wide notion of content. As Chomsky continues,

The same is true when Marr writes that he is studying vision as a ˜mapping from
one representation to another, and in the case of human vision, the initial repre-
sentation is in no doubt “ it consists of arrays of image intensity values as detected
by the photoreceptors in the retina™ (Marr 1982:31) “ where ˜representation is
not to be understood relationally, as representation of.™

We will return to these points in assessing Egan™s interpretation below.21
At the heart of Egan™s view of Marr is a particular view of the nature of
Marr™s computational level of description. Commentators on Marr have
almost universally taken this to correspond to what others have called
the “knowledge level” or the “semantic level” of description, that is, as of-
fering an intentional characterization of the computational mechanisms
governing vision and other cognitive processes. Rather than ignoring
Marr™s computational level, Egan rejects this dominant understanding
of the computational level, arguing instead that what makes it a compu-
tational level is that it speci¬es the function to be computed by a given
algorithm in precise, mathematical terms. That is, while this level of de-
scription is functional, what makes it the ¬rst stage in constructing a
computational theory is that it offers a function-theoretic characterization
of the computation, and thus abstracts away from all other functional
characterizations.22
Thus, while vision might have all sorts of functions that can be spec-
i¬ed in language relatively close to that of common sense “ it is for
extracting information from the world, for perceiving an objective world,
for guiding behavior “ none of these, in Egan™s view, form a part of Marr™s
computational level of description. Given this view, the case for Marr™s
theory being individualistic because computational follows readily:

A computational theory prescinds from the actual environment because it aims to
provide an abstract, and hence completely general, description of a mechanism
Individualism and Externalism in Cognitive Sciences
160

that affords a basis for predicting and explaining its behavior in any environ-
ment, even in environments where what the device is doing cannot comfortably
be described as cognition. When the computational characterization is accompa-
nied by an appropriate intentional interpretation, we can see how a mechanism
that computes a particular mathematical function can, in a particular context,
subserve a cognitive function such as vision.23

According to Egan, while an intentional interpretation links the compu-
tational theory to our common sense“based understanding of cognitive
functions, it forms no part of the computational theory itself. Egan™s view
naturally raises questions not only about what Marr meant by the com-
putational level of description but more generally about the nature of
computational approaches to cognition.
There are certainly places in which Marr does talk of the computational
level as simply being a high-level functional characterization of what vision
is for, and thus as simply orienting the researcher to pose certain general
questions. For example, one of his tables offers the following summary
questions that the theory answers at the computational level: “What is
the goal of the computation, why is it appropriate, and what is the logic
of the strategy by which it can be carried out?” Those defending the
claim that Marr™s theory is externalist have typically rested with this broad
and somewhat loose understanding of the computational level of the
theory.24
The problem with this broad understanding of the computational
level, and thus of computational approaches to cognition, is that while it
builds a bridge between computational psychology and more folksy ways
of thinking about cognition, it also creates a gap within the computa-
tional approach between the computation level and the algorithmic level.
For example, if we suppose that the computational level speci¬es simply
that some visual states have the function of representing edges, others
the function of representing shapes, and so on, there is nothing about
such descriptions that guides us in constructing algorithms that gener-
ate the state-to-state transitions at the heart of computational approaches
to vision. More informal elaboration of what vision is for, or of what it
evolved to do, do little by themselves to bridge this gap.
The point here is that computational speci¬cations themselves are a
very special kind of functional characterization, at least when they are
to be completed or implemented in automatic, algorithmic processes.
Minimally, proponents of the broad interpretation of computational ap-
proaches to cognition need either to construe the computational level as

<<

. 26
( 62 .)



>>