Minds and Computers : An Introduction to the Philosophy of Artificial Intelligence

partly (although not entirely) informed by beliefs concerning the
nature of mental representation.
18.1 INTENTIONALITY
Intentionality is the technical philosophical term for the representa-
tional nature of mental states. Intentional states are those which are
about something, which represent something.
181

---

The terms ‘intentionality’ and ‘intentional’ are not to be confused
with the verb ‘intend’ and its cognates. Whether or not a state is inten-
tional, in the technical philosophical sense, has nothing to do with it
being intended by some agent. Rather, a mental state is intentional
just in case it is about something. Intentionality is the property of
mental states such that they are directed towards an object of repre-
sentation (a thing which is represented).
It is mental representations which are the primitive bearers of
intentionality. Our mental states are intentional by virtue of having
mental representations as constituents. So, for instance, my belief that
‘my dog is a fine companion’ involves, inter alia, a token mental rep-
resentation of ‘my dog’. My belief is about my dog by virtue of having
a constituent mental representation which is about my dog. My
mental representation of ‘my dog’ is directed towards (about) its
object of representation – namely, my dog.
This brings us to one of the important features of mental repre-
sentations I want to highlight here. Mental representations are cat-
egorial. My mental representation of ‘dog’ picks out all and only dogs.
In other words, it serves to categorise – in my mental life – those things
which I take to be dogs and distinguish them from those things which
I take not to be dogs. Similarly, my mental representation of ‘brown’
picks out all and only the things I take to be brown and my mental
representation of ‘chair’ picks out all and only the things I take to be
chairs.
The other important feature of mental representations I want to
highlight is that they are compositional. Mental representations
compose into more complex mental representations. Given, for
instance, my possession of mental representations of ‘brown’ and
‘dog’, I need nothing further to compose the more complex mental
representation of ‘brown dog’. This more complex mental represen-
tation picks out all and only the things I take to be brown dogs.
This compositionality of mental representations allows for one part
of an account of how it is that mental representations are conferred
with their intentional content. Complex mental representations
inherit their intentionality from the primitive mental representations
of which they are composed.
The crucial – and most di
fficult – question to answer, however, is
how it is that primitive mental representations are conferred with their
intentional content. How is it that our atomic mental representations
come to be about their objects of representation? In other words, what
is the nature of the relation between mental representations and the
(categories of) objects they represent?
182
  

---

18.2 CATEGORIES AND CONTENT
To give an account of the relation between mental representations
and their intentional objects is to give part of a semantics for mental
representation. There are numerous theories of the semantics of
mental representation but I’m not going to give a balanced exposition
of the available theories here. Instead, I want to give just the barest
sketch of two kinds of theories.
On one hand we have theories according to which mental represen-
tations are essentially discrete. On the other hand we have theories
according to which mental representations are essentially interre-
lated.
Theories of this first kind fit well with the symbolic paradigm in
artificial intelligence research. According to this kind of theory,
mental representations are symbols.
Theories of the second kind fit well with the connectionist para-
digm in artificial intelligence research. According to this kind of
theory, mental representations are distributed patterns.
A commitment to symbolic representation, on the one hand, or
distributed representation on the other, brings with it a raft of corol-
lary commitments. These include commitments concerning the mech-
anisms by which representations are conferred with their intentional
content, the nature and structure of the categories represented and
the ways in which mental representations interact.
Proponents of symbolic representation take representations to be
essentially discrete in a number of ways. The mechanism by which
symbols are conferred with their content is understood as some kind
of direct relation between tokens of the symbol and objects of repre-
sentation. Crucially, this mechanism is such that the content of a
symbol in no way depends on the content of other symbols. Each
symbol is discretely conferred with its intentional content.
Furthermore, symbolic representations are understood to remain
discrete in their interactions with other representations. The compos-
itionality of mental representation is understood to be simple syn-
tactic concatenation. When symbols compose to give more complex
representations, each symbol always brings the same content to the
complex in which it participates. In other words, the content of
symbols is taken to be contextually insensitive.
It is a further feature of symbols that their presence is binary – a
symbol token is either present or not. If a symbol token is present, it
is fully present and if it is absent, it is fully absent. Symbols, if you
will, are either on or o
ff, with no scope for anything in between. This

183

---

binary nature of symbolic representation has implications pertaining
to the nature of the categories which they represent.
If mental representations are symbolic then the categories which
they represent must admit of sharp borders and no internal structure.
In other words, if mental representations are symbolic, then the cat-
egories they represent are like boxes – objects are either in the cat-
egory or not and there are no better or worse cases of category
membership.
Proponents of symbolic representation, to recap, take the content
conferring mechanism on representations to be discrete, the cat-
egories they represent to be binary and unstructured, and the com-
position of mental representation to be contextually insensitive
syntactic concatenation.
Advocates of distributed representation, on the other hand, have a
di
fferent understanding of each of these elements of the semantics of
mental representation.
18.3 SYMBOLS AND PATTERNS
Theorists who endorse an account of mental representations as dis-
tributed patterns understand representations to be essentially interre-
lated. The semantics of mental representation that such a theorist will
advance are such that the mechanism by which content is conferred
on a representation is essentially mediated by relations to other rep-
resentations.
There are a number of ways to flesh out this mediation but the
mechanics of particular theories need not concern us here. What is
important for our purposes is the commitment to the interrelated
nature of mental representations, in stark opposition to the view held
by proponents of symbolic representation. The way in which a mental
representation is conferred with its intentional content, according to
distributed accounts of representation, is essentially bound up with
the way in which other mental representations are conferred with
their intentional content.
The composition of distributed mental representations is also
taken to be somewhat more complex than syntactic concatenation. It
is crucial that there be an account of the composition of mental rep-
resentation that is su
fficient to secure the systematicity of cognition,
since this is generally held to account for the productivity of the lin-
guistic facility.
Those who endorse a view of mental representation as distributed
understand the composition of representations to be the highly
184
  

---

complex interaction of patterns of activation in a distributed
network. Precisely what this means will become clearer in the follow-
ing chapter when we discuss artificial neural networks. For present
purposes, it su
ffices to appreciate that the way distributed representa-
tions compose is contextually modulated. In other words, the content
that a particular representation brings to the complex representation
in which it participates will vary in a way that is dependent on the
other particular representations also participating in the complex.
It is a further feature of distributed representations that they can
be partially tokened. Since token representations are taken to be pat-
terns of activation widely distributed across an interconnected
network of nodes, these patterns can be partially activated. Again,
precisely what this means will become clearer in the following chapter.
The important point here is an appreciation of the implications that
the possibility of partial tokening of mental representations has with
respect to the nature of the categories represented.
If mental representations are distributed patterns which can be