The paper presents a paradoxical feature of computational systems that suggests that computationalism cannot explain symbol grounding. If the mind is a digital computer, as computationalism claims, then it can be computing either over meaningful symbols or over meaningless symbols. If it is computing over meaningful symbols its functioning presupposes the existence of meaningful symbols in the system, i.e. it implies semantic nativism. If the mind is computing over meaningless symbols, no intentional cognitive processes are available prior to symbol grounding; therefore no symbol grounding could take place since any such process presupposes intentional processes. So, whether computing in the mind is over meaningless or over meaningful symbols, computationalism implies semantic nativism.
We will discuss an aspect of the problems a digital computational system has to
acquire meaningful symbols and what these problems mean for a computational theory
of the mind, in particular its relation to semantic nativism. The classical
computational theory of the mind, or “computationalism” for short, holds that the mind is a
digital computer, in particular that it is a computational information processor. The
view of the mind as a computer, especially as a computer processing symbols
according to rules, is the basis of classical cognitive science and artificial intelligence. As
Fodor puts it: “The cognitive science that started fifty years or so ago more or less
explicitly had as its defining project to examine a theory, largely owing to Turing, that
cognitive mental processes are operations defined on syntactically structured mental
representations that are much like sentences.” [
Computationalism is typically (but not necessarily) a version of the view of mental
states as physical states with a specific causal functional role, as proposed by the
earlier Putnam. If the mind is described not at the a basic physical level, but described
at the level of these functional roles and if these are taken as realizations of a Turing
machine, as computational states, then we have the theory commonly known as
Machine Functionalism, which includes the thesis of the necessity of computing for
mentality: “Mentality, or having a mind, consists in realizing an appropriate Turing
machine” [
We shall only discuss computationalism in the sense that computation is sufficient for mental states and that it is the cause of mental states in humans, not in the stronger sense that computation is necessary and sufficient (or only necessary). It must be noted, however, that computationalism is not just the weak thesis that some or all mental processes can be modeled on a digital computer. If a hurricane can be modeled on a computer, this is not to say that the hurricane is a computational system. (NB, it is doubtful whether such modeling is strictly speaking possible on a digital computer, since a hurricane is not a discrete state phenomenon). Despite the distinction of computationalism from this weaker thesis, there is the possibility, however, that minds might be special cases such that modeling a mind actually is producing a mind— given that it has sufficient functional properties [e.g. 5].
Computationalism directly implies the possibility of strong Artificial Intelligence:
“… computers can think because, in principle, they can be programmed with the same
program that constitutes human thought.” [
The notion of computing used her is the classical one as defined by Turing, i.e.
computing means a mechanical procedure (e.g. the manipulation of symbols)
according to algorithms, i.e. explicit non-ambiguous rules that proceed step by step and that
can be carried out in finite time, leading to a definite output—what is also called
“effective computing.” The Church-Turing thesis says that a Turing machine can
compute all and only the effectively computable functions. (I ignore the theoretical
possibility of hypercomputing in this paper, but see [
The main theoretical options within computationalism depend on whether the symbols on which the computer operates (and that constitute its program) are meaningful or not. I shall call the option of operating on meaningful symbols “Language of Thought Computationalism” or LOCO.
Symbol Grounding in Computational Systems: A Paradox of Intentions 3
The tradition of Fodor’s “Language of Thought” focuses on “cognition” or, even
more narrowly, “thought”, and it claims that thinking is computing over mental
representations. Fodor’s slogan could be said to be “There is no computation without
representation” [
The symbols are also taken to be closely related to natural language concepts, in
what Smolensky calls the “Newell/Simon/Fodor/Pylyshyn view”: the programs of this
computational system “are composed of elements, that is, symbols, referring to
essentially the same concepts as the ones used to consciously conceptualize the task
domain.” [9, cf. 10]. One consequence of this approach is “the view that propositional
attitudes (such as believing, noticing, preferring) are to be regarded as computational
relations to semantically valuable representations that are encoded in the brain or
other hardware of the thinker.” [
Fodor’s emphasis on the syntactical nature of the computational process should not be
taken to mean that his position is anything other than language of thought
computationalism. It just so happens, that thinking is a computation over symbols that are
representations:
“First, all mental processes are supposed to be causally sensitive to, and only
to, the syntax of the mental representations that they are defined over; in
particular, mental processes aren’t sensitive to what mental representations mean.
This is, I think, at the very heart of the Classical [Fodor’s] account of
cognition.” [
This brings us to the problem. How is it possible that these symbols of a
computational system have meaning? Fodor himself appears to see that this is problematic, at
least sometimes: “How could a process which, like computation, merely transforms
one symbol into another guarantee the causal relations between symbols and the
world upon which … the meanings of symbols depend?” [
The situation invites a very short line indeed: If language of thought computationalism is the manipulation of meaningful symbols, then the functioning of the language of thought (the “cognition” or the “thinking”) presupposes the existence of meaningful symbols in the system. In other words, the system must have meaningful symbols before the language of thought can function. The acquisition of these meaningful symbols can thus not be the work of a language of thought.
So, if a newborn child’s mental activity is within language of thought, then a child must be born with meaningful symbols: language of thought computationalism presupposes meaningful symbols. Fodor himself has been supporting the idea of innate meaning for some time, of course, but many in the field want the language of thought computationalism without the nativism. Nativism is typically taken as optional but as the ‘short line’ shows, it is not. This “short line” is a simple argument against language of thought computationalism without semantic nativism—an argument we lack so far as far as I can see [see 1, 13, 15, 16].
Given the problem described in the above “short line”, it may be plausible to revert to a more modest version of computationalism: Mental computation is (or could be) purely syntactic. Of course, this does not exclude that the symbols could be interpreted by some observer; it just says that they have no meaning for the system. At first glance, this is what is the case with any conventional digital computing machine: For example, the operation of a set of switches that constitute an XOR-gate could be interpreted as be doing a logical operation, or as computing an addition, or as doing various other things. (The logical gates for exclusive or are the same as those for binary addition plus a “carrying over” of surpluses to the next digit.) The switches have no meaning for the system itself. When my pocket calculator displays the output “844$” or my washing machine displays “End”, this means something to me, but not to the computer.
Symbol Grounding in Computational Systems: A Paradox of Intentions 5
In order to understand the proposal of computationalism without presupposing meaning in the system, it is useful to gain a deeper understanding of what a computer, really does. The main characteristic of a digital computer is that is algorithmic. Any calculator can “carry out” a particular algorithm (and mechanical calculators were already constructed in the 17th Century by Schickard, Pascal and Leibniz). A computer, however, is programmable, that is which algorithm it carries out can be changed. The universal Turing machine is a model for a computer that can run any program, essentially by giving numbers to all the other simple Turing machines that can compute only one algorithm.
To understand computation, it is important to see that we can describe a computer on three (plus) levels of description:
Physical level: Some physical objects such as toothed wheels, holes in cards, states of switches, states of transistors, states of neurons, etc. are causally connected with each other—such that a state of one object can alter the state of another.
Syntactical level: The physical objects are taken to be tokens of a type (e.g. charge/no charge) and are manipulated according to algorithms. These algorithms are also stored and changed in the computer via some set of physical tokens (typically the same set). The manipulation follows the algorithms and only concerns these tokens as tokens, not their physical realization or their interpretation; it is “purely syntactical.” To do this, the computer needs to recognize each token as of a type, as a basic symbol for this system, e.g. a 0 or 1 in a binary system.
What I call the syntactical level could also be called the “form” of a computational
procedure. This way of talking is aptly criticized by [
Horowitz makes the related proposal that we need “computational externalism
without relying on semantic externalism” [
Symbolic level: The physical objects that are manipulated on the logical level are taken to represent; they are (parts of) letters, numbers, words, images, vectors, concepts, … One could thus have one algorithm (on the syntactical level) that carries out several functions (on the symbolic level).
Piccinini’s terminology, who discusses the problem of how to “individuate
computational states” [
I propose to have “3+” levels rather than “3” because each of the symbols on the symbolic level can symbolize something else in turn. Accordingly, one might distinguish several further levels within the symbolic level when describing a computational cognitive system, for example, the distinction between nonconceptual content and conceptual content, or the distinction between symbols and concepts [for the latter, see 21, ch. 7].
If we now describe a conventional van Neumann machine, e.g. a PC, at its syntactical level, rather than at its physical (realization) level, we will see basic operations on bits of main memory such as read (is this bit on or off?) and write (to this bit). These operations are combined by building in logical (Boolean) switches where one bit takes a particular state, given the state of two other bits. With the help of such switches, one can construct algorithms of switching patterns that perform particular tasks on the symbolic level, e.g. compare, add, … The computing process is a long sequence of such basic operations resulting in a memory state. Note that it is irrelevant for the syntactical description of the computer how a particular operation is carried out—one way to see this is to conceive of the computer as operating a Turing machine [see e.g. 22].
After this initial clarification we can return to the proposal of syntactic computationalism. Some have claimed that this is per se impossible, that there could not really be a computing system, without any meaningful symbols. One prominent objection is that the system must be able to carry out programs, programs that are themselves encoded in symbols, and typically stored in memory. Does this not require following rules and understanding at some level? For example, in Searle’s famous computation in the “Chinese Room” [23, cf. 24], Searle sits in the room and manipulates Chinese symbols according to manipulation instructions given in English: A language that he understands!
John Haugeland claims that in any computing system there are primitive
operations of which the system knows how to carry them out [
If this was right, in any computing system we would be back at our original problem: If there are “meaningful primitives” in any computing machine—where do they get their meaning? Our ‘short line’ would show that computationalism implies semantic nativism, generally. Or rather, semantic nativism must be true for any computer, given that we have working computers. All our computers would already have meaningful symbols built in!
I think it will be apparent form the discussion of descriptive levels above, that purely syntactic machines are in fact possible, however. We just need to be more careful when we say that the system “follows rules”, or “executes programs.” Wittgenstein famously distinguished between following a rule and acting according to a rule—and only the former requires that one understands the rule (gives it an interpre
Symbol Grounding in Computational Systems: A Paradox of Intentions 7 tation). The computer does not literally follow a rule. Being in a particular state, given a particular input, it will perform a series of steps (e.g. switches) and produce a particular output, a memory state. The same happens when it is programmed, i.e. its switches are set (this even happens in the same central memory, in the case of a ‘stored program’ von Neumann machine). This is a purely causal, mechanical procedure that requires no understanding of a rule. It is no different from a can vending machine taking a particular input (my coins and my pressing a button), processing, and producing a particular output (the can).
The computing machine is just constructed in such a way that it will mechanically do what we call “carrying out a program”, on the logical or even the symbolic level. We can describe the computer as “following a rule” and some of its states as “symbols” but that is entirely irrelevant to its functioning. A computer can be described on the symbolic level, but it must not have such a level. It may also, to repeat, be described differently on the symbolic level. The widespread resistance to calling computing “purely syntactical” [e.g. 11, 22, 29, 30] perhaps due to the fact that this process is, of course, causal. It is not so much a formal procedure, but rather the syntactical properties of a physical procedure. On the syntactical level, one can say that the computer operates on meaningless symbols with programs that are meaningless to it.
Accordingly, the solution to symbol grounding cannot be to give basic rules, as does for example Hofstadter in his discussion of the matter. For his MU and MIU systems you assume that rules have meaning [31, chs. I & II, pp. 170, 264]. If you do not, then you have to postulate that “absolute meaning” comes about somehow by itself, in “strange loops” (ch. VI and passim).
So, how does syntactic computationalism, thus understood, fare with our problem of symbol grounding? The problem for a computationalist is that she has to construct a causal chain that does not involve any mental process at any stage that is other than purely syntactic. Meaning-involving processes such as attention, object tracking, object-files, interest, intention, etc. are not permitted.
Let us look at some lessons from history to understand the difficulty: I take the discussion about the so-called “causal theory of reference”, originally developed by Putnam and Kripke in the early 1970ies, to have shown two things:
A) We want to grant causal connections between tokens of some kinds of symbols and their reference a role in the determination of the meaning of the symbols—in particular, we want to do this in the case of natural kind terms, such as “gold”, where the stuff they refer to, the element gold, plays a role in the determination of what counts as gold and what does not. This is what Putnam called the “contribution of the environment.” I say, “We want to grant” because it is important to see that Putnam’s and Kripke’s discoveries are discoveries about our linguistic intuitions.
B) The causal relations between, for example, the tokens of the word “gold” and the element gold are immensely complex and it is extremely hard to figure out the particular causal relation that should connect a particular token to its referent. A given token stands in any number of causal relations and none of these by itself distinguishes itself as the right one (for example, “gold” does not refer to jewelers shops or to chemistry textbooks or to metal or to undiscovered fake gold). What we need is a notion of “explanatory cause”, the cause that is relevant for our explanatory intentions.
What is relevant here is not so much semantic externalism (that has lead to
externalism about mental states) but Putnam’s later critique of his own earlier causal
theories of reference. This critique shows that a successful story of the causal relation
between my tokens of “gold” and gold has to involve my desire to refer to that
particular metal with that particular word. Putnam has tried to show this in his
modeltheoretic argument [
The Putnam/Kripke story shows that the causal relation of a linguistic symbol to its referent must involve the intention of speakers to refer to a specific object or kind: otherwise it is underdetermined due to the multiplicity of causal chains.
Fodor himself seems to see an issue when he argues against language acquisition
by non-linguistic thinkers as follows: “Plausibly, for example, learning English
requires learning that the form of words ‘it’s raining’ is properly used to communicate
the thought that it’s raining. How do you learn that sort of thing if you have the kind
of mind that can’t, even in principle, think about thoughts?” [
So, the problem is, how can a system acquire meaningful symbols without making use of cognition? Could there be a theory of language acquisition (or machine learning) that assumes a language can be learned by a system that has no cognitive processes? I propose that to develop such a theory is more than just a challenge: it cannot be done.
The same point can be illustrated in the terms used in Searle’s “Chinese room argument” [23, cf. 24]. Searle’s central notion is “understanding” (of Chinese and of stories) and he claims, 1) that the symbol manipulator in the Chinese room should not be said to understand Chinese by virtue of his handling the symbols correctly and thus producing correct output, also that he has no chance of learning Chinese [both of this everybody agrees with], 2) that the whole system containing the Chinese room, with manipulation manuals and all, cannot be said to understand Chinese [the “systems reply”], not even if “sensory organs” (cameras, microphones, etc.) are added [the “robot reply”], since these supply “just more Chinese.” He sometimes expresses this as saying that the system has syntax but no semantics for its symbols: that symbols in a system cannot acquire meaning due to mere symbol manipulation.
Symbol Grounding in Computational Systems: A Paradox of Intentions 9
As several people have pointed out, 2) does not follow from 1). This does not mean that his argument fails, however. The upshot of the argument is, in my view, that Searle sets the task to explain how a system can understand Chinese given that the central symbol manipulator does not. After the Chinese Room Argument the belief that a symbol manipulating system can “understand” is in doubt and would require positive support.
Searle’s claim is that he cannot learn Chinese by manipulating the symbols in his room, even if he tries hard—and then he expands this point to the whole system. But he already grants too much: Searle in the room does understand the symbols in the instructions for manipulation, wants to learn Chinese, knows that Chinese is a language, that some of its symbols refer and which world they refer to. None of these is given in an actual purely syntactic computational system. Given that there is literally no understanding, desire and knowledge in the actual Chinese Room of a syntactic system (there are no intentional states), there is even less reason to believe that there is in the whole system.
The argument presented above thus goes some way towards closing the gap in Searle’s argument by explaining why symbol manipulation, even under causal interaction with the environment, cannot produce intention. The system will not acquire meaningful symbols because it lacks everything necessary, specifically it has no desire to do so (it has no desires directed at anything). The situation is thus worse than in Searle’s “Chinese Room”, where Searle tries to show that an intelligent agent operating a purely syntactical system cannot acquire meaning. We only need to claim that a purely syntactical system itself will not acquire meaning—even if it could.
On a cautionary note, just like Searle, we do not claim to have found any bounds as to what can be done with purely syntactic computing. Clearly, advanced AI systems (and perhaps “lower” animals) have achieved impressive feats without the “meaningful symbols” we have been asking for and which humans surely possess.
This look into the Chinese room might leave a paradoxical air; one might wonder what that magical bit is which allows humans and other animals what computers cannot have. My suggestion here is that this bit has to be something that is not computational—and I think desire is a good candidate.
What we have seen so far is that: 1) A language of thought computational system presupposes innate meaning, 2) A purely syntactical computational system is possible, 3) A purely syntactical computational system could only acquire meaning if that process does not involve any mental states with intention (e.g. desires, beliefs, attention, …).
What we have not seen is whether there is another version of computationalism that could save the day. Perhaps there is computation without symbols or there is information processing in ways other than computing? Let us take a brief look at the options.
Searle has repeatedly said that whether a system is a computer or not depends on its interpretation by some observer, a syntactic property is an observer-relative notion This is why he comes to the prima facie surprising conclusion that “The brain is a computer, in the sense that it instantiates computer programs...” because “everything is a digital computer at some level of description” [cf. 30, 37].
Whether this view is true or not (I tend to think it is not [cf. 20]), as Searle knows, this makes computationalism vacuous. Clearly, computationalism cannot be the claim that, if an observer likes to see it that way, the brain is a computer, and so is a train, a tree or a bumblebee.
There are cognitive scientists that use the word “computational” in a much weaker
sense than the one defined above—in fact, the plethora of definitions is depressing: I
counted 9 different ones, most of which are obviously either too narrow or too broad,
in a recent exchange between Pinker and Fodor [
3) Having “the architecture of a Turing machine or some other serial, discrete, local processor” [attributed to Fodor by 15]. False attribution, since in 2000, Fodor did not mention the possibility of other processors. Suggests that “architecture” means physical setup (tape and reader), after all—see problems in 2).
4) Being ‘Turing-equivalent’, in the sense of ‘input-output equivalent’ [
6) Being supervenient “on some syntactic fact or other”—“minimal CTM” [
8) “In this conception, a computational system is one in which knowledge and
goals are represented as patterns in bits of matter (‘representations’). The system is
designed in such a way that one representation causes another to come into existence;
and these changes mirror the laws of some normatively valid system like logic,
statistics, or laws of cause and effect in the world.” [
9) “…human cognition is like some kind of computer, presumably one that engages
in parallel, analog computation as well as the discrete serial variety” [
One prominent idea is that computing is somehow “information processing.” [cf. 38] But information processing could take many forms, some of which are not computational. There are many systems that could be used to compute but should not be
Symbol Grounding in Computational Systems: A Paradox of Intentions 11
called a computer. Dynamical systems in the sense of van Gelder [
I would therefore make the terminological suggestion to distinguish between “computationalism” and “information processing” as paradigms for cognitive science.
The pages above present a reason to believe that the mind is not a computational
system, unless it has semantics built in. However, there is still good reason to think
that some parts of the human mind are computational, even if the problems explained
show that it is not only that. Perhaps the picture that emerges is that of a hybrid and
modular mind where some modules are computational but many are not [