(Presented at the 1997 Convention of the Canadian Psychological Association)
Nearly 40 years of work in this research program seems to have led to a dead end. Physical symbol systems such as these cannot be made powerful enough or flexible enough to plausibly simulate human-level cognitive competencies. In addition a strong new competitor has come along to challenge the physical symbol system hypothesis: parallel distributed processing, or as I will call it here, connectionism.
In connectionist models the individual physical tokens do not typically represent anything at all. The representation is distributed over the activity of the entire network. For instance, if a connectionist model were instantiating a proposition such as "If A then B," there would be no individual part representing A, no other part representing B, and no third part representing the "If then " relation. The proposition would be represented "holistically," as it were, across the whole network. Instantiations of different propositions would consist in different patterns of activity spread across the very same network. That is to say, the representations of various states are stored in superposed form-one on top of the other, as it were-in one and the same physical system.
Functionalists such as Fodor and Pylyshyn (1988) have argued that it is precisely for this reason that connectionist networks cannot be take seriously as theories of cognition. Nevertheless, connectionist networks have been shown to learn domains of knowledge more easily, deal with incomplete information more flexibly, and generalize their previous experience to new cases more realistically than the symbolic models that went before them. These successes have led to their enormous increase in popularity over the last decade, despite the theoretical problems to which Fodor and Pylyshyn object.
In the mid-to-late-1980s, many connectionists were content to say that symbolic models are reasonable approximations of cognitive activity, though one would have to look to the connectionist level if one wants to know what is really going on. In the 1990s, however, many connectionists have shown less and less tolerance for the symbolic paradigm, opting for an eliminative materialist standpoint. This is the position that there are no beliefs and desires; that such "folk psychology" is just a bad old theory of mind to be tossed out with the theories of phlogiston and caloric.
Ramsey, Stich, and Garon (1991) have gone one step further, arguing that acceptance of the connectionist paradigm logically commits one to the acceptance of eliminativism as well. Their argument runs roughly as follows: Because all the beliefs and desires of the model are stored in superposed form across the entire network, and because the activity of the whole network must go into any given behavioral output, no one belief or desire, or small subset of beliefs and desires, can ever be said to be causally responsible for a given output; they are in a sense all responsible for any output. Thus, no finite set of beliefs or desires can be said to explain any behavior: for example, Cornelius didn't go into the bar because he wanted a beer, only the organization of his entire cognitive network can be said to have been causally responsible. Thus, explanation of behavior will only be found in an analysis of the network's structure. Since beliefs and desires can never explain behavior, Ramsey, Stich, and Garon argue, what scientific purpose can they possibly serve? They should simply be eliminated like other hypothetical constructs that have turned out to have no explanatory function. Thus, they conclude, the adoption of a connectionist approach to cognition logically commits one to eliminativism with respect to beliefs and desires. It is interesting to note here that the Ramsey-Stich-Garon argument does not eliminate beliefs and desires tout court; it only eliminates them from what they take to be the causal chain. Exactly how one determines what is and is not part of the causal chain is not made explicit. It seems that it is simply assumed that it runs straight from the physical activity of the network to the physical activity of the organism in which the network is housed.
Some philosophers and connectionists have resisted the Ramsey-Stich-Garon conclusion. Philosopher Andy Clark (1993), for instance, concedes the point that beliefs and desires do not literally cause our behavior, but argues that explanatory power is not always grounded in causal power. Beliefs and desires still provide us with counterfactual-supporting accounts of behavior, even if they do not play a role in the causation of that behavior. Just as I can say that the match lit because it was struck (and that it wouldn't have lit if it had not been struck) even though this does not describe what Clark calls the "causal microstructure" of combustion, so I can say that Cornelius went into the bar because he wanted a beer (and that he wouldn't have gone in if he didn't) even though this does not describe the causal microstructure of cognition.
This is a very dangerous move, however. First, most people would be perfectly happy to accept the match-lighting example as a straightforwardly causal one, even if it fails to describe the physical microstructure of the situation. Second, no one is very sure what gives counterfactual-supporting accounts their explanatory power. It is only an observation that they seem to have such power, not an explanation of that power. Second, one of the most influential accounts of causation-viz., that of David Lewis (1973)-amounts to the argument that accounts of causal relations between events are little more than true counterfactual conditionals involving the propositions that describe those events. Thus, even if Lewis' account is wrong, it would seem incumbent upon Clark to give us his theory of causation first. (Actually he starts to in his mention of causal microstructure, but surely he can't believe that microstructure is where causation is always located: square-pegs cannot fit into round holes of the same area, but this is not caused by their microstructure.)
Smolensky, a founding connectionist who was once (1988) among those who said that symbol-level accounts of cognition are only approximate accounts, has recently (1995a) modified his position significantly. Unlike many of his connectionist colleagues, he has explicitly rejected eliminativism-particularly the argument of Ramsey, Stich, and Garon (Smolensky, 1995b). He now argues that although the casual account of cognition is pitched at the connectionist level, the true explanation of behavior is to be found at the symbolic level. That is, if you want to know how Cornelius went to the bar, look to the connectionism. If you want to know why, then you should look to his beliefs and desires. The problem here should be immediately apparent. If explanations of behavior are to be pitched at an entirely different level of analysis from the causal story, what exactly is it that makes the explanatory level explanatory? Smolensky does not say.
People on both sides of this debate seem to implicitly hold a theory of causation in which physical bits of matter must literally be seen bumping into one another. This is quite naïve, however. Indeed, it seems to derive almost directly from the Hume's theory that cause is nothing more than regularly observed concomitance between objects. That is, we know the red billiard ball caused the blue billiard ball to move only because we saw the red one move, then make contact with the blue ball, and then we saw the blue ball move, and that we have seen this pattern regularly in the past. The lie was put to Hume's regularity hypothesis by his contemporary, Thomas Reid, who noted that day follows night with a regularity unmatched by most any other event, but no one has ever supposed that one causes the other.
Many have tried to salvage Hume's regularity hypothesis, and it has its descendants today, but many others have argued that it is hopeless and must be supplanted by something that takes into account causation's modal nature; that is, its dependence on what is necessary and sufficient to make something happen. The best known of these theories is Mackie's (1965) proposal that the cause C, of an event E, is an insufficient but necessary part of an unnecessary but sufficient condition for E's occurrence (INUS condition). For instance, when we say that a particular short circuit caused a particular fire, we are saying that the short circuit although Insufficient to start a fire by itself, was a Necessary part of a combination of conditions, such as requisite oxygen and combustible material, that could start a fire. Although this combination of conditions was Unnecessary to start the fire (e.g., it may have been started by a match) together they were Sufficient to start it.
This paradigm makes beliefs and desires perfectly good causal agents without their having to be ordinary physical objects that "bump" into one another: Cornelius' desire for a beer is an insufficient but necessary part of a set of conditions (e.g., his being able to move), that are collectively sufficient for his going to a bar. This set of conditions are themselves unnecessary, however. He may go to the bar simply to meet a friend.
While Mackie's theory seems to be a step down the right path, a lot of water has gone under the bridge in the last 30+ years and various problems with it have emerged. One of the most significant of these problems is that is that INUS conditions seem to capture objects that would appear to be of the wrong scale. For instance, let us say that the short-circuited wires that caused the fire were, as a matter of fact, made of copper. The short circuiting of these copper wires constitutes an insufficient but necessary part of an unnecessary but sufficient condition for the fire, but we would be hesitant to say that the fire was caused by a short circuit in copper wires. This claim seems to miss the point somehow. The details of the material composition of the wires (assuming they can conduct electricity) are irrelevant.
Philosopher Stephen Yablo (1992) has recently tried to address this problem by substantially reworking the conditions of causation to capture the right scale. As he puts it, causes must be commensurate with their effects. Yablo argues that:
Yablo also argues that effects bear parallel obligations to their causes:
Now let's return to the cause of our hypothetical fire. The short circuit in the copper wires is adequate for the effect of causing the fire, but it is not just enough. Adding the detail of "copper" is too much. "Copper wires" is an example of a C+. Any conductive wires at all short circuiting are enough. By contrast, the fire was contingent on the short circuit, but it did not require that the wires be copper. Here any conductive wires-copper or no-would count as a C- that would have been able to cause the fire, so the their copperness is a superfluous detail. In sum, it is most appropriate to say that the fire was caused by a short circuit tout court, not a short circuit in a certain kind of wire. It is interesting to notice here that the appropriateness of causal explanations does not seem to be a back-and-white affair. It seems to admit of certain "shades of grey." Saying that the fire was caused by a short circuit in copper wires is not out-and-out wrong; it just seems to be somewhat less correct-less exact-than saying simply that it was caused by a short circuit.
What impact does all of this have for a connectionist theory of mind? It seems to undercut the implicit theory of causation underlying both the eliminativist argument that the causal chain runs solely from the physical network to the physical behavior, and the Clark-Smolensky retort that even though beliefs and desires are not causal they can still somehow explain behavior. It turns out that beliefs and desires are casual, even if they are not straightforwardly physical. To wit: Cornelius' desire for a beer is adequate for his going into a bar, and it is also just enough. His having wanted, say, a Heineken would have been too much; a C+. This might have been an appropriate explanation for his wanting to go to a particular bar, known to serve Heineken, but not so appropriate to explaining his going into just any bar. By contrast, Cornelius would not have gone into the bar, other things being equal, if he had not wanted a beer. There is, of course, a problem of overdetermination of effect here. If he had, in the original case gone to the bar both because he wanted a beer and to see his friend, then he would still have gone even if he didn't want the beer. Overdetermination problems plague many theories of causation, however, and I do not propose to solve them here. The ceteris paribus clause blocks this problem for the time being, but must eventually be worked out. Some combination of Mackie's and Yablo's criteria might help here. Finally, the going in to the bar requires wanting a beer, again ceteris paribus. There may be other reasons for going into a bar, but these are not currently under consideration.
What I have tried to show here is that the implicit theory of causation underlying the debate among eliminativist and non-eliminativist connectionists is the source of the debate among them. The Ramsey-Stich-Garon argument tries to show that beliefs and desires do not play a role in the causal chain leading to behavior, but this does not fully eliminate them. Further, their argument about where the causal chain lies assumes a naïve theory of causation. Smolensky's and Clark's efforts to save beliefs and desire for the purpose of explanation, even if not causation, are also grounded in this same naïve theory. Beliefs and desires are explanatory of behavior precisely because they are causal, though you have to have a more sophisticated theory of cause than they do in order to see it.
Clark, A. (1993). Associative engines: Connectionism, concepts, and representational change. Cambridge, MA: MIT Press.
Fodor, J. A. & Pylyshyn, Z. W. (1988). Connectionism and cognitive architecture: A critical analysis. Cognition, 28, 3-71.
Lewis, D. (1973). Causation. Journal of Philosophy, 70, 556-567
Mackie, J. L. (1965). Causes and conditions. American Philosophical Quarterly, 2, 245-255.
Ramsey, W., Stich, S. P., & Garon, J. (1991). Connectionism, eliminativism, and the future of folk psychology. In W. Ramsey, S. P. Stich, & D. E. Rumelhart (Eds.), Philosophy and connectionist theory (pp. 199-228). Hillsdale, NJ: Lawrence Erlbaum.
Smolensky, P. (1988). On the proper treatment of connectionism. Behavioral and Brain Sciences, 11, 1-73.
Smolensky, P. (1995a). On the projectable predicates of connectionist psychology: A case for belief. In C. MacDonald & G. MacDonald (Eds.), Connectionism: Debates on psychological explanation (Vol. 2, pp. 357-394). Oxford: Basil Blackwell.
Smolensky, P. (1995b). Constituent structure and explanation in an integrated connectionist/symbolic architecture. In C. MacDonald & G. MacDonald (Eds.), Connectionism: Debates on psychological explanation (Vol. 2, pp. 223-290). Oxford: Basil Blackwell.
Yablo, S. (1992). Cause and essence. Synthese, 93, 403-449.