175 ©2007 The Institute of ~1md and Behavior, Inc. The Journal of ~'lmd and BehavIor Summer and Autumn 2007. V~llllme 28. Numbers 3 and 4 Page, 175-188 ISSN 0271-0137 Why History Matters: Associations and Causal Judgment in Hume and Cognitive Science Mark Collier University of Minnesota, Morris It is commonly thought that Hume endotses the claim that causal cognition can be fully explained in tetms of nothing but custom and habit. Associative learning does, of course, playa major role in the cognitive psychology of the Treatise. But Hume recog­ nizes that associations cannot provide a complete account of causal thought. If human beings lacked the capacity to reflect on rules for judging causes and effects, then we could not (as we do) distinguish between accidental and genuine regularities, lind Hume could not (as he does) carry out his science of human nature. One might reply that what appears to be rule-governed behavior might emerge from associative systems thar do not literally employ rules. But this response fails: there is a growing consensus in cognitive science that any adequate account of causal learning must invoke active, controlled cognitive processes. Keywords: David Hume, associative learning, rules for judging causes and effects Hume's account of causation is his most important legacy in the history of philosophy. His first major contribution to the philosophy of causation involves the definitional question: How should we analyze the concept of cau­ sation? Hume worries that we often use causal terms without the foggiest notion of what they mean. When philosophers are pressed to define causa­ tion, for example, they typically maintain that "A causes B if and only if A produces B or brings about B or necessitates B." But these causal terms are all synonymous, and thus we have been taken around in a circle (1739(1978, p. 157). Hume makes significant progress, therefore, when he offers a deflation­ ary analysis of causation in terms of invariable regularities. Requests for reprints should be sent to Mark Collier, Ph.D., Department of Philosophy, University of Minnesota, Morris, 600 East Fourth Street, Morris, Minnesota 56267. Email: mcollier@morris.umn.edu I _I 176 COLLIER WHY HISTORY MATTERS 177 Hume recognizes thar it is one thing to analyze what the term causation 2005b), how we make probabilistic inferences (Collier, 2005a), and why we means, however, and another thing to explain how we discover causes and believe in object permanence (Collier, 1999). effects. Hume's second major contribution to the philosophy of causation What about causal cognition? Does Hume really maintain that causal involves the epistemological question: How can we make causal inferences? thought can be fully explained in terms of associations? The traditional inter­ The definitional and epistemological issues are, of course, closely related: if pretation is that his associationism is intended to be "complete" (Pears, 1991, causes are constituted by constant conjunctions, then we can discover causes p. 70) and "comprehensive" (Jessop, 1966, pp. 46-47). But the truth is that by searching for invariable sequences. For example, every time the ignition Hume does not regard his custom-habit account as a finished product. He is key is turned, the car starts. Since one event always follows another, we can sensitive to the fact that it must be supplemented with a rule-based system in pronounce them cause and effect. By taking the mystery out of causation, order to paint a complete picture of human causal reasoning. There are two Hume has apparently taken the difficulty out of causal inference. main lines of support for this revisionary interpretation. The first involves This is precisely how Hume characterizes our everyday causal inferences. Hume's explicit remarks about experimental methods and the "rules for judg­ We ordinarily make causal attributions whenever we observe a frequent repe­ ing causes and effects." The second appeals to the actual experiments per­ tition between events (p. 87). Why do we make causal inferences in this way? formed by Hume in his science of human nature. Hume's answer is that these events have become associated in our imagina­ tions; in short, we have become conditioned to do so. Hume's Methodological Pronouncements When the mind ... passes from the idea or impression of one object ro the idea or belief Hume observes that genuine causal conditions are often surrounded by of another, it is not derermin'd by reason, but by certain principles, which associate "superfluous circumstances." together the ideas of these objects, and unite them in the imagination. (p. 92) In almost all kinds of causes there is a complication of circumstances, of which some are Let us call this the "custom-habit" account. It has several important features. essential and others superfluous; some are absolutely requisite to the production of the First, it does not postulate any explicit reasoning. We do not ordinarily make effect, and others are conjoin'd by accident. Now we may observe, that when these superfluous circumstances are numerous, and remarkable, and frequently conjoin'd with causal inferences through reflection, thought, or judgment; rather, our every­ the essential, they have such an influence on the imagination, that even in the absence day causal attributions depend solely upon associative propensities of the of the latter they carry us on to the conception of the usual effect .... (p. 148) imagination. Second, ordinary causal reasoning is automatic and implicit; when we approach a river, for example, we do not consciously deliberate For example, imagine that we repeatedly observe two potential causes fol­ about whether it is dangerous; rather, "custom operates before we have time lowed by an effect, where one is genuine, and the other is superfluous. Let us for reflex ion" (p. 104). Finally, this account does not distinguish between borrow a contemporary example in order to flesh this out (Glymour, 1998, p. human and animal learning; there is no difference between the way in which 40). Suppose that we are searching for the causes of lung cancer. We examine we ordinarily make causal inferences and the way in which a dog comes to fear punishment from the tone of his master's voice (p. 178). In sum, Hume a population of smokers and observe that everyone who smokes filterless cig­ arettes (A) develops lung cancer (E). Imagine that it is also the case that ) appears to be a behaviorist: that is, he seems to hold that causal inferences can everyone in this population has yellow fingers (X). Hume maintains that, in be fully explained in terms of associations. these circumstances, human beings will have a tendency to confuse causation It is tempting to think that Hume endorses radical associationism. After all, and correlation. After all, X has been repeatedly paired with E, and as a result, one can hardly turn a page of the Treatise without encountering the laws of these events will become closely associated in our minds. Thus, custom and association. Hume is quite explicit about the prominent role that associations habit would lead us to infer that people with yellow fingers will get lung can­ play in his theory of mind; he writes at one point that "if any thing can enti­ cer, whether or not they smoke. tle the author so gloriOUS a name as that of an inventor, 'tis the use he makes If Hume maintained that causal thought involves nothing but associations, of the principle of the association of ideas, which enters into most of his phi­ then this would be all there is to the story. He would have to concede that we losophy" (pp. 661-662). Indeed, associative learning plays a leading role in cannot discover causes in situations where there are confounding variables. everyone of Hume's psychological explanations in Book One: it accounts for But this is not what he says. Hume proceeds to argue that we can correct this the origin of ideas, the nature of general language and thought (Collier, propensity by employing what he calls "rules for judging causes and effects." I _I .. - J 178 COLLIER WHY HISTORY MATTERS 179 We shall afterwards take notice of some general rules by which we ought to regulate our judgment concerning causes and effects .... By rhem we learn to distinguish the acci­ sions in a section of Book Two of the Treatise entitled "Experiments to Confirm dental circumstances from the efficacious causes; and when we find that an effect can this System." Hume observes that we typically feel passions such as pride and be produc'd without the concurrence of any particular circumstance, we conclude that humility whenever we stand in close relation to agreeable or disagreeable that circumsmnce makes not a part of the efficacious cause, however frequently con­ join'd with it. (p. 149) objects. For example, we feel pride when we contemplate our prized posses­ sions, and we feel shame when we reflect upon our embarrassing habits. In These rules enable us to screen off accidental regularities. They direct us to order to explain these complex emotions, Hume proposes that they are caused search for instances in which an effect occurs without one of its associated con­ by what he calls the "double association" of impressions and ideas (p. 332). ditions; if we discover that a particular circumstance is not always conjoined His hypothesis is that we feel indirect passions (E) whenever there are objects with an effect, we can eliminate it from the list of efficacious conditions. To that (A) are closely related to us and (X) give us sensations of pleasure or pain. return to the smoking example, if we observe someone who develops lung can­ cer (E) without yellow fingers (X) - suppose she wore gloves - then we will Let us call this the "double association" hypothesis. In order to test it, have eliminated this superfluous circumstance from the genuine regularity Hume devises a series of new experiments that examine what would happen between smoking (A) and lung cancer (E). if each of these conditions was varied. For example, what if objects give us Sometimes nature is kind and provides us with this information. In other pleasure or pain (X), but we are unrelated to them (-A)? And what if we are cases, we must set up "new experiments" in order to discover the genuine causal closely related to these objects (A), but we are indifferent towards them (- X)? conditions. In these situations, Hume maintains, we would not feel any of the indirect passions (-E) [pp. 333-335]. There is no phenomenon in nature, but what is compounded and modify'd by so many different circumstances, that in order to arrive at the decisive point, we must carefully This is the reasoning 1 form in conformity to my hypothesis; and am pleas'd to find upon separate what is superfluous, and enquire by new experiments, if every particular cir­ trial that every thing answers exactly to my expectation .... This exact conformity of cumstance of the first experiment was essential to it. (p. 175) experience to our reasoning is convincing proof of the solidity of that hypothesis on which we reason. (p. 338) This method requires perseverance. We must vary each of the associated con­ ditions and observe whether the effect still obtains. And we must continue Hume isolates the cause of pride and humility, then, by considering situations with this procedure, as Hume puts it, until we have "enlarged our sphere of in which their associated conditions are altered. These test cases provide him experiments as much as possible" (p. 175). It is only after we have conducted with the requisite confidence that he has isolated a genuine regularity in nature. these elaborate investigations that we are entitled to assert that causal rela­ ti()ns "really are so" (p. 173). Jonathan Bennett is mistaken, then, when he Hume does not rely upon the custom and habit, then, when he carries out writes that Hume "exaggerates our intellectual passivity" and "ignores causal his experimental investigations into the passions. But what is .perhaps even judgments which look interrogatively rather than confidently towards the more remarkable is that Hume does not rely upon custom and habit when he future" (1971, p. 302). Hume recognizes that genuine regularities often lie came up with his custom-habit account! Hume's sophisticated approach to heneath the surface of things, and that we must use sophisticated experimen­ causal reasoning is on clear display when he examines the psychological tal procedures in order to discover them. processes that underlie our everyday causal attributions. Let us look carefully, then, at his methodological approach to the psychology of causal inference. Hume's Experimental Investigations The first step involves observation. Hume notices the following fact about human behavior: we make causal attributions whenever we see constant con­ There is another important reason why we should not interpret Hume as junctions. This observation can itself be viewed as a second-order constant endorsing radical associationism about causal cognition: he does not rely upon conjunction between (A) cases where subjects witness constant conjunctions, custom and habit when he conducts his own experimental investigations. and (E) cases where subjects make causal attributions. What Hume observes, in other words, is that A is usually followed by E. When Hume conducts his science of human nature, he subjects the regulari­ ties that he observes to careful examination. Perhaps the clearest illustration of If Hume's own causal reasoning depended upon custom and habit, then he Hume's sophisticated causal reasoning lies with his account of the indirect pas- would infer that A causes E. But this is not what happens. Hume never asserts that the correlation between A and E is sufficient to make causal judgments. _I ---:--uIII 180 COLLIER WHY HISTORY MATTERS 181 Rather, he begins the second step of his experimental investigation and puts it, the rules for judging causes and effects are "extremely difficult" to apply attempts to understand what goes on in our minds when we make such infer­ and require "the utmost stretch of human judgment" (p. 175). ences. That is, Hume seeks to discover the psychological faculty (X) which, 3. Hume's causal reasoning cannot be explained solely in terms of associa­ together with experience (A), produces these attributions (E). Notice that tions. Indeed, Hume's sophisticated experimental reasoning requires that we causal discovery in this case moves in the opposite direction from the smok­ regulate the associative propensities of the imagination. As we have seen, ing example. In the case of accidental regularities, one must subtract causally these associations are what lead us to confuse accidental and genuine circum­ irrelevant facrors in order to locate the genuine regularity; in the case at hand, stances, and it is only when we "correct" this "bias" that we have any hope of one must add causally relevant factors to do so. This additional stipulation is making genuine causal discoveries (p. 148). necessary because the psychological processes that underlie these causal attri­ butions are hidden from view; in contemporary terms, they are hidden vari­ Hume's Official Position Regarding Causal Reasoning ables. Hume proceeds to formulate two hypotheses concerning the nature of these As we have seen, Hume does not endorse radical associationism either in psychological faculties. According to the first, which we might call the "ratio­ theory or practice. He is quite sober about the fact that associations cannot nalist hypothesis," it is the faculty of reason which, together with the senses, tell us all there is to know about how human beings think about causation. If leads us to make the causal attributions that we do. However, Hume argues that were the case, then we could not (as we do) distinguish between acciden­ that this hypothesis cannot possibly be correct. If reason was the faculty that tal and genuine regularities, and Hume could not (as he does) carry out his we are looking for, then we must have non-circular justifications for drawing science of human nature. In short, Hume acknowledges that human beings are the conclusions that we do. But there are no such justifications (pp. 89-92). qllite capable of sophisticated causal reasoning. Therefore, reason cannot be responsible for these inferences. The only We can now see that there is a significant tension in Hume's attempt to remaining explanation is the custom-habit hypothesis. On this alternative become the "Newton of the mind." There are two distinct senses in which it account, our everyday causal attributions are explained in terms of the inter­ might be said that Hume's science of human nature is Newtonian. The first is action between sensory experience and custom and habit. We do not ordinar­ methodological: Hume maintains that the experimental methods of modern ily make causal inferences by considering reasons or arguments; rather, we are physics can be imported to the human sciences. The second is substantive: associatively primed to do so. The relevant psychological process is not reflec­ Hume hopes to show that the dynamics of the "mental world" can be com­ tion, but conditioned reflex. pletely understood in terms of laws of association, just as changes in the mate­ In sum, Hume's own causal reasoning has little in common with the cus­ rial world have been fully explained by laws of motion. tom-habit account: 1. Hume's causal reasoning is active rather than passive. On the custom­ Here is a kind of ATTRACTION, which in the mental world will be found to have as extraordinary effects as in the natural, and to shew itself in as many and as various habit account, one sees a constant conjunction and, without knOWing why, one forms. (pp. 12-13) makes a causal attribution. But Hume never embraces the na'ive view that cor­ relations are sufficient for causal attributions. Hume qua scientist of human At first glance, these two projects seem to fit together quite well: Hume would nature does not just sit back and observe case after case of A followed by E, use experimental methods in order to discover the universal laws of the mind. each time becoming more and more confident until he attributes a causal con­ But when one probes deeper, it becomes apparent that these two commit­ nection to them. Rather, Hume regards the conjunction between A and E as ments are not easily reconciled. the starting point for further inquiry. The problem is that, if the substantive claim is true, then Hume could never 2. Hume's causal reasoning involves "reflection" (p. 148). Our everyday have carried out his experimental investigations. Consider his examination of causal attributions take place implicitly and automatically; as he puts it, "cus­ our everyday causal inferences. Hume would have learned to associate A tom operates before we have time for reflexion" (p. 104). But this is not the case events (i.e., cases where we observe constant conjunctions) and E events (i.e., with Hume's experimental reasoning. He consciously formulates hypotheses, cases where we make causal attributions). But he could never have moved to devises new experiments, and determines which conditions to vary. Sometimes the second step of the investigation and discovered the deeper regularity these experiments screen off accidental factors; other times they postulate gen­ (A+X~E) that explains why we make the attributions that we do. Indeed, we uine factors. This method requires careful and tireless investigation. As Hume might pose this as a general paradox about radical associationism: if human -, 182 COLLIER WHY HISTORY MATTERS 183 cognition was governed solely by the laws of association, then nobody could discov­ showed that subjects can also make these judgments when the order of trials er that this was the case. is reversed (Le., "backwards blocking"). If subjects are shown trials in which The trouble arises from the self-reflexive nature of Hume's science of A + X are followed by E, and subsequently presented with trials in which A human nature. Hume must rely upon his own faculties in order to study alone is followed by E, they inevitably judge that there is no causal relation human cognition. But we can now see that this imposes constraints on what between X and E. he can say about the operations of the human mind; in particular, he cannot These blocking experiments demonstrate that subjects disregard potential make substantive claims that would undermine his capacity to carry Qut his causes if they are statistically redundant. Indeed, these studies suggest that our investigations. If the laws of association were presented as a complete account untutored causal attributions conform quite closely to the prescriptions of the of human cognition, however, it would violate this condition. Hume cannot rules for judging causes and effects. In the forward blocking experiments, for therefore combine these two aspects of his Newtonian project - something example, subjects observe instances where A alone is followed by E. The rules has to give. In the end, what gives is the thesis that associations provide a for judging causes and effects would direct them to conclude that A causes E. comprehensive account of causal cognition. And this is precisely what subjects do in subsequent trials: they judge that X One of Hume's fundamental insights is that human beings have more in is an accidental factor. This is true of the backwards blocking experiment as common with the rest of the animal kingdom than we are willing to admit. In well. Subjects observe that E can occur without X. The rules for judging caus­ order to support this deflationary picture of human nature, he sets out to show es and effects would direct them to conclude that X is an unnecessary part of that much of our cognitive lives can be explained in terms of elementary prin­ the genuine causal regularity. Once again, subjects conform to this normative ciples of association. But Hume recognizes, by the end of Book One, that recommendation and judge that A is the cause of E. there are essential limits to this reductive project. Causal judgment is one The contemporary controversy over causal learning concerns the question place where reason cannot completely yield to custom and habit. The prob· of how subjects manage to solve these problems (Shanks, Holyoak, and lem is that human beings not only learn associations, but we also reflect upon Medin, 1996). What are the computational processes that underlie behavior them. We are not entirely creatures of habit. in these tasks? On one side of the debate are those who defend an "associative approach." These researchers maintain that the causal learning experiments Contemporary Associative Learning Theory to the Rescue? can be fully explained in terms of excitatory and inhibitory connections in associative memory; some of them appeal to the Rescorla-Wagner model of In cognitive science, the question of how we ordinarily think about causa­ conditioning (Dickinson, Shanks, and Evenden, 1984; Shanks, 1985), where­ tion has been turned into an empirical research program. Whereas Hume as others invoke connectionist models (Read and Marcus-Newhall, 1993; relies upon informal observations about human behavior, contemporary psy­ Van Overwalle and Van Rooy, 1998). On the other side of the debate are chologists have developed a quantitative experimental paradigm with which those who defend a "rule-based approach." Some of these researchets main­ to investigate causal reasoning. In these experiments, subjects are presented tain that subjects use statistical rules in order to compute causation (Cheng with trials in which potential causes are followed by an effect. Subjects are and Holyoak, 1995); others postulate that subjects employ deductive rules in then asked to evaluate the causal relationship between these various events. order to do so (De Houwer and Beckers, 2002). The results of these experiments are unequivocal: subjects are capable of mak­ Ockham's razor dictates that, all else being equal, we should choose the sim­ ing sophisticated causal judgments. plest hypothesis. If the causal learning experiments can be accounted for in Consider the phenomenon of blocking. Anthony Dickinson and his col­ terms of nothing but associations, then, we ought to prefer this explanation leagues demonstrated that subjects are quite adept at selecting among poten­ over ones that posit more elaborate computational processes. But is it actual­ tial causal factors when making causal judgments (Dickinson, Shanks, and ly the case that the associative hypothesis can save the phenomena? As Hume Evenden, 1984). In their "forward blocking" experiments, subjects are initial­ pointed out, it is rather difficult to see how this could be so. Associations sim­ ly presented with trials in which A events are followed by E events, and they ply do not seem up to the task of explaining ho~ we manage to distinguish are subsequently presented with trials in which A + X are followed by E. between accidental and genuine causal regularities; quite the contrary, it Subjects are then asked to make a judgment about the relationship between appears that they are directly responsible for our propensity to collapse corre­ X and E. The results were consistent: subjects deny that X causes E, even lation and causation. In order to correct these mistakes, Hume maintains, we 1 ~~ though these events were repeatedly paired together. David Shanks (1985) must invoke rules for judging causes and effects. .~ _I 184 COLLIER WHY HISTORY MATTERS 185 Hume's official pOSItion is that sophisticated causal thought cannot be Higher-Order Causal Reasoning explained in terms of associations. But is this merely a lack of imagination on his part? One might object that Hume drastically underestimates the power Contemporary psychologists appear to have proven Hume wrong. Hume and complexity of associative learning theory. After all, associative psychology maintains that the laws of association do not allow us to distinguish between has made significant advances since the time that Hume wrote (Gormezano accidental and genuine regularities. In retrospect, we can see that he was and Kehoe, 1981). Moreover, it has been shown that simple associative systems pushed towards this negative conclusion simply because his own version of can perform extraordinary complex computations (Siegelmann and Sontag, associationism did not allow for cue competition. It seems that a slight mod­ 1995). Indeed, it is a truism in cognitive science that what seems like rule-gov­ ification in associative learning theory is all that is required in order to erned behavior can emerge from systems that do not literally employ rules explain sophisticated causal reasoning. (Bereiter, 1991; Rumelhart and McClelland, 1986, p. 32). There is reasonable But associationism is not out of the woods yet. Recent causal learning hope, then, that our capacity to make sophisticated causal judgments can be experiments have presented a serious challenge to the associative approach reduced to the level of associations. (see De Houwer, Beckers, and Vandorpe, 2005, for an overview). For exam­ This is not idle speculation. Contemporary psychologists have labored hard ple, De Houwer and Beckers (2002) performed an experiment in which sub­ over the last twenty years to show that the causal learning experiments can be jects were required to solve recursive backwards blocking tasks. Subjects were accommodated by the associative approach. For example, Dickinson, Shanks, initially presented with trials in which A+ B were followed by E and subse­ and Evenden (1984) showed that the forward blocking studies can be quently shown trials in which B+C were succeeded by E. In the final stage of explained in terms of the principles of classical animal conditioning. The cru­ the experiment, subjects were split into two groups: one group was presented cial point is that classical conditioning involves cue competition: when mul­ with trials in which A was followed by E, and the other was shown trials in tiple potential causes are followed by an effect, they will compete to predict which A was not followed by E. The crucial result was that subjects in the for­ its.next occurrence (Rescorla and Wagner, 1972). It is also significant that mer group were much more likely to judge that C, but not B, is a genuine associative learning in this model is driven by error-correction. This means cause of E. De Houwer and Beckers maintain that this result cannot be that when a potential cause reliably predicts an effect, there will be no change explained in terms of conditioning models, including those that incorporate in associative weights. Given these principles, it becomes relatively easy to absent cues (p. 149). These models might explain why these subjects deny account for the forward blocking data: the association formed between A and that B is a genuine cause, since A had been paired with B during the initial E in the first trial will block any association between X and E in subsequent stage of the experiment, and thus B would have been an absent but expected trials. After all, A has zero competition during the initial trials, and as a result, cue. But they cannot account for the fact that these subjects would draw any the connection between A and E will achieve maximal strength; since A reli­ conclusions at all about C, since it was never paired with A at any point dur­ ably predicts E in the later trials, the association between X and E will remain ing the experiment. unchanged. This study suggests that subjects use higher-order reasoning in order to The backwards blocking experiments are more difficult to explain in terms make causal judgments. Indeed, De Houwer and his colleagues conclude that of classical conditioning. The problem is that the standard model does not subjects could not have solved the recursive blocking task unless they con­ allow absent cues to enter into cue competition. The subsequent trials in sciously reflected upon rules for judging causes and effects (De Houwer, which A alone is followed by E, therefore, cannot weaken the initial associa­ Vandorpe, and Beckers, 2005, p. 46). The thought is that subjects must have tions that were formed between X and E. This problem led Van Hamme and reasoned as follows: A+B is followed by E; but A alone is followed by E; thus Wasserman (1994), among others, to modify the classical model of condition­ B is a superfluous factor; yet B+C is followed by E; thus C must be an effica­ ing so that the strengths of absent cues can be lowered, as long as they were cious cause of E. Even though A and C were never paired together, in other previously paired with a cue that is present. This slight revision enables the words, subjects can use deductive rules to draw conclusions about their rela­ associative hypothesis to account for the backwards blocking studies. Since A tionship. De Houwer and Beckers (2003) provide further evidence for the and X were paired during the initial trials, X would be expected to accompa­ claim that subjects use conscious reasoning strategies in order to make causal ny A on subsequent trials; as a result, its absence would weaken the connec­ judgments. Their experiments demonstrate that the ability to solve recursive tion between X and E. backwards blocking problems is severely attenuated when subjects are asked -, -'-,-=­ 186 COLLIER WHY HISTORY MATTERS 187 to perform difficult secondary tasks. This implies that causal learning is effort­ Conclusion ful and involves controlled cognitive processes. Even if these recent blocking experiments could be accommodated by the There is a growing consensus among contemporary psychologists that asso­ associative approach (which seems unlikely at the moment), there are inde­ ciations cannot fully explain causal learning (De Houwer, Vandorpe, and pendent reasons to think that this approach cannot provide an exhaustive Beckers, 2005; Dickinson, 2001; Pineno and Miller, 2007; Shanks, 2007). account of causal reasoning. One problem is that associative learning theory Researchers have begun to direct their attention to hybrid models that incor­ can only explain our ability to select among potential causal factors when we porate associative and rule-based components. This enables researchers to ask are provided with the appropriate negative instances. But as Hume pointed out, important new questions, such as: In which contexts do subjects rely upon nature is not always so kind. It might very well be the case that the rooster associations or rules? How are the rules for judging causes and effects acquired? always crows before the sun rises. In cases such as these, we must intervene in the How are they represented in the mind? In the end, the interesting issue is no order of things to make causal discoveries. Indeed, recent evidence suggests that longer whether causal cognition should be explained in terms of either associ­ hoth adults (Lagnado and Sloman, 2004) and children (Sommerville, 2007) ations or rules (Allan, 1993). As Hume taught us, this is a false dilemma: both carry out such interventions in order to distinguish between accidental and gen­ associations and rules are required for a complete and comprehensive account uine regularities. of how we think about causation. Another problem is that associative learning theories fail to explain how subjects can discover unobserved causal factors. The blocking experiments References involve situations where subjects must subtract causally irrelevant factors; but what about cases where they must add causally relevant features? As Hume Allan, L.G. (1993). Human contingency judgments: Rule based or associative? Psychological Bulletin, 114, 435-448. pointed out, if we are to make causal judgments on the basis of partial infor­ Bennett, J. (1971). Locke, Berkeley, Hume: Central themes. Oxt~lrd: Clarendon Press. mation, we must use background knowledge and generate hypotheses about Bereiter, C. (1991 ).lmplications of connectionism for thinkIng about rules. Educational Researcher, the complete set of factors at work. In fact, recent experiments have shown 20,10-16. Cheng, P.W., and Holyoak, K.J. (1995). Complex adaptive systems as intuitive statisticians: that people often rely upon sophisticated causal models when asked to make Causality, contingency, and prediction. In H.L. Roitblat and j.A. Meyer (Eds.), Comparative causal judgments in such circumstances (Waldmann, Hagmayer, and Blaisdell, approaches to cognitive science (pp. 271-302). Cambridge, Massachusetts: MIT Press. 2006). As these researchers put it, "people ... do not stick to the superficial Churchland, P.M. (1989). A neurocomputatumal perspective: The nature of mind and the structure of science. Cambridge, Massachusetts: MIT Press. level of event covariations, but reason and learn on the basis of deeper causal Churchland, P.M. (1995). Engine of reason, seat of the soul. Cambridge, Massachusetts: MIT Press. regularities" (p. 307). Collier, M. (1999). Filling the gaps: Hume and connectionism on the continued existence of The problem is that the associative approach makes causal learning an unperceived objects. Hume Studies, 25,155-170. entirely passive affair. This lesson is particularly damaging for those like Paul Collier, M. (2005a). A new look at Hume's theory of probabilistic inference. Hume Studies, 31. 21-36. Churchland (1989) who defend an associative account of scientific reasoning. Collier, M. (2005b). Hume and cognitive science: The current status of the controversy over Churchland maintains that recurrent connectionist networks can explain abstract ideas. Phenomenology and the Cognitive Sciences, 4, 197-207. "'::' De Houwer, J., and Beckers, T (2002). Higher-order retrospective revaluation in human causal how scientists recognize events "qua instance of a type of causal or law-gov­ learning. Quarterly Journal of Experimental Psychology, 55, 13 7-151. erned process" (1995, p. 105). As he puts it, these regularities would be rep­ De Houwer, J., and Beckers, T (2003). Secondary task difficulty modulates forward blocking in resented as "lines" rather than "points" in the network's state-space. But it is human contingency learning. Quarterly Journal of Experimental Psychology, 56B, '45-357. De Houwer, J., Beckers. T, and Vandorpe, S. (2005). Evidence for the role of higher order rea­ difficult to see how this could be so. These networks simply learn to associate soning processes in cue competition and other learning phenomena. Learning ar,,:! Behavior, whatever events follow one another in time. This makes them excellent mod­ 33. 239-249. els of how we ordinarily make probabilistic inferences (Collier, 2005a). But it De Houwer, J., Vandorpe, S., and Beckers, T (2005). On the role of controlled cognitive process­ es in human associative learning. In A.J. Wills (Ed.), New directions in human associative learn­ makes them poor models of sophisticated causal reasoning: the networks lack ing (pp. 41-63). Mahwah, New Jersey: Lawrence Erlbaum Associates. the capacity, as it were, to interrogate nature and discover her secrets. The Dickinson, A. (2001). Causal learning: An associative analysis. Quarterly Journal of Experimental networks extract statistical regularities from the data, but they cannot deter­ Psychology, 54B, 3-25. mine which ones are genuinely causal. Dickinson, A., Shanks, D, and Evenden, J. (1984). Judgment of act-outcome contingency: The role of selective attention. Quarterly Journal of Experimental Psychology, 54, 29-50. Glymour, C. (1998). Learning causes: Psychological explanations of causal explanation. Minds and Machines, 8,39-60. I _I ,-------=:::ilII 188 COLLIER .)4.1. 189 1 •• Gormezano, I., and Kehoe, J.E. (1981). Classical condirioning and the law of contiguity. In ©2007 The Institute of Mind and BehaVIor, Inc. P. Har~em and MD. Zeiler (Eds.), Predictability, correlation, and contiguity (pp. 1-45). New The Journal of Mmd and BehaVIOr :' !:i Summer and Aut,umn 2007, Volume 28, Numbers 3 and 4 York: John Wiley and Sons. Pages 189-202 ,~~ ~::fi Hume, D. (1978). A treatise of human nature. Oxford: Oxford University Press. (Original work ISSN 0271-0137 i~ published 1739) ~~ Jessop, T.E. (1966). Some misunderstandings of Hume. In v.c. Chappell (Ed.), Hume: A collec­ ::~ t ·~ tion of critical essays (pp. 35-52). New York: Anchor Books. Lagnado, D., and Sloman, S. (2004). The advantage of timely interevention. Journal of Experimental Psychology: Learning, Memory, and Cognition, 30, 856--876. Pears, D. (1991). Hume's system: An examination of the first book of his Treatise. Oxford: Oxford , University Press. Pineno, 0., and Miller, R. (2007), Comparing associative, statistical, and inferential reasoning accounts of human contingency learning. The Ql41rterly Journal of Experimental Psychology, The Phenomenology of Freedom ,'~' Tomis Kapitan .,.~ 60,310-329. Read, S.J., and Marcus-Newhall, A. (1993), Explanatory coherence in social explanations: A parallel disrributed processing account. Journal of Personality and Social Psychology. 65, Northern Illinois University 429-447. Rescorla, R.A., and Wagner, A.R. (1972). A theory of Pavlonian conditioning: Variations in the effectiveness of reinforcement and nonreinforcement. In A.H. Black and W.E Prokasy (Eds.), Classical conditioning 1/: Current research and theory (pp. 64-99), New York: Appleton Books. John Searle describes our sense of freedom as an experience of a "gap" between an Rumelhart, DE, McClelland, J.L., and the PDP research group. (1986). Parallel distributed pro­ intentional action and its psychological antecedents, specifically, our reasons. Since the cessmg: Explorations in the microstructure of cognition. Volume I. Cambridge, Massachusetrs: gap is itself understood as a lack of causation, then no agent can accept the antecedent MIT Press. determination of voluntary action except at the price of "practical inconsistency." I Shanks, D.R. (1985), Forward and backward blocking in human contingency judgment. argue that despite Searle's insightful discussion, the sense of freedom is not an experi­ Ql41rterly Journal of Experimental Psychology, 37, 1-21. ence of a gap as he describes it but, instead, is a higher-order attitude concerning one's Shanks, D.R. (2007). Associationism and cognition: Human contingency learning at 25. limited grasp of causes. As a result, a determinist can engage in voluntary action with­ Ql41rterly Journal of Experimental Psychology, 60,291-309. out falling into inconsistency. Shanks, D.R., Holyoak, K,J., and Medin, D.L. (Eds.). (1996). The psychology of learning and moti­ vation (Volume 34): Causal learning. San Diego, California: Academic Press. Keywords: freedom, deliberation, gap Siegelmann HT, and Sontag ED. (1995). On the computational power of neural networks. Journal of Computer and System Sciences, 50, 132-150. Sommerville, J.A. (2007). Detecting causal structure: The role of interventions in infants' under­ As practical beings, we act with a sense of freedom or, to use Kant's memo­ standing of psychological and physical causal relations. In A. Gopnik and L. Shulz (Eds.), rable phrase, "under the idea of freedom." This feeling derives from our expe­ Causal learning: Psychology, philosophy, and computation (pp. 48-57). Oxford: Oxford University rience of alternative courses of action as being open to us, and from our envis­ Press. Van Hamme, L.J., and Wasserman, E,A. (1994). Cue competition in causality judgments: The agement of a past as partly shaped by our own voluntary doings, and an incom­ role of nonpresentation and compound stimulus elements. Learning and Motivation, 25, plete future partly subject to our deliberations. There is tremendous emotion­ 127-151. al investment in this attitude. It is the basis for pride about our past achieve­ Van Overwalle, E, and Van Rooy, D. (1998), A connectionist approach to causal attribution. In S.J. Reed and L.c. Miller (Eds.), Connectionist models of social reasoning and behavior (pp. ments, regrets about what we have done or failed to do, and hopes that we can 143-171). Mahwah, New Jersey: Lawrence Erlbaum Associates. change our ways for the better. It is central to the belief that we might have Waldmann, M.R., Hagmayer, Y., and Blaisdell, A.P. (2006). Beyond the information given: acted otherwise and, thereby, are responsible for the way we have acted. Causal models in learning and reasoning. Current Directions in Psychological Science, 15, 307-311. Without it, it is difficult to see why we would plan, give advice, establish nor­ mative systems, or put forth any effort at all. What exactly is the sense or experience of freedom? Is it a type of perceptu­ al awareness? Is it a purely intellectual or conceptual cognition? Is it a con­ ;'.; scious doxastic state? And what exactly is its content? John Searle, agreeing that feeling of freedom derives from "a sense of alternative possibilities" (2001, p. 67), speaks of a causal gap between our reasons and our decisions- Requests for reprints should be sent to Tomis Kapitan, Department of Philosophy, Northern Illinois University, DeKalb, Illinois 60115. Email: Kapitan@niu.edu i _, \Jil