In our target article, we put forward the claim that all associative learning in humans is mediated by the non-automatic truth evaluation of propositions about relations in the world. This implies that there is no need to postulate a second learning mechanism that is based on the unconscious formation of links between symbolic representations. One positive outcome of our target article was that several commentators expressed their agreement with the basic tenet of our position (Beckers & Vervliet; Chater; Morsella, Riddle, & Bargh [Morsella et al.]; Gopnik; Greenwood; Lagnado; Li; Lyn & Rumbaugh; and Newell). However, it is clear that we were unable to convince everybody. Many commentators explicitly expressed their belief in a dual-system model of learning in which both propositional and link-formation processes can produce learning (Baeyens, Vansteenwegen, & Hermans [Baeyens et al.]; Boakes; Dawson & Schell; Dwyer, Le Pelley, George, Haselgrove, & Honey [Dwyer et al.]; Gawronski & Bodenhausen; Livesey & Harris; Matute & Vadillo; McLaren; Miles, Proctor, & Capaldi [Miles et al.]; Olsson & Öhman; Schultheis & Lachnit; Uleman). This, too, is a positive outcome of our target article, because it confirms that we are not fighting a straw man, but a widely held position in psychology. Few researchers had previously made explicit in writing their belief in a dual-system model of learning.
The arguments made by the commentators for rejecting the propositional approach can be grouped into three categories: (1) the evidence that supports the propositional approach can also be explained by link-formation models; (2) there are empirical findings that contradict the propositional approach and thus support the idea of a second, link-based, learning mechanism; and (3) there are conceptual grounds for rejecting the possibility that all learning is propositional. Some of the arguments presented in the commentaries are based on an interpretation of the existing empirical evidence with which we disagree. Others seem to reflect a misunderstanding of the propositional approach, which suggests that we were not completely clear in the target article. From our analysis, we conclude that although the current propositional approach needs further refinement, the basic assumption that all associative learning is mediated by propositions is conceptually sound and is supported by the available data.
R1. Can evidence for the propositional approach also be explained by link-formation models?
In section 3 of the target article, a large number of empirical findings were presented that confirm predictions of the propositional approach. We first describe exactly why the propositional approach predicts these findings, and then look at the extent to which these findings can be explained by link-formation models. The core assumption of the propositional approach is that associative learning depends on the non-automatic truth evaluation of propositions about relations in the world. Some commentaries revealed a misunderstanding of what we meant by “propositions” and how propositions differ from associative links. Propositions are statements about the events in the world and how they are related. These statements may be quite specific as to the nature of the relation between events (e.g., “X causes Y”). However, they can be quite general; they may specify only that particular events are related, but not how they are related. We agree with Gawronski & Bodenhausen that the crucial distinction between propositions and links is that only propositions imply a truth value. People can thus differ in the degree to which they believe that statements about the world are true or false. Although propositions are described as statements, they do not, as Shanks seems to imply, necessarily involve abstract representations. As Barsalou (Reference Barsalou2008) pointed out, propositions can also involve embodied, grounded representations. Propositions are very different from associative links. Associative links are not statements about the world. They are assumed to be states of the world, namely, states through which one representation can activate another representation.
The idea that associative learning is a function of the non-automatic truth evaluation of propositions leads to two main predictions. First, everything that influences truth evaluation can influence associative learning. This means that apart from the actual occurrence of events, factors such as prior knowledge, instructions, intervention, and deductive reasoning also matter. Second, because truth evaluation depends on non-automatic processes (in the sense that they are dependent on awareness, cognitive resources, time, and goals), associative learning should also be non-automatic. There is substantial evidence to support these predictions (see our target article and De Houwer [Reference De Houwer2009] for reviews). Because of this evidence, many commentators accept the conclusion that some associative learning phenomena are the result of propositional processes (e.g., Baeyens et al.; Boakes). However, they do not agree with the more far-reaching conclusion that all learning is propositional. They argue that there are some “true” forms of learning that are driven by link formation. Although it is not clear to us what the defining characteristic of “true” learning is, we do agree that evidence in support of the propositional approach is also compatible with dual-system models. Thus, such findings do not exclude the possibility that link formation exists as a second learning mechanism that supplements propositional learning. The reasons for why many commentators maintain a belief in the existence of link formation will be analyzed in the next section. In the remainder of this section, we will discuss a second response to the evidence in support of the propositional approach, namely, that link formation models can also account for (some of) that evidence.
A first group of commentators (Dwyer et al.; Livesey & Harris; Witnauer, Urcelay, & Miller [Witnauer et al.]; see also section 2.1 of the target article) points out that link formation is not necessarily a fully automatic process. For instance, influential link-formation models do postulate that attention modulates the formation of links (e.g., Mackintosh Reference Mackintosh1975; Pearce & Hall Reference Pearce and Hall1980). One could also argue that link formation leads to awareness of relations in the world (e.g., Davey Reference Davey1992). Hence, evidence for the non-automatic nature of associative learning could in principle be accommodated by link-formation models by adding assumptions about the way in which link formation is non-automatic. However, one of the important reasons why many researchers believe in the existence of a link-formation mechanism is precisely because it might account for seemingly automatic, irrational types of learning. This is probably why many commentators explicitly characterize link formation as primitive and automatic. Another problem with the argument that link formation might be non-automatic is that there is no a priori reason for this assumption. There are, however, good reasons to assume that the truth evaluation of propositions is a largely non-automatic process. Hence, the failure to observe automatic learning is a prediction of, and therefore supports, the propositional approach.
Other commentators reject the dual-system idea, but argue that the single system is based on link formation. They suggest that the evidence for the propositional approach can be accounted for by (link-based) connectionist models of learning. In this way, propositional processes can be reduced to link formation. We did foresee this argument in our target paper when we discussed the nature of connectionist models (sect. 6.2). Our main point was that connectionist models differ from link-formation models because they involve sub-symbolic, rather than symbolic, representations. The nodes in a connectionist model do represent information (as Penn, Cheng, Holyoak, Hummel, & Povinelli [Penn et al.] point out), but they do not represent a discrete stimulus or event in the world (as they do in link-formation models). In other words, the information represented by symbolic nodes (representations) is specified by the model, but the information represented by sub-symbolic nodes (in connectionist networks) needs to be inferred from how the system operates. Our quarrel is only with symbolic link-formation models. We are aware of the fact that the brain resembles a sub-symbolic connectionist network. In fact, connectionist models can be regarded as simulated miniature brains (Clark Reference Clark1990). It is also obvious that the truth evaluation of propositions must somehow be implemented in the brain. Hence, we do not exclude the possibility that truth evaluation might be implemented in a connectionist model (although current models still seem to have a long way to go before this promise can be realized). Such an implementation would not, however, constitute evidence that propositional processes can be reduced to symbolic link formation or that there is a need to assume the existence of symbolic link formation.
R2. Are there empirical findings that contradict the propositional approach?
A second category of arguments focuses on evidence for associative learning that cannot be explained by propositional models. The evidence within this category can itself be divided into three classes. The first class concerns evidence for associative learning in the absence of relevant propositions (e.g., learning without awareness). Second, several commentators refer to learned behavior that is arational in that it cannot be inferred logically from the propositions that people entertain. They claim that the propositional approach does not provide an explanation for the production of Pavlovian conditioned responses (CRs). The third class includes evidence for arational beliefs about relations in the world. Some beliefs seem to result from factors other than truth evaluation and, it is argued therefore, constitute evidence for the formation of links. In this section, we explain why we are not convinced by these arguments.
R2.1. Evidence for learning in the absence of relevant propositions
A number of commentators cited studies that they considered to be demonstrations of unconscious learning (Boakes; Dawson & Schell; Dwyer et al.; Gawrosnki & Bodenhausen; Livesey & Harris; McLaren; Miles et al.; Morsella et al.; Olsson & Öhman; Penn et al.; Uleman). The majority of these studies have already been considered in the several reviews we cited in our target article, and shown to be artefactual, unreplicable, or subject to alternative interpretations. In many cases, direct supporting evidence for specific alternative interpretations has been reported. Other studies nominated by the commentators have been published subsequent to these reviews, but have used procedures subject to the same criticisms as the earlier studies. The commentary process provides a unique insight into the results that are considered to provide the most robust evidence for unconscious learning by dual-system theorists; therefore, we consider it important to briefly review the main difficulties with these studies and provide references with additional detail for the interested reader.
Before turning to the specific research procedures, it is important to first emphasize one general point about this line of inquiry. Apparent unaware conditioning can be generated simply by the use of an insensitive or noisy measure of awareness, because some aware participants will be misclassified as being unaware. As a consequence, studies that successfully show a relationship between awareness and conditioning should be given greater weight than those that fail to find such a relationship, unless there is independent evidence that the null result was not due to lack of sensitivity of the awareness measure (Lovibond et al., submitted).
Studies involving fear conditioning to backwardly masked (or subliminal) fear-relevant stimuli (cited by Dawson & Schell, Olsson & Öhman, and Penn et al.) suffer from three primary limitations. The first two are insensitive measures of awareness and the use of fixed stimulus durations which are suprathreshold for some participants. The third is a little more subtle. Sometimes the measurement of awareness relies on report of stimulus identity. However, participants who cannot identify the stimuli may nevertheless be aware that CS+ and CS- differ in some way – this is all that is required for differential conditioning to be observed (Cornwell et al. Reference Cornwell, Echiverri and Grillon2007; Lovibond & Shanks Reference Lovibond and Shanks2002; Pessoa et al. Reference Pessoa, Japee and Ungerleider2005). In addition, several early studies have since been shown to be an artifact of the experimental procedure; they could be explained by a specific correlated hypothesis account based on restrictions on trial order (Wiens et al. Reference Wiens, Katkin and Öhman2003). Finally, studies using a more sensitive measure of awareness, such as a trial-by-trial measure of subjective US expectancy, have shown a concordance between this measure and objective measures such as skin conductance (e.g., Öhman & Soares Reference Öhman and Soares1998). These findings directly support a single-system account. The same is true for the resistance to extinction observed with supraliminal fear-relevant stimuli (Dawson et al. Reference Dawson, Schell and Banis1986; Lovibond et al. Reference Lovibond, Siddle and Bond1993).
Studies involving anesthesia and spinal preparations (cited by Dwyer et al. and Miles et al.) have mostly been conducted in animals, so implications for the role of awareness are hard to draw. By contrast, there has been a great deal of research on learning in humans suffering from amnesia. Early reports of conditioning in amnesics (cited by Miles et al.) were largely anecdotal, with little detail regarding the nature of the assessment of conscious knowledge. Those studies which have been more fully reported (e.g., Bechara et al. Reference Bechara, Tranel, Damasio, Adolphs, Rockland and Damasio1995; Clark & Squire Reference Clark and Squire1998) have often used a post-experimental assessment of awareness, which is particularly likely to underestimate contingency knowledge in amnesics at the time of conditioning (Lovibond & Shanks Reference Lovibond and Shanks2002). A recent study by Speekenbrink et al. (Reference Speekenbrink, Channon and Shanks2008), however, was specifically designed to assess the relationship between task performance and relevant explicit knowledge during learning, in a probabilistic category learning task. They found a slightly slower learning rate in amnesics compared to control participants. But, importantly, there was no evidence for a qualitatively different learning process and no dissociation between task performance and explicit knowledge of cue-outcome association.
As noted by Miles et al. and Olsson & Öhman, there have been several reports that eyeblink conditioning with a delay conditioning procedure (in which CS and US overlap) is unrelated to awareness, unlike conditioning with a trace procedure (in which CS and US are separated by a brief interval; see Li's commentary). However, there have been specific criticisms of some of these studies (Lovibond & Shanks Reference Lovibond and Shanks2002; Shanks & Lovibond Reference Shanks and Lovibond2002). Furthermore, other researchers have successfully demonstrated a relationship between awareness and differential eyeblink conditioning (with no evidence of conditioning in unaware participants), regardless of whether a trace or delay procedure is used (Knuttinen et al. Reference Knuttinen, Power, Preston and Disterhoft2001; Lovibond et al, submitted).
The use of startle modulation as an index of conditioning (cited by Olsson & Öhman) has not, in general, yielded evidence of unconscious conditioning. The early study by Hamm and Vaitl (Reference Hamm and Vaitl1996) is limited by the fact that contingency awareness was assessed after an extinction period (Lovibond & Shanks Reference Lovibond and Shanks2002), and more recent studies by Weike et al. (Reference Weike, Schupp and Hamm2007) have been criticized by Dawson et al. (Reference Dawson, Rissling, Schell and Wilcox2007) as using an insensitive recall-based awareness measure. More importantly, however, several studies have shown a reliable relationship between contingency awareness and startle modulation, with no evidence of modulation in the unaware group (Dawson et al. Reference Dawson, Rissling, Schell and Wilcox2007; Lipp & Purkis Reference Lipp and Purkis2005; Purkis & Lipp Reference Purkis and Lipp2001).
There have been numerous reviews of evaluative conditioning (cited by Baeyens et al.), and specifically the claim that such conditioning can occur without contingency awareness (see also Bliss-Moreau & Barrett). As noted in our target article, more recent studies have started to provide clearer evidence of an association between contingency awareness and evaluative conditioning, both for the picture-picture procedure (Pleyers et al. Reference Pleyers, Corneille, Luminet and Yzerbyt2007) and the flavor-flavor procedure (Wardle et al. Reference Wardle, Mitchell and Lovibond2007). Other commentators (Boakes; Livesey & Harris) focused on the more general point that conditioning with tastes, odors, and internal bodily consequences, whether considered as an example of evaluative conditioning or not, may be independent of cognitive factors such as instruction and awareness. Again, much of this evidence in humans is anecdotal. Curiously, Boakes also cites placebo effects in this context, but of course this is an instructional effect that suggests bodily reactions are in fact sensitive to cognitive factors. One exception, as noted by Lovibond and Shanks (Reference Lovibond and Shanks2002), is the odor-taste learning demonstrated by Stevenson et al. (Reference Stevenson, Boakes and Prescott1998). These findings are suggestive of an associative process that is independent of explicit contingency knowledge, and this avenue is worthy of further investigation to see if it supports the idea of a separate gustatory learning module isolated from cognitive processes.
Finally, many commentators (Gawronski & Bodenhausen; Miles et al.; Morsella et al.; Penn et al.; Uleman) raised studies from the cognitive and social literatures that do not involve direct recording of conditioned responses but that are, nonetheless, clearly associative in nature. Many of these studies report effects that have been labeled implicit learning, such as sequence learning, motor learning, and artificial grammar learning. The most comprehensive review of this literature was carried out by Shanks and St. John (Reference Shanks and St. John1994), who reported that virtually all of the studies of implicit learning at that time failed either their sensitivity or their informational criterion, or both. Subsequent research has strengthened and extended Shanks and St. John's conclusion and extended it to other claims of implicit learning, including sequence learning (Wilkinson & Shanks Reference Wilkinson and Shanks2004), continuous tracking (Chambaron et al. Reference Chambaron, Ginhac, Ferrel-Chapus and Perruchet2006), probabilistic category learning (Newell et al. Reference Newell, Lagnado and Shanks2007), and contextual cueing (Smyth & Shanks Reference Smyth and Shanks2008).
An alternative approach to demonstrating non-propositional learning, taken by Perruchet (Reference Perruchet1985) and Perruchet et al. (Reference Perruchet, Cleeremans and Destrebecqz2006), is to show effects of learning that directly contradict relevant propositions (see sect. 5.1 of the target article). Such evidence would be much more difficult to discount than the evidence for unaware learning, because demonstrating the presence of a proposition that contradicts the behavior is easier than demonstrating the absence of propositions altogether.
Some commentators argued that the Perruchet effect alone is sufficient evidence to postulate the existence of link formation as a second learning mechanism (Dwyer et al.). We believe, however, that the evidence is not yet conclusive. For instance, recent findings suggest that at least the reaction time version of the Perruchet effect does not result from the operation of two associative learning systems. Mitchell et al. (in revision) presented half of their participants with the same 50% tone-square partial reinforcement schedule used by Perruchet et al (Reference Perruchet, Cleeremans and Destrebecqz2006). The remaining participants received the same schedule but with the tones removed. In their experiment, just as in the one Perruchet et al. (Reference Perruchet, Cleeremans and Destrebecqz2006) conducted, reaction times were fastest following runs of square-present trials, and slowest following runs of square-absent trials. This was true for all participants, regardless of whether or not they had been presented with the tone (CS). This suggests that the pattern of reaction times does not indicate the operation of an associative learning process; it is simply the consequence of the recent square presentations. Because the behavioral effect is not due to the relation between the tone and the target event, there is no longer a contradiction between propositional beliefs about the tone-square relation and a behavioral effect of that relation. Therefore, the dissociation between RTs and expectancy cannot indicate the operation of two distinct associative learning systems.
The research on the Perruchet effect shows that the propositional approach can be falsified in principle. However, it also illustrates how one should be careful in drawing conclusions prematurely. Many years of research on learning in the absence of relevant propositional knowledge has led to few (if any) convincing findings. This calls for a cautious approach toward any new piece of evidence that claims to demonstrate learning in the absence of relevant propositional knowledge.
R2.2. Learning can result in arational behavior
Some commentators (e.g., Baeyens et al.; Livesey & Harris) point out that when participants do have relevant propositional knowledge about relations in the world, one cannot explain, on the basis of this knowledge, why those relations have the behavioral effect that they do. For instance, why does the proposition “the tone predicts the shock” lead to an increase in skin conductance (the CR) after the presentation of the tone? Why would a proposition such as “this flavor was sometimes followed by a bad aftertaste” lead to a disliking of the flavor when presented on its own? In the target article, we acknowledged that the propositional approach does not explain why propositions about relations in the world have certain effects but do not have others. This is because propositional models are not models of behavior. They are models of one determinant of behavior: associative learning. There are many other factors that go to determine the way in which learning translates to performance.
This line of reasoning may give the impression that the propositional approach is unfalsifiable; any behavior that does not seem to follow logically from the propositions entertained can be attributed to some performance factor or other. However, although we do not have a psychological model of this translation process, we do know quite a lot about the effects that particular propositions will have, and we do assume some stability in this process. For example, as Olsson & Öhman concede, we know that (all else being equal) the proposition “the tone signals shock” will lead to an expectancy of shock when the tone is presented, and this expectancy will generate the CR of increased skin conductance (see sect. 3.2 of the target article). At some point during evolution, it would appear that the cognitive expectancy of negative events became a trigger for the activation of genetically determined defensive response patterns that include an increase in skin conductance. We can, given that we know this process is in place, make predictions about the effects of propositions on behavior.
It is very important to note that, in fact, the link-formation models do not explain CR production any better than does the propositional approach. Most link-formation models remain silent about how links are expressed in behavior. Furthermore, the psychological mechanism provided by the link model for the production of CRs relies on the old idea that conditioned responses are simply indirectly activated unconditioned responses. Thus, for instance, when a tone is paired with a shock, it might evoke an increase in skin conductance because activation caused by the presentation of the tone spreads to a representation of the shock, which then activates the responses associated with shock. But, as we pointed out in our target article (see sect. 2.2), conditioned responses often differ substantially from unconditioned responses. And these differences cannot always be explained in terms of the nature of the CS. Link-formation models also do not explain why learning can be selective. For instance, why does a flavor that is paired with a bad aftertaste become negative, whereas a color paired with the same aftertaste does not (see comment of Baeyens et al.)? In sum, neither propositional models nor link-formation models provide a full account of learned behavioral responses. Hence, we see no reason why evidence for arational learned behavior should favor link-formation models.
R2.3. People sometimes have arational beliefs about relations in the world
The studies discussed in sections R2.1 and R2.2 of this rejoinder can be seen as attempts to discount propositional models of learning by showing behavioral effects of learning that cannot be due to relevant propositions, either because relevant propositions are absent (sect. 2.1) or because the behavioral effects of learning cannot be inferred logically from the relevant propositions (sect. 2.2). Some commentators (Dwyer et al.; McLaren) discuss a third set of studies that followed a different approach. They suggest that some propositional beliefs about relations in the world are arational, and, therefore, might result from the operation of a link-formation mechanism. The evidence presented in the target article on this issue (sect. 4.3.2) came from the studies of Le Pelley et al. (Reference Le Pelley, Oakeshott and McLaren2005a) on unblocking, Karazinov and Boakes (Reference Karazinov and Boakes2007) on second-order conditioning (see Uleman for very similar studies in the context of social psychology), and from Shanks's (Reference Shanks2007) example of the judgment of event probabilities. Additional studies of this kind are presented in the commentaries (e.g., McLaren cites Spiegel & McLaren's [Reference Spiegel and McLaren2006] work on sequence learning).
As we made clear in our target article (see sect. 5.2), evidence for seemingly arational beliefs about relations in the world does not convince us of the need for a separate link-formation mechanism. First, there is a very important general point that, even if certain findings cannot be explained by propositional models, this is not evidence for the existence of link-formation processes. Many other explanations are also possible. Second, detailed research is needed before one can exclude with confidence propositional processes as a source of the observed phenomenon. Human reasoning is not necessarily normatively correct. Hence, demonstrations of arational or irrational beliefs do not fall outside of the scope of truth evaluation on the basis of reasoning. Thus, we see no need to evoke the idea of link formation in order to account for such beliefs.
We do acknowledge that these concerns complicate attempts to provide evidence in support of the link-formation mechanism. Hence, we understand the complaint of certain commentators that the propositional approach runs the risk of becoming unfalsifiable (McLaren; Nolan). However, those who want to argue for the existence of link-formation processes should examine carefully all possible explanations that do not rely on link formation before they conclude that link-formation processes do operate. It is unlikely that such issues can be settled by isolated demonstrations of effects that at first sight appear to contradict the propositional approach, or (as Witnauer et al. point out) seem consistent with a particular associative model (see McLaren; Dwyer et al.). The burden of proof in the first instance is on those who want to claim the existence of link-formation processes to rule out all other possibilities. That being said, we do accept that once efforts have been made to rule out these other possibilities, proponents of propositional models cannot simply dismiss potentially problematic findings on the basis of the argument that some kind of propositional process might be crucial. They should always back up their alternative explanations with empirical evidence or sound arguments (see the work of Mitchell et al. [in revision] on the Perruchet effect). If they cannot, they should accept the conclusion that some findings fall beyond the scope of the propositional approach.
R3. Are there conceptual arguments for rejecting the propositional approach?
The evidence discussed in section R2 of this response relates to possible inconsistencies between the propositional approach and empirical knowledge about associative learning. In this section, we discuss comments that relate to possible inconsistencies between the propositional approach and our knowledge about evolution, the brain, and normative criteria that good theories should meet.
R3.1. Propositional models are not in line with the principles of evolution
Some commentators (Miles et al.; Schultheis & Lachnit) suggested that the propositional model of human learning requires postulation of a “magical cut-point” in evolution, when propositional learning appeared and reflexive learning disappeared. Similarly, these and other commentators (Matute & Vadillo; Olsson & Öhman; Penn et al.) asked why evolution would simply abandon an effective and adaptive reflexive mechanism. However, in our target article, we explicitly argued for a “continuum of cognitive complexity” (sect. 6.3) between primordial creatures, on the one hand, and humans, on the other hand. We suggested that early representational structures were likely S-R in nature, that these were elaborated to allow S-S structures that separated knowledge from action, and that further quantitative elaboration of knowledge structures provided the necessary foundation for abstract reasoning and artificial symbolic systems (language). This continuity position is entirely consistent with the known principles of evolution, which as noted by Olsson and Öhman is “a slow accumulative process . . . that incorporates successful adaptations at one level into more complex functions at higher levels.” Therefore, our answer to the question of what happened to primitive reflexive mechanisms is not that they shriveled up and died (Penn et al.), but rather that they evolved into the rich representational system that humans possess. As Morsella et al. note in their commentary, this system is suitable for the solution of both complex and simple problems.
Furthermore, contrary to the claims of several commentators (e.g., Hall; Matute & Vadillo), the removal of a primitive learning mechanism during the course of evolution might well have been adaptive. A primitive mechanism that creates links whenever certain inputs co-occur might be adaptive for simple organisms that register only limited and simple inputs but maladaptive for more complex organisms that receive a lot of complex input. In more complex organisms, an automatic link system would lead to an overload of associations that would result in chaotic behavior (as in the case of schizophrenia; see Carson et al. Reference Carson, Peterson and Higgins2003). This problem can be managed if propositional encoding needs to occur before a relation can be coded in memory and influence behavior. Hence, in addition to the benefits offered by a more complex, propositionally based learning mechanism, reducing the negative effects of a primitive learning mechanism provides a second evolutionary reason for the gradual evolution from primitive to more complex learning mechanisms.
Ironically, it is the dual-system model that suggests an implausible discontinuity in evolution, because it denies that earlier reflexive systems formed the precursor to representational abilities and, subsequently, to human reasoning and language. Instead, proponents of this model generally assume that reasoning emerged independently and late in mammalian evolution, for example, in primates. In the extreme case, those who propose that reasoning relies upon language (e.g., Castro & Wasserman) are forced to argue that reasoning is unique to humans (and perhaps, to a very limited extent, certain primates) and therefore emerged de novo at “one minute to midnight” in biological time. But emergence of a novel capacity in such a short time-frame is entirely implausible given the accumulative and opportunistic nature of evolution, which is much more likely to appropriate existing mechanisms to new purposes rather than develop a new one from scratch. We agree with Christiansen and Chater (Reference Christiansen and Chater2008) that it is more plausible to suppose that language was enabled by the prior evolution of a critical level of cognitive capacity, and that language was, therefore, moulded by the operating characteristics of the cognitive system.
The only alternative for dual-system theorists is to concede that propositional reasoning, or the representational capacity that preceded it, developed well before the evolution of humans. But this strategy raises another set of problems for dual-system theorists. They must now confront seriously the role that the nascent cognitive system might play in learning and other tasks in the animal laboratory. According to the propositional model, this is exactly what animal researchers have been studying successfully for many years, and indeed they sometimes describe their work as “animal cognition.” But from a dual-system perspective, it would now be necessary to force a distinction between cognitive processes and co-existing reflexive or link-formation processes. We have already reviewed the vanishingly small evidence base for such a distinction in humans, where some predictions of the dual-system model can be tested through verbal report. It seems unlikely that independent evidence will emerge from the animal laboratory that clearly supports a dual-system architecture. Rather, we consider that the learning capacities of particular species studied in the laboratory will reflect the nature and complexity of their cognitive system. Species more closely related to humans (e.g., other mammals and primates, in particular) will have cognitive systems and hence learning capacities correspondingly more similar to our own (see Lyn & Rumbaugh).
It is through this lens that we approach the comments of Miles et al. and Schultheis & Lachnit concerning the learning abilities of honeybees. We suggest that during human evolution there was a continuum of development of cognitive capacity from simple organisms to modern humans. However, not all of these earlier species survive today, and of those that are extant, many are on branches of the evolutionary tree that diverged quite early from human evolution. Therefore, we are not able to make strong predictions as to the cognitive abilities of a particular species. We would not be surprised if a species such as the honeybee showed some characteristics indicative of capacity to represent the environment, such as cognitive maps, but we would also not be surprised to find strong limitations to such capacities. It is also likely that species with substantially different nervous systems will have solved the same problem in different ways, due to convergent evolution. For example, different species may show blocking on the basis of different learning mechanisms, just as they achieve other capacities such as locomotion through different mechanisms (e.g., flying, swimming, walking).
Some commentators made the case for preservation of a reflexive learning system on the basis that this system is fast and automatic (Matute & Vadillo), and that evolution favors the concurrent existence of multiple, complementary systems with different strengths (Hall; Olsson & Öhman; Schultheis & Lachnit). The implication of this argument is that the cognitive system is slow and effortful and hence will be out-performed by a reflexive system in situations requiring rapid reactions. However, it is not at all clear that a reflexive system, if it existed in humans, would in fact be faster in such situations. This is because situations requiring rapid responses (e.g., predator avoidance) are ones that depend on performance based on prior learning, not on new learning per se. As we noted in our target article, performance based on existing propositional knowledge is fast and automatic – regardless of whether the knowledge is acquired through direct contingency exposure or is learned symbolically. Furthermore, when it comes to acquisition of new learning, it is debatable whether reflexive learning would be faster than propositional learning. Many S-R theories propose a relatively slow and gradual process of establishing and strengthening of links, in contrast to the rapid “insight” learning associated with propositional knowledge. What is clear, however, is that there would be considerable selection pressure to favor the ability to represent the environment in richer ways and to put together pieces of existing knowledge to draw new conclusions – precisely the characteristics we ascribe to the human learning system.
R3.2. Propositional models are not in line with what is know about the brain
Several commentators considered that we had ignored or understated the importance of biological data, in particular brain data. It is true that the primary conclusions we reached were based on behavioral data. However, it is not the case that we ignored brain data. Rather, we argued that at our present state of neuroscientific understanding, the brain data are not definitive with regard to the central debate between single- and dual-system models. That is, the available brain data are to some extent open to interpretation. Of course, neuroscientists have had much more experience in fitting their data to the prevailing dual-system model of learning, so it is perhaps not surprising that several critiques were based on long-standing assumptions regarding the mapping between brain structures and psychological mechanisms.
The most common assumption implicit in the present commentaries was that consciousness and higher cognitive functions are exclusively mediated by the cerebral cortex (Hall; Miles et al.; Olsson & Öhman). Such a “corticocentric perspective” has been challenged recently in this journal by Merker (Reference Merker2007). He reviews evidence that natural or experimental loss of cortical tissue does not eliminate conscious and goal-directed behavior, and points out that mid-brain structures conform more closely than cortical structures to the limited-capacity, executive role associated with reasoning and decision making. In general terms, the graceful degradation of performance that is observed with increasing damage to brain tissue is strongly suggestive of distribution, rather than localization, of function. And, as Chater notes in his commentary, many dissociations observed in lesion and activation studies are not inconsistent with a single-processing system. There is certainly little reason to believe that the functions of cortical and subcortical structures map neatly onto the cognitive and reflexive systems proposed in dual-system models of learning. Our single-system model makes no specific claims regarding particular brain mappings. What our model does predict, however, is that those neural systems underlying reasoning (whether they are organized on a topological, chemical, or some other basis) also subserve associative learning. None of the data put forward by the commentators, either from lesion studies in animals or neuropsychological case studies in humans, contradicts this prediction.
R3.3. Propositional models do not meet the normative criteria that good models should meet
Some commentators argue that good models should be falsifiable, parsimonious, and formalized and that propositional models do not possess these characteristics. In this section, we explain why we believe that propositional models are falsifiable and parsimonious. They lack formalization, but that does not impede their heuristic and predictive value.
R3.3.1. Propositional models are not falsifiable
As stated in section 2.1 of this response, reliable demonstrations of associative learning in the absence of relevant propositions (i.e., propositions about the relations in the world that drive learning) would raise serious doubts about the validity of the propositional approach (but would not necessarily provide evidence for link formation). We noted that such demonstrations are only convincing if sensitive measures are used to probe all relevant propositional knowledge that participants might have. Most past research does not satisfy this criterion. A more powerful approach, taken by Perruchet (Reference Perruchet1985) is to demonstrate that participants have relevant propositional knowledge that contradicts the observed learning effects. The Perruchet effect, if it is demonstrated to be an effect of associative learning (and not due to a performance effect or some other factor) could not be accommodated within the propositional approach.
Tests of the propositional approach must also satisfy one other requirement. As pointed out in section 3.1 of the target article, because propositions can be stored in memory after their truth has been evaluated, they can influence behavior as the result of memory activation even at times when their truth is not evaluated. Hence, in order to demonstrate learning in the absence of relevant propositional knowledge, one should take into account not only propositions that are truth evaluated at the time the learned behavior is emitted, but also those evaluated earlier in training, before the behavior is observed. Baker, Baetu, & Murphy [Baker et al.] argue that, because propositions can be coded in memory and can then influence later behavior in an automatic way, the propositional approach is rendered unfalsifiable. However, carefully controlled experiments can be set up in which new relations are introduced and propositions about those relations are tested online, at regular intervals during the experiment. The available evidence suggests that learned behavior will be observed only after participants report propositions about the relations that were implemented (e.g., Dawson & Biferno Reference Dawson and Biferno1973). When retrospective rather than online measures of propositional knowledge are used, researchers should make sure that the measures are sensitive enough to pick up knowledge that participants might have held in the past. This can be done by minimizing the delay between learning and testing.
Finally, in contrast to Livesey & Harris, we do believe that research on learning in the absence of relevant propositional knowledge (e.g. outside of awareness) is crucial in testing the propositional account. If convincing evidence for this type of learning can be found, it would raise serious doubts about models in which propositional processes are the only ones that can support associative learning.
R3.3.2. Propositional models are not parsimonious
In our target article (sect. 6.1), we argued that the propositional approach to associative learning is more parsimonious than a dual-process approach because “no approach that needs two systems can be more parsimonious than an approach that proposes only one of those systems, no matter how parsimonious the second system might be.” Nevertheless, some commentators have still suggested that the propositional approach is not parsimonious.
Dwyer et al. argue that the propositional approach lacks parsimony because it needs additional systems in order to account for phenomena such as habituation and perceptual learning. We agree that, within the propositional approach, learning cannot take place by propositional processes alone, but also requires perception and memory. However, a very similar argument seems also to apply to link models. There may still be some researchers who assume that link formation can provide a full account of learning or behavior. It does seem more realistic, though, to assume that link models, like propositional models, need to be extended by other processes in order to explain (non-associative influences on) behavior. These necessary additional processes of perception and memory provide mechanisms for the phenomena referred to by Dwyer et al.
Livesey & Harris argue that a propositional approach to learning is not parsimonious because propositional processes are effortful. A system in which all learning depends on effortful processes might end up using more resources than a system in which learning is sometimes outsourced to an effortless link-formation process. There are, however, several counterarguments to this position. First, as we pointed out earlier in this response (see sect. R3.1), having a primitive, effortless link-formation process that registers relations in the world could lead to an overload of information, resulting in chaotic behavior. It could also lead to action tendencies that are contrary to those arising from the cognitive system. The propositional encoding of relations in the world could provide the necessary buffer against such chaotic and hence wasteful behavior. Second, one can overestimate the expenditure of resources that propositional learning processes would require. Although truth evaluation is an effortful process, it can also be conducted in a “quick and dirty” manner. When there is little opportunity or motivation, not all information that is relevant for truth evaluation will be taken into account, thus minimizing the resources that are used to evaluate propositions about relations in the world. Moreover, once a truth evaluation has been completed, the resulting proposition can be stored in memory. On later occasions, this proposition can influence behavior without being evaluated again. In sum, propositional processes could provide the right balance between learning what needs to be learned and expenditure of cognitive resources.
R3.3.3. Propositional models are not formalized
There are two reasons why we do not mind the current lack of formalization of propositional models. First, we do not believe that formalization is a normative criterion that all models should meet. Second, those who prefer formalized models are free to formalize the aspects of propositional models that can be formalized. With regard to the first argument, the quality of a model cannot be judged by its degree of formalization. It is easy to imagine a fully formalized model that is unable to account for, or correctly predict, any instance of associative learning. The quality of a model is determined by its heuristic and predictive value, that is, its ability to (a) account for and organize existing knowledge and (b) to correctly predict new empirical results. Formalization can help achieve these aims, but it is neither necessary nor sufficient.
Witnauer et al. argue that, because propositional models are not formalized, precise predictions are difficult to derive from propositional models. In response, we would like to point out that formalized (link) models of associative learning also do not necessarily allow for precise, unequivocal predictions. Most of these models include a variety of free parameters. Various predictions can be derived from a model by varying the value of parameters or by adding new parameters.
Despite these arguments, we acknowledge that formalization can also have benefits such as making explicit the assumptions and inferences that are made when accounting for or predicting a certain finding. It is therefore important to realize that various aspects of propositional models can and have been formalized. Formalization can be coined both in terms of logical operations (e.g., the modus tolens inference; see Beckers et al. Reference Beckers, De Houwer, Pineno and Miller2005; Mitchell & Lovibond Reference Mitchell and Lovibond2002) or mathematical expressions. As we pointed out in our target article, mathematical models of link formation such as the Rescorla-Wagner model can be regarded as mathematical formalizations of the operating principles of propositional processes. As Beckers & Vervliet correctly argue, such mathematical models have a heuristic and predictive value. What is important to realize, however, is that these models retain their value even when it is assumed that they describe propositional processes rather than link formation (see sect. 6.1 of the target article).
Other ways to formalize propositional models is by means of connectionist models and Bayes nets. It should be noted, however, that at present none of the available avenues for formalization is able to capture all aspects of propositional processes. For instance, as Gopnik correctly points out, Bayes nets and Baysian inferences do not model the fact that truth evaluation is a non-automatic process. Some commentators (e.g., Baker et al.; Schmajuck & Kutlu; Shanks; Witnauer et al.) are optimistic that future connectionist models will be able to mimic all aspects of propositional processes, but a brief look at the current status of connectionist models suggests that much progress still needs to be made.
R4. Conclusion
In this rejoinder paper, we have attempted to address the main arguments that the commentators have raised against propositional models. In our opinion, the assumption that all associative learning is mediated by propositional processes is supported by the great majority of the available data, and is not significantly undermined by the conceptual arguments that have been raised against it. We recognize that we have not responded to all of the points made in the commentaries. The diversity of views expressed, and the different disciplines from which they come (philosophy, cognitive psychology, animal learning, social psychology, and neuroscience), are testament to the central importance of learning to understanding human and animal behavior. But this diversity also means that many very interesting comments and suggestions are beyond the scope of the present response. We do believe, however, that we have addressed the main arguments raised against the central claims made in the target article, that learning is propositional in nature and there is little reason to postulate the existence of an additional link-formation mechanism.
We also acknowledge that many aspects of the propositional approach require further development. As Newell and Sternberg & McClelland point out (see also Mandler), the way in which propositional processes interact with perception and memory, and whether memory processes can as such support associative learning, needs clarification. Evidence that memory processes can produce learning independently would undermine the propositional approach. However, such evidence would not constitute evidence for a link-formation mechanism. There are also other aspects of the propositional approach that need refinement. We agree with Sternberg & McClelland that research on human learning would benefit greatly from integration with reasoning research.
Just as was the case with the target article, we anticipate that this response article will not have persuaded everyone. We would like to finish, therefore, by outlining what we see as some significant future challenges to the dual-system view. It is clear from the commentaries provided here that the postulated link-formation mechanism has a variety of forms. At one end of the spectrum, it is a simple S-R system that we share with Aplysia (Hall; Matute & Vadillo; Penn et al.; Schultheis & Lachnit). At the other end of the spectrum is quite a different mechanism, which is affected by attention and seems to be comparatively rich and complex (e.g., Dwyer et al.; Livesey & Harris; McLaren). It is sometimes advantageous to propose a simple mechanism (often on grounds of parsimony). At other times it is advantageous to propose a more complex mechanism (perhaps because the data demand it). But the system in which the links are formed cannot have the characteristics of both the Aplysia-like S-R mechanism and those of the more complex, attention-demanding, S-S mechanism – these characteristics are incompatible. Proponents of the dual-system view must, therefore, decide on the nature of the link-formation mechanism; otherwise, they may be accused of wanting to “have their cake and eat it too.”
Those who propose the Aplysia-like system must then explain, for example, why learning does not occur outside of awareness. A different problem arises for those who opt for the more complex link-formation mechanism described by the models of Rescorla and Wagner (Reference Rescorla, Wagner, Black and Prokasy1972), Mackintosh (Reference Mackintosh1975), Pearce and Hall (Reference Pearce and Hall1980), Wagner (Reference Wagner, Spear and Miller1981), and Miller and Matzel (Reference Miller, Matzel and Bower1998). This version of the dual-system approach proposes that two very complex learning systems, which have many characteristics in common and which are sensitive to the same environmental variables, operate side by side. Clarification is needed as to the crucial differences between the link-formation system and the propositional system. It is not at all clear why both of these systems are needed.
Whatever the characteristics of the link mechanism, clarification is also required as to the way in which the link system and the propositional system interact. That is, how is conflict between the output of these systems resolved? In our target article, we noted the lack of detail regarding conflict resolution in dual-system theories (although see Gawronski & Bodenhausen [Reference Gawronski and Bodenhausen2006] for an exception), and regrettably many of the present commentaries continue this noncommittal tradition. For example, Olsson & Öhman stated that the two levels of learning are “partially independent but also interacting.” Lack of testability therefore remains a fundamental weakness of the dual-system approach. We reiterate our challenge to dual-system theorists to clarify the laws or mechanisms by which they consider the output of the two systems to be gated to produce observed behavioral outcomes. We believe that such an exercise will reveal that major concessions are needed to allow dual-system models to fit with existing data (e.g., regarding awareness, cognitive load, and verbal instruction) and that, in this light, the coordinated propositional system we propose might be viewed more favorably.
We would like to end with a comment on the evidence for the link-formation mechanism. As we pointed out in the target article and reaffirmed in this response to commentaries, a close examination of the data reveals only one or two isolated phenomena that might indicate the presence of a non-propositional (perhaps link-based) learning mechanism. These include the eyeblink version of the Perruchet effect (Perruchet Reference Perruchet1985) and the odor-taste learning work of Stevenson et al. (Reference Stevenson, Boakes and Prescott1998). As Dwyer et al. concede, evidence for the link-formation mechanism is not widespread. Thus, even the proponents of the dual-system approach accept that the link mechanism is of somewhat limited explanatory value. It seems to us that, if we do indeed possess two separate learning mechanisms, then we should see evidence for both mechanisms everywhere. Why, therefore, is the evidence for the second mechanism so weak and so vanishingly small? We keep an open mind, but there seems to be an obvious and almost unavoidable conclusion, that no such mechanism exists.
ACKNOWLEDGMENTS
Preparation of this paper was supported by a Discovery Project grant DP0774395 to Chris Mitchell and Peter Lovibond from the Australian Research Council, and Grant BOF/GOA2006/001 of Ghent University and Travel Grant K.1.424.06.N.01 of the Research Foundation – Flanders (FWO – Vlaanderen, Belgium) to Jan De Houwer. Thanks also goes to Gavan McNally and Fred Westbrook for discussions on issues in neuroscience.
