Despite criticisms dating back to the 18th century, the concept of association has remained central in psychology, and the core idea has remained largely unchanged. This chapter describes a fundamental rethinking of the concept of association and of how associative models are used in psychology. While I intend my view to apply generally, associative models are most commonly used in comparative psychology, and a fundamental revision of the concept of association would lead to significant changes in the methods and practices there.
The core that has remained unchanged is the general, if not universal, assumption that association is a kind of psychological process, also called “associative processing.” This assumption has caused problems. The alternative view I advocate solves these problems, while also opening up associative models to play a more constructive role in psychology generally. It is not my goal here to argue for the view in depth,1 but to describe it and its motivations.
The general idea behind my view is that “association” is a generic term that refers to any causal relationship between representational states in a psychological process. As such, an association can be implemented by any number of mechanisms, not only by the specific mechanism of associative processing. Associative models, then, are highly abstract characterizations of the sequence of representational states that become active in a process as it operates.2 In principle, they can be applied to any kind of process. This view allows for associative models to be compatible with other kinds of models, like cognitive models. In fact, the two kinds of models can work together: each can be helpful in characterizing the other in greater detail. Associative models, so conceived, are helpful when we do not understand a process well. And in psychology, this is often the case.
In the current literature, association is treated as a kind of process. Associative models are treated as denoting a member of that kind: using an associative model to describe some process implies that the process is a member of this kind. In comparative psychology, associative models, which describe members of the class of associative processing, are usually contrasted with cognitive models that describe members of the class of cognitive processing. So, for instance, an associative process could be distinguished from a process that allows an animal to simulate another’s perspective (this will be my main example of a cognitive process here). The ability to simulate others’ perspectives allows flexible engagement in social situations, and an ability to respond to cues that are not directly present to the simulating animal, which associative processing does not. It has been assumed that the distinction between associative and cognitive processes generalizes straightforwardly: associative processes are simple and inflexible, while cognitive processes are more complex and more flexible. But this way of distinguishing the kinds of process has become problematic. I’ll discuss two arguments, one based in comparative psychology, and one in human psychology.
Buckner (2011 and Chapter 39 in this volume) has done an excellent job of developing the concerns that have arisen in comparative psychology, and bringing them to philosophical attention, so I describe them rather quickly here.3 The basic setup comes when the dichotomy between associative and cognitive processes is paired with a widespread methodological principle known as “Morgan’s Canon.” In its modern interpretation, Morgan’s Canon is taken to dictate that, ceteris paribus, models that posit simpler processes should be preferred. This means that whenever an associative model gets the behavioral predictions right, it will be the preferred model; even if a cognitive model also predicts, the associative model describes a simpler process.
Suppose, to start, that some behavior is predicted by an associative model. This, as noted, is taken to be reason to think that the process itself is associative. Research proceeds from here roughly as follows: to further test the model, the experimental task is made more complex, such that the associative model does not predict success. If the animal succeeds in this new task, the associative model is considered to be falsified, and it is replaced with a cognitive model of some kind. However, it is often the case that the associative model can be modified in a way that predicts this new behavior as well. In such a case, the process repeats. It can do so indefinitely.
For instance, research on social learning in pigeons has gone through several iterations. At each iteration, new complications are built into the task, and in turn, the associative model. In the initial experiment, pigeons acted as if they had learned that pressing a lever produced a food reward simply by watching another pigeon be rewarded for doing so. This seemed to go beyond pure associative learning, perhaps indicating an ability to simulate the perspective of the demonstrator. But then a new associative model was created to predict the effect (Zentall 1996). In a subsequent experiment, pigeons were able to differentiate between rewards that came specifically when pecking the lever or when stepping on the lever, which the modified associative model did not predict (Zentall, Sutton, and Sherburne 1996). But again, a new associative model was built to explain this behavior (Meltzoff 1996), and again, subsequently shown to be too simple by another task (Akins and Zentall 1998). This process has continued to iterate since.4
I’d like to point out here that this practice implicitly assumes that all of these tasks probe a single psychological capacity. This is usually not mentioned because each experiment is only a minor variant of the previous. But it should be made explicit: the assumption that there is a common capacity justifies the proposition that one experiment can falsify the associative model as a model of performance in the previous experiments as well as performance in the experiment where the model actually failed to predict. The evident goal is to produce a model of the process that predicts how it will respond across conditions.
We are now in position to understand how the distinction between associative and cognitive processes has become problematic. As mentioned, this distinction has traditionally been cast in terms of complexity. But this iterating, back-and-forth dynamic produces more complex associative models at each turn. As associative models become more complex, the line is blurred. Some commentators have suggested that this has undermined, or at least threatened, the distinction (e.g. Allen 2006; Penn and Povinelli 2007; Papineau and Heyes 2006). At the same time, there are concerns that the “is it associative or is it cognitive?” orientation misses the really interesting issues: what information is encoded and how the animal interprets the task itself (Smith, Couchman, and Beran 2014). Putting these concerns together, it looks bad for the distinction between association and cognition: there are no real standards for drawing the distinction, and it is not clear what we even gain by doing so.
A parallel argument appeals to research on priming and associative learning in humans. Priming is when presentation of one stimulus facilitates some task involving a second, related stimulus. For instance, if the word “lion” is flashed, a subject will read the subsequently presented word “tiger” more quickly. While the basic phenomenon of priming looks associative, there are good reasons to believe that the processes responsible are too complex to reasonably be counted as associative processes. For instance, it matters what kind of list a specific prime-target pair appears in (list context effects; e.g. McKoon and Ratcliff 1995), and it matters what task participants are instructed to perform on the prime (prime task effects; e.g. Smith, Bentin, and Spalek 2001). Associative models don’t predict either of these effects.
Associationists reply to evidence like this by pointing to instances in which associative models do predict. For instance, when Mitchell, DeHouwer, and Lovibond (2009) argue that human associative learning is propositional in a target article, the phenomenon of affective priming is mentioned repeatedly in this role in the commentary. In these replies, the fact that associative models predict some basic priming phenomena is taken to be evidence that the basic processes responsible for priming are associative, while complex-looking effects come from other processes running in parallel, as in dual-process theories generally.
The problem for this reply is that the evidence of complexity in human priming and associative learning is too pervasive to easily cordon off as effects of a distinct process.5 But it also remains the case that associative models predict the behavior of these processes in some cases. I argue that both of these facts should be taken seriously. The standard view forces a dilemma: either deny that the processes are complex, or deny that associative models are properly applied. Both horns come with costs. On the first horn, one is forced to explain away evidence of complexity that is so pervasive that it suggests that, even if there are multiple processes, the simplest processes involved are not associative (or so I argue). On the other horn, one is forced to reject potentially useful associative models; we should not simply dismiss their predictive success as accidental or irrelevant due to presuppositions about what associative models do.
Moreover, because it is not clear what behavioral evidence can settle the question of how many processes are present (or how to attribute effects to specific processes if there are many), these debates tend to trade on competing appeals to parsimony (as in Mitchell, DeHouwer, and Lovibond 2009 and commentary). It is difficult to make progress in debates over clashing parsimony claims like this (Dacey 2016b). So once we find ourselves facing this dilemma, both options have real costs, and fruitful debate becomes difficult. A view that avoids this dilemma (as mine does) would be better than the current standard view, which does not.
I have presented two arguments that the standard view of association is problematic. Really, these are two sides of the same coin. It turns out that some processes look associative in some contexts, but not in others. On the standard view, these two pieces of evidence are in tension; associative models are assumed to only describe simple processes, and if they apply at all, they apply across contexts. We see two, effectively opposite, responses to this tension. Sometimes, evidence of complexity is taken to trump previous success of associative models; sometimes the previous success of associative models is taken to trump evidence of complexity. The responses given dictate how each literature has progressed. In comparative psychology, the practice of taking evidence of complexity to falsify associative models has driven the iterating back-and-forth dynamic that has blurred the distinction between association and cognition while obscuring other questions of interest. In the human literature, researchers take both sides, though both have costs, and the debate can make little progress.
The way to avoid these problems is, I hope, clear: reject the view that associative models must denote a particular kind of process. The trouble is that there is no systematic alternative in the current literature. I now describe an alternative, and the role it gives association in psychology.
The view I advocate treats “association” as a highly abstract, generic term that could be realized by many different mechanisms. An association itself is simply a causal relation (any causal relation) between representations that become active in a process. This view of association comes from a simplified interpretation of associative models. Associative models describe only sequences of representational states (often in terms of spreading activation) and/or the influence that variables in learning have on those sequences. Representations, in turn, are specified purely by content, so no specific kind of representation is required. This is all the models themselves say. I propose that we take this at face value. The claim that associative models denote associative processes is not necessitated by the most basic interpretation of the models. We can simply jettison this claim, and with it, the commitments that have caused the problems discussed above. Associative models are partial descriptions of a process; they are highly abstract, and only tell us about these specific features of the process.
So specified, an association could be underwritten by any number of neural or psychological processes: in principle, it could be produced by the application of a rule in an algorithm, the manipulation of a mental map, or a mental model (to name a couple cognitive process). Returning to the pigeon social learning example, an associative model could describe the (cognitive) process of simulating the perspective of conspecifics. It could do so in two ways: either by describing the sequences of representations that occur within the simulation, or by describing the representational states that trigger the simulation (more on this below). Similarly, any variables can be included in an associative learning model. An associative learning model that includes variables like contact, latency between events, and patterns of movement as relevant to learning might provide a precise, mathematical description of a process for learning mechanical-causal relations between objects (e.g. Spelke 1990).
Cognitive models generally include greater causal detail, so they still indicate kinds of processes (the specific kind depends on the specific model), while associative models do not. This is why both can be applied to the same process; they can describe the same causal structure at different levels of abstraction. Associative models and cognitive models are not differentiated by the kind of process they describe; they are differentiated by what they say about the process, and as a result, the descriptive and explanatory work they do. Whether you use an associative model or a cognitive model depends on the question you are asking and the information you have, not on the process you are describing.
In any single task, this view does not change the behavioral predictions that a specific associative model makes. The difference comes when we consider how a model applies across conditions. As discussed above, if we take associative models to denote associative processes, it implies that an associative model should predict behavior across conditions in which the relevant process drives behavior. Instead, I argue that we should restrict the scope of each associative model to the specific task for which it was designed. So if an associative model fails to predict behavior in a new task, that does not falsify it as a model of the process in previous tasks where it does predict. It still may be the case that the model accurately describes the sequence of representational states that the process follows, and the learning variables it responds to, in those original tasks. There may be a different associative model that predicts behavior in the new task, or several, or none.
If we take these different tasks to probe the same capacity, the goal is to integrate the associative models that predict behavior in different tasks into a model of a single mechanism that can produce them all. In principle, it is possible that multiple associative models predict behavior even in a single task. In these cases, only one of these can be the right one. The way to tell which is to find a set of associative models, one in each task, that could be a product of a single mechanism. This requires constructing a model of the mechanism, and determining which sequences it would produce and which learning variables it would respond to. Again, in principle, there may be multiple such sets. Building these mechanism models and arguing for one of them is difficult, likely involving an inference to the best explanation, including background theory. So performance on one task does bear on the evaluation of an associative model of performance in another task, but not as a direct confirmation or falsification. Instead, the question is whether two different associative models of performance on different tasks could be products of the same mechanism.
For example, we can build associative models that predict behavior in all of the experiments that we believe probe pigeons’ capacity to learn socially. Some of these may include a simulation, and some not. We then compare these models, and look for a set that could be the product of a single mechanism. If the best explanation posits a mechanism that simulates the perspective of conspecifics, the resulting characterization of that mechanism would show what sequences of states are (and are not) included in the simulation, and what conditions do (and do not) trigger the simulation. Thus, we would not only have systematic evidence the capacity is present (contrasted with the current back-and-forth that emphasizes single experiments over integrating evidence), and we would have a much more detailed characterization of the capacity (contrasted with the vagueness of many current appeals to cognitive processes).
So associative models set abstract constraints on more detailed models of the mechanism. Any more detailed model must follow the sequence of representations and/or respond to the learning variables specified by the associative models. Far from excluding other kinds of models, like cognitive models, associative models become an important part of the process of developing those models. We do not need to draw a distinction between association and cognition in order to use both kinds of models.6 This view also explains why associative models can predict the behavior of a complex process (like priming) in some conditions, but not all. So it solves both problems with associative models discussed above.
No discussion of associative processing can proceed without addressing the issue of simplicity (or parsimony) as a scientific virtue. This issue bears on the discussion in two ways. Firstly, one might argue that associative models must denote associative processes by applying Morgan’s Canon when interpreting the models (not just when choosing them). Associative processes are the simplest processes that an associative model could describe, and thus, the standard view would be reinstated. However, this would require a very strong, universally applicable version of Morgan’s Canon. There are good reasons to reject such a view, which I cannot canvas here (see Fitzpatrick 2008 and Chapter 42 in this volume; Dacey 2016b). Given the problems caused by the resulting view of association, along with these more general concerns about Morgan’s Canon, this argument cannot justify a reinstatement of the standard view.
Secondly, on my view simplicity remains a virtue of associative models. This is not because the process is simple, but because the model is simple. The model is simple because it abstracts away from details of the way the process operates. This is valuable for three reasons. First, we often don’t need those details; for instance, an associative model is often sufficient for pragmatic concerns like prediction and control. Second, because associative learning models focus on specific variables, we can characterize their influence in precise, mathematical terms (models like this are often derived from the Rescorla-Wagner model [Rescorla and Wagner 1972]). Many, probably most, psychological models lack this precision, so the precision of associative models is valuable, and not readily replicable. My view of association allows associative models to lend their mathematical rigor to other kinds of models, following the process described in the last section. Third, treating associative models as partial models avoids theoretical commitments about the nature of psychological processes that we may not have good evidence for. This is valuable when we do not understand processes well. This is often the case in psychology.
This view is a fundamental revision of the way association is viewed in contemporary cognitive science. By way of concluding, I would like to stress that it is not, as it may seem, a break from the history of the concept; it is, in many respects, a return to a view as old as association itself. Views like the current standard view have traditionally been the most common interpretation of association (as I said above, this core of the concept has changed little), but it is only since roughly the mid-20th century that this has been the only interpretation. One can find views of association much like mine in both the empiricist associationist and behaviorist traditions.
Among the associationists, the view is made explicit by Thomas Brown (1840). Brown first published his main work on the topic in the year of his early death in 1820, and was prominent for several decades after. He occupied a unique place at the intersection of different traditions, but he was a close follower of Hume in many respects: he was an associationist and a staunch Humean about causation. At the time of his writing, the dominant view was that association was a real link between thoughts that drove them to follow a sequence. He disagreed, largely because of his Humeanism: just as there are no causal “links” between successive events in the world, there are no associative “links” between successive thoughts in the mind. Association (like causation) is just invariant sequence. The “associative link,” argued Brown, is explanatorily vacuous and metaphysically dubious.
In this respect, Brown was more Humean than Hume. For his part, Hume does struggle with the obvious inconsistency between his rejection of causal links in the world and his acceptance of links in the mind in an appendix to the Treatise (Hume 1978). He never resolved this concern in print, and his writing betrays an ambiguity between a view like Brown’s and the associative link view. This ambiguity was never really resolved in general, and remained, in various guises, through the associationist tradition.7
When behaviorism rose to prominence in the 1920s, the behaviorists retained association as the core concept of psychology, but reframed it as a relation between external, observable stimuli and responses rather than mental representations. Even then, the same ambiguity remained. In this tradition, it was a divide between those who argued that the specific physical stimuli or muscle movements were the relata of association, and those who argued that association had to be a relation between more abstract patterns of stimulation and response.
The distinction between two interpretations of association is made most clearly by Edwin Guthrie, Edward C. Tolman, and Edward S. Robinson. Guthrie (1959) sees his emphasis on behavior as an abstraction designed to exclude complicating factors, much like the use of constructs such as frictionless planes in physics. Thus, for Guthrie, behaviorism abstracts away from details about mental states, rather than denying their existence. He also argues that views that treat association as a “mechanism” (by which he means an invariant sequence of specific physiological/chemical changes) are much too specific to actually explain behavior. Tolman (1932; 1948) attacks the same view, characterizing it as treating the brain as a neural “switchboard” which simply connects stimulus to response. For both Tolman and Guthrie, associations are formed between abstract patterns of stimulation and response that can be realized (to put it in modern terms) by many different specific mechanisms. Robinson (1932) agreed, and (perhaps presciently) argued that associationism was dying out because the mistaken view dominated.
While the details vary with historical context, the spirit is the same: association is not a particular, concrete mechanism that drives sequences in thought, it is something more abstract. My view updates this insight. Association is not a kind of psychological process; it is a generic causal relation between representational states of the system. Associative models, in turn, describe the sequence of states a process moves through and the variables it responds to in learning at a very abstract level. This view of association avoids problems with the current standard view, and gives association a more productive role in psychology.
1 I do so elsewhere: Dacey 2016a; in prep.
2 I distinguish associative models from neural network models like neural circuit models and distributed connectionist models. It is common to treat these kinds of models as the same (e.g. Bechtel and Abrahamson 1991; Clark 1993), a view which I call reductive associationism (see Dacey 2016a, and note 3 below). There are certainly interesting similarities between these kinds of models, but they tell us different things about the target system. Associative models describe the relationships between representational states (i.e. localist associative models); neural network models describe relationships between distributed “parts” of representations, activation in neurons, or activation in neural areas. Because they tell us different things about the system, they do different work, and they should not be equated.
3 In this volume (Chapter 39), Buckner sketches his own solution to these problems as well. In effect, we have made opposite moves: he treats associative models as being less abstract than cognitive models, while I treat them as being more abstract. The way I see it, Buckner’s view amounts to a version of reductive associationism (see note 2). I reject this view because I take associative models and neural network models to each do different, valuable work. And so they should be kept separate. Buckner has helpfully characterized a role for neural network models, while I characterize the role for associative models here.
4 See Dacey 2016a for more detail.
5 I develop this argument in more detail in Dacey (in prep). I discuss human psychology because the argument would be substantially more controversial if applied to any nonhuman animal. For similar arguments, see Mitchell, DeHouwer, and Lovibond (2009) and Mandelbaum (2015).
6 One should not mistake my view for eliminativism (e.g. Gallistel 1990, 2000) or instrumentalism about association. Associations are real features of psychological processing. Moreover, there may be processes that can only be described with associative models (exactly as associative processes are thought to be). But these processes are not defined by being describable by associative models, they are defined by their failure to be describable by cognitive models. A name that better reflects this, like “non-cognitive processing,” would lead to less confusion (see also Buckner, Chapter 39 of this volume). There is a related question about whether processes cluster such that there is a distinct kind here. Buckner (2013) is optimistic that they do, while I am not. But we won’t know until much more empirical work is done, and the view of association I advocate will help answer this question, rather than presuming that there is such a class.
7 I discuss this aspect of the associationist tradition in more detail in Dacey (2015).
Akins, C. K., and Zentall, T. R. (1998) “Imitation in Japanese quail: The role of reinforcement of demonstrator responding,” Psychonomic Bulletin & Review, 5(4), 694–697.
Allen, C. (2006) “Transitive inference in animals: Reasoning or conditioned associations,” in Hurley, S., and Nudds, M. (eds.) Rational animals, Oxford: Oxford University Press.
Bechtel, W., and Abrahamson, A. (1991) Connectionism and the mind: An introduction to parallel processing in networks, Cambridge, MA: Blackwell.
Brown, T. (1840) Lectures on the philosophy of the human mind, Hollowell: Glazier, Masters, and Smith.
Buckner, C. (2011) “Two approaches to the distinction between cognition and ‘Mere Association,’” International Journal of Comparative Psychology, 24, 314–348.
——— (2013) “A property cluster theory of cognition.” Philosophical Psychology, 28(3), 307–336.
Clark, A. (1993) Associative engine: Connectionism, concepts, and representational change. Cambridge, MA: MIT Press.
Dacey, M. (2015) “Associationism without associative links: Thomas Brown and the associationist project,” Studies in History and Philosophy of Science Part A, 54, 31–40.
——— (2016a) “Rethinking associations in psychology,” Synthese, 193(12), 3763–3786.
——— (2016b) “The varieties of parsimony in psychology,” Mind & Language, 31, 414–437.
——— (in prep) “Associative models and the mechanisms of priming”
Fitzpatrick, S. (2008) “Doing away with Morgan’s Canon,” Mind & Language, 23, 224–246.
Gallistel, C. R. (1990) The organization of learning, Cambridge, MA: Bradford Books/MIT Press.
——— (2000) “The replacement of general-purpose learning models with adaptively specialized learning modules,” in Gazzaniga, M. S. (ed.) The cognitive neurosciences, 2nd ed., Cambridge, MA: MIT Press.
Guthrie, E. R. (1959) “Association by contiguity,” in Koch, S. (ed.) Psychology: A study of a science Vol. 2, New York: McGraw-Hill.
Hume, D. (1978) A treatise of human nature, Oxford: Clarendon Press.
McKoon, G., and Ratcliff, R. (1995) “Conceptual combinations and relational contexts in free association and in priming in lexical decision and naming,” Psychonomic Bulletin & Review, 2(4), 527–533.
Mandelbaum, E. (2015) “Attitude, inference, association: On the propositional structure of implicit bias,” Noûs 50(3), 629–658.
Meltzoff, A. N. (1996) “The human infant as imitative generalist: A 20-year progress report on infant imitation with implications for comparative psychology,” in Heyes, C. M., and Galef, B. G. (eds.) Social learning in animals: The roots of culture, New York: Academic Press.
Mitchell, C. J., De Houwer, J., and Lovibond, P. F. (2009) “The propositional nature of human associative learning,” Behavioral and Brain Sciences, 32(2), 183–198.
Papineau, D., and Heyes, C. (2006) “Rational or associative? Imitation in Japanese quail,” in Hurley, S., and Nudds, M. (eds.) Rational animals, Oxford: Oxford University Press.
Penn, D. C., and Povinelli, D. J. (2007) “Causal cognition in human and nonhuman animals: A comparative, critical review,” Annual Review of Psychology, 58, 97–118.
Rescorla, R. A., and Wagner, A. R. (1972) “A theory of Pavlovian conditioning: Variations in the effectiveness of reinforcement and nonreinforcement,” in Black, A. H., and Prokasy, W. F. (eds.) Classical conditioning II, New York: Appleton-Century-Crofts.
Robinson, E. S. (1932) Association theory to-day, New York: The Century Co.
Smith, J. D., Couchman, J. J., and Beran, M. J. (2014) “Animal metacognition: A tale of two comparative psychologies,” Journal of Comparative Psychology, 128(2), 115–31.
Smith, M. C., Bentin, S., and Spalek, T. M. (2001) “Attention constraints of semantic activation during visual word recognition,” Journal of Experimental Psychology: Learning, Memory, and Cognition, 27(5), 1289–1298.
Spelke, E. S. (1990) “Principles of object perception,” Cognitive Science, 14, 29–56.
Tolman, E. C. (1932) Purposive behavior in animals and men, New York: Appleton-Century-Crofts.
——— (1948) “Cognitive maps in rats and men,” Psychological Review, 55(4), 189.
Zentall, T. R. (1996) “An analysis of imitative learning in animals,” in Heyes, C. M., and Galef, B. G. (eds.) Social learning in animals: The roots of culture, New York: Academic Press.
Zentall, T. R., Sutton, J. E., and Sherburne, L. M. (1996) “True imitative learning in pigeons,” Psychological Science, 7(6), 343–346.