How Can Functional Neuroimaging Inform Cognitive Theories?

Localization

Skepticism has often been expressed about whether localizing any specific cognitive process in the brain could tell you anything at all about how that process actually works (i.e., could tell you anything about cognition):

“Finding a cell that recognizes one’s grandmother does not tell you very much more than you started with: after all, you know you can recognize your grandmother. What is needed is an answer to how you, or a cell, or anything at all, does it. The discovery of the cell tells one what does it, but not how it can be done” (Mayhew, 1983, cited in Boden, 1988, p. 6).

“If the mind happens in space at all, it happens somewhere north of the neck. What exactly turns on knowing how far north?” (Fodor, 1999).

More specifically, Coltheart (2004, p. 28) argued that “In order to localise the modules of a cognitive system, we must first know what the system’s modules actually are. So we must begin with a model and then seek to do localisation research. Imaging is not contributing to modelling here; instead, imaging depends on (prior) modelling.” This exact point is echoed by Wixted and Mickes (2013, this issue); their view is that research using functional neuroimaging to localize specific cognitive processes in the brain always requires reference to some particular cognitive-level theory about cognition for the interpretation of the neuroimaging data. So, as Wixted and Mickes say, such localization studies “do not shed light on cognitive theories but instead depend on cognitive theories to interpret the data—an interpretation that is only as valid as the cognitive theory on which it is based” (p. 104).

What are the implications of this point for Blumstein’s article (Blumstein & Amso, 2013, this issue), which is entirely about localization? The general conclusion of that article is that cognitive operations recruit networks rather than specific regions and that areas considered to be domain-specific are actually recruited in a domain-general fashion so that similar behavioral profiles may depend on different neural architectures. The target here appears to be Fodor’s conception of modularity—Blumstein and Amso write “There are a number of defining characteristics of a theory of modularity. We provide fMRI evidence that challenges two of these properties: domain specificity and fixed neural architecture (Fodor, 1983)” (p. 44). But Fodor explicitly states in his book that he is not proposing any defining characteristics of modularity. In his view, a module can still be a module even if it has no fixed neural architecture (see Coltheart, 1999, for further discussions of this point). Thus, any evidence about any cognitive module not having a fixed neural architecture is evidence about the brain and not about cognition—it is evidence about how cognitive modules are implemented in the brain, not evidence against any cognitive theory proposing modularity. Their conclusion, “evidence from cognitive neuroscience challenges a theory of modularity” (p. 46), may be true if they are addressing “anatomical modularity,” but it is not true if they are addressing “cognitive modularity.” For example, finding that a single region—the inferior frontal gyrus—is activated when a variety of different linguistic operations are being performed (syntactic processing, speech processing, semantic processing) does not license the conclusion that there is a single cognitive module responsible for all three different kinds of linguistic processing, and it is compatible with the view that, in all senses of the term cognitive module, syntax, semantics, and speech depend upon separate cognitive modules.

A similar point can be made regarding Wager and Atlas’s article, which, like Blumstein and Amso’s article, is largely about localization. Their first sentence makes an extremely important point: “Neuroimaging can inform cognitive theories to the extent that particular patterns of brain activity are sensitive and specifically associated with particular types of cognitive processes” (p. 91). To explore this point, imagine we had two conflicting cognitive theories about how people perform some cognitive task. According to Theory P, all the cognitive processing that people carry out when performing this task is parallel processing. According to Theory S, some of the processing is serially operating. Now suppose we could identify a particular pattern of brain activity—say, activity in a particular specific brain region—that was sensitive to and specifically associated with a particular type of cognitive process, namely, serial processing. If we found that this region was activated when the task is being performed, that would be strong evidence against Theory P and would offer support to Theory S. Notice here that the data being used to test the cognitive theories are localization data; but Coltheart (2004) and Wixted and Mickes have argued that cognitive theories cannot be tested by localization data.

The solution to this apparent conflict is, of course, that we do not have any evidence showing that any particular form of brain activation is a sign that some particular type of cognitive operation is being performed. Perhaps cognitive neuroscience in the future might succeed in obtaining such evidence. That might make it possible for localization data to be used for testing cognitive theories in the future. But that is certainly not the current situation.

It is helpful to note a distinction drawn by Bergeron (2007, 2008) between cognitive workings of some brain region and cognitive uses of that brain region. Bergeron’s idea is that we can specify the cognitive workings of any brain area by identifying the several kinds of cognitive computations that area is able to perform, and we can specify the cognitive uses of any brain area by identifying the cognitive functions that require the performance of any of these kinds of cognitive computations. If a particular cognitive theory claims there is some brain region that is the site of some cognitive subsystem, and if we knew what the workings of that region were (i.e., what kinds of computations it can or cannot perform), then if the kinds of computations proposed in the cognitive theory did not correspond to what we know about the cognitive workings of that region, we would have a case in which localization data could be used as evidence against a cognitive theory. Of course, it is not currently possible to argue along these lines because cognitive neuroscience is not yet able to characterize the cognitive workings of any regions of the brain believed to be associated with cognition, even if it can characterize cognitive uses of specific brain regions.

In general, then, existing fMRI work that has aimed at localization of cognitive functions is not relevant to this special section, because such work does not inform cognitive theory. Instead, it is informed by cognitive theory.

Testing Cognitive Theories With Functional Neuroimaging Data

There are two different ways in which cognitive neuroimaging data can be used to test cognitive theories: “Attempts at testing cognitive models with brain imaging data can take either of two forms. The work might involve just a single model, with its aim being to confirm or refute the model; or instead the work might involve competing models, with its aim being to adjudicate between these” Coltheart (2010a, p. 62).

A crucial issue here concerns what Mole and Klein (2010) refer to as the consistency fallacy. Their point is that data that are consistent with some theory cannot, just in virtue of this consistency, be offered as evidence in support of that theory—something additional is needed. For a body of data to provide evidence for a hypothesis, the data must not only be consistent with the hypothesis, they must also count against the contradictory of the hypothesis. It is very common for articles involving cognitive neuroimaging to end with statements solely about consistency between cognitive theory and neuroimaging data: for example, “these two studies support the possibility that memory reactivation can be induced by environmental cues during sleep” (Levy & Wagner, 2013, this issue, p. 76), and “functional neuroimaging evidence accumulated during the last two decades . . . is more consistent with the components framework than with the other two frameworks” (Cabeza & Moscovitch, 2013, this issue, p. 49). The only way to avoid the consistency fallacy is to point out explicitly what pattern(s) of neuroimaging data that are inconsistent with the theory could plausibly have been obtained in the neuroimaging study. If there are no plausible outcomes of the study that, if obtained, would have been inconsistent with a theory, then consistency between theory and data cannot be offered as evidence in support of the theory.

When a neuroimaging study seeks to evaluate a single cognitive theory, avoidance of the consistency fallacy is achieved by identifying at least one plausible outcome of the study that, if it had been obtained, would have been inconsistent with what that theory predicts. When, instead, the aim is to adjudicate between competing cognitive theories, one must perform this exercise for all of the theories concerned. Several articles in this issue are about adjudication between competing cognitive theories. How do they fare when considered in the light of the consistency fallacy?

Cabeza and Moscovitch discuss three competing cognitive-level hypotheses about the nature of associations and dissociations between performance on explicit versus implicit memory tests. They argue in favor of one of these hypotheses, the “components framework,” but only on the grounds that it is more consistent with the results of neuroimaging work, and as noted above, mere consistency with neuroimaging cannot in itself be offered as support for any cognitive theory. They do argue that this framework makes predictions that could be falsified: “the component process theory, in conjunction with relational memory theory, predicts that [the hippocampus] should be activated for relational memory tasks but not for nonrelational memory tasks (both declarative and nondeclarative)” (p. 51). So, at least in principle, an experiment could have a possible outcome that would be inconsistent with component process theory—namely, an outcome in which the hippocampus is activated by relational memory tasks. Demonstrating the potential for theory falsification in this way helps to deal with the consistency fallacy, and so it would be a useful practice for cognitive neuroimaging researchers to adopt.

Both Rugg and Thompson-Schill (2013, this issue) and Wixted and Mickes also emphasize just how strong the assumptions must be to get some cognitive theory to generate predictions about what should happen in functional neuroimaging experiments. The general point is that, if neuroimaging data are to be used in this particular way to test any cognitive theory, the cognitive theory has to be embellished by a proposal of the form: Cognitive Process C is implemented in Brain Region X and nowhere else in the brain, and Brain Region X subserves Cognitive Process C and no other cognitive process. This is what Wager and Atlas referred to as the combination of sensitivity (only one brain region is sensitive to a particular form of cognitive processing) and specificity (activation in this brain region is specific to only one particular form of cognitive processing). Are there findings that illustrate just this kind of relationship between cognition and brain? Blumstein and Amso argue that, at least in the case of language, there can be absence of specificity, as shown by findings that a particular brain region is implicated in various distinct forms of language processing, and also absence of sensitivity, in the sense that more than one brain region can be implicated in a single specific form of language processing,

Several other articles in this issue also apply neuroimaging data to the task of adjudicating between cognitive theories: Levy and Wagner on two competing cognitive theories of how memory-based inferences support generalization of knowledge from multiple prior experiences, White and Poldrack on two competing cognitive theories of category learning, and Park and McDonough on three competing cognitive theories of ageing. Space does not permit a detailed analysis of each of these articles, but as an exercise one might like to answer the following questions about each of these three papers:

With respect to Question 2, Wixted and Mickes point out that not all theories of cognition make any predictions about brain activity, and of course, whenever this is the case, no neuroimaging data can inform the cognitive theory.

They do offer one example of a cognitive theory that does make predictions about brain activity: the theory of embodied cognition, which holds that cognition is grounded in perception. This cognitive theory predicts that brain regions active when the relevant form of perception is occurring will also be active when the corresponding form of cognition is occurring. The particular form of cognition they consider is episodic memory, and they refer to Wheeler, Peterson, and Buckner’s (2000) finding that remembering the visual and auditory aspects of studied items elicits activity in the same regions of visual and auditory cortex that were activated during the initial perception of those items. What is not clear, however, is whether the cognitive theory predicts merely that there will be some regions in common between activations elicited by perception and activations elicited during performance of the episodic memory task, or whether it predicts that the regions showing activations in the two tasks will be identical. So we don’t know what to make of Wheeler and colleagues’ finding that there was bilateral activation near the superior temporal gyrus during the perception of auditory material, whereas such activation was seen only in the left hemisphere during the recall of auditory material. Is this evidence against an embodied-cognition account of episodic memory or not?

Rugg and Thompson-Schill also offer embodied-cognition theory as an example of a cognitive theory, and they agree with Wixted and Mickes that embodied-cognition theory can be tested by neuroimaging data. However, they disagree with what the outcome of that testing has been. Wixted and Mickes believe that this theory is supported by the neuroimaging data, whereas Rugg and Thompson-Schill believe that the theory (at least in its strong form) is disconfirmed by such data. They focus on the domain of color in their exploration of embodied-cognition theory, discussing in particular the neuroimaging results of Hsu, Kraemer, Oliver, Schlichting, & Thompson-Schill (2011), who compared a color-perception task to a color-memory task.

The color-perception task required judgments of the lightness or darkness of chromatic or achromatic visual stimuli; subtraction of brain activations in the achromatic condition from brain activations in the chromatic condition were used to identify brain regions involved in color perception.

The color-memory task involved presentation of two printed words (names of objects with characteristic colors) at the bottom of a computer monitor, followed 4 s later by a third printed word (also the name of an object with a characteristic color) at the top of the screen. The subject’s task was to decide which of the two bottom words was most similar in real world color to the top word.

The question of interest is to what degree are the regions identified as being involved in color perception also active in the color-memory task. The cognitive theory being tested states that accessing a conceptual feature engages the same processes that are active when the feature is directly expressed. However we need to clarify what we mean by “same processes.” The embodied-cognition framework does not predict that the brain regions activated in the color-memory task will be identical to the brain regions associated with color perception in the color perception task, because there are various cognitive operations required in the color memory task that are not carried out in the color perception task. The prediction instead is that the brain regions activated in the color-memory task will include the brain regions associated with color perception. But does this require that the brain regions activated in the color-memory task should include all of the brain regions associated with color perception or just some of them? The actual result was that some brain regions were activated in the color-memory task but not in the color-perception task, some brain regions were involved in the color-perception task but not in the color-memory task, and a small amount of brain tissue was activated in both tasks (10 voxels in the left fusiform gyrus; just 2 voxels in the left lingual gyrus). Is this pattern of results consistent or inconsistent with the embodied-cognition theory?

Wixted and Mickes and Rugg and Thompson-Schill have developed a neural embellishment of a cognitive theory to try to make predictions about what should happen in functional neuroimaging experiments. It appears that neither attempt at this difficult task has been fully successful, because in both cases it is not clear whether what these studies actually found is consistent or inconsistent with the predictions of the theories being tested.

Conclusions

Contributors were invited to provide examples of research with fMRI offering either findings from neuroimaging research that importantly informs a cognitive theory or findings that have been touted as informing a cognitive theory but that do not actually tell us anything new about that theory.

As it turned out, no contributor offered any example of the second kind. However, the question of whether the first kind of finding has occurred or can occur was much discussed.

Two key points were made by Wixted and Mickes. The first was that neuroimaging studies aiming solely at the localization of cognitive processes do not inform cognitive theories. Instead, they are informed by cognitive theories, because such theories are needed if the neuroimaging data are to be successfully used for localization of cognitive functions.

Their second key point was that some cognitive theories make no predictions about the results of any functional neuroimaging experiments. Given this, what can we do to try to bring functional neuroimaging data to bear on the informing of cognitive theory? Such theories need to be embellished in ways that allow predictions to be made about the results of neuroimaging experiments. This is far from easy to do: Wixted and Mickes and Rugg and Thompson-Schill attempt this but do not seem quite to have succeeded.

One way to neurally embellish cognitive theories is advocated by Wager and Atlas: The investigator must be willing to make proposals about the mapping of mind to brain that involve both sensitivity (a claim that Brain Region X will always be active when Cognitive Process C is being executed) and specificity (a claim that Brain Region X will not be active except when Cognitive Process C is being executed). Do we find such claims about the mapping of mind to brain implausibly strong? If we do, is there any satisfactory way of deriving predictions from cognitive theories about the results of neuroimaging experiments that do not require such strong claims?

Whatever solutions to this problem are adopted, we must still be aware of and avoid the consistency fallacy. It is common for cognitive neuroimaging articles to end with a statement of the form “Our data therefore are consistent with Theory C.” It is rare for such articles to contain such sentences as “Our experiment might have yielded Result X, which would have been inconsistent with Theory C”. It would be good to see this done routinely.