Research ReportFrontopolar cortex mediates abstract integration in analogy
Introduction
Analogical reasoning is a complex form of reasoning in which concepts from one situation are mapped onto another situation resulting in new inferences and explanations. More specifically, an analogy is a mapping between the abstract structure of one situation and the abstract structure of another situation. Analogical reasoning, such as the reasoning involved in comprehending the analogy, “The atom is like the solar system,” is a relational form of reasoning that is essential for learning, understanding our environment, and generating novel ideas (Dunbar and Blanchette, 2001, Gentner, 1999, Holyoak, 2005). Analogical comprehension has been regarded as a key component of intelligence (Sternberg, 1977), inductive reasoning (Holyoak and Thagard, 1995), and everyday discourse (Blanchette and Dunbar, 2002). Like all forms of reasoning, analogical reasoning involves a complex array of cognitive processes. Two processes that are central to analogy are manipulation of component terms in working memory and integration of relations in order to abstract a schema for the whole analogy (Holyoak, 2005). Here, using fMRI, we investigate the neural correlates of abstract relational integration in analogical thinking while keeping constant the demand for manipulation of component terms in working memory.
Starting with Milner's research on concept attainment in frontal lobe patients (Milner, 1963), neuroscientific research on complex human reasoning has shown that reasoning involves brain-based mechanisms for temporary maintenance of information as well as manipulation of this information. Temporary maintenance and manipulation of information are thought to be primary functions of working memory (Baddeley and Hitch, 1974, Curtis and D'Esposito, 2003, Owen, 1997). Disentangling working memory processes from other cognitive components of reasoning has become an important issue for cognitive neuroscience. For example, researchers have sought to isolate working memory processes such as goal-directed maintenance (Boroojerdi et al., 2001, Koechlin et al., 1999, Ruff et al., 2003, Wharton et al., 2000), attentional switching (Koechlin et al., 1999), and inhibiting distractor interference (Kroger et al., 2002). These investigations have revealed that working memory is a key component of the reasoning process. Furthermore, each of these studies identified reasoning processes mediated by prefrontal cortex (PFC) that are cognitively and neuroanatomically distinct from working memory processes.
Analogical reasoning is a form of relational reasoning in that it requires understanding how terms, and the relations between terms, are related to each other (Holyoak, 2005). In processing an analogy, a reasoner must first identify the relations that are present within each of the items/situations being compared; that is, the reasoner must be able to see how component elements relate to each other within each item/situation. These relations within items/situations have been called conventionalized semantic relations (Gentner, 1998) because they usually refer to a conventional way in which two things are related to each other. For example, in the analogy, “The atom is like the solar system,” there is a conventionalized semantic relation between component elements of the atom (i.e., electrons revolve around the nucleus), and there is a conventionalized semantic relation between component elements of the solar system (i.e., planets revolve around the sun).
Additionally, in order to successfully appreciate an analogy, a reasoner must comprehend that the two conventionalized semantic relations both represent the same abstract relation. In the example analogy, a reasoner must comprehend that the two conventionalized semantic relations both represent the abstract relation, revolves around. Gentner (2000) delineates, in computational terms, that the abstract relations that tie analogies together are higher-order relations in that they take other, lower-order relations (e.g., conventionalized semantic relations) as arguments. As noted by Blanchette and Dunbar (2000), these higher-order relations are abstract in that they do not depend on the specific surface-level properties of the elements of the analogy. Rather, they are similarities at the level of underlying structure. Thus, they can be abstracted from or applied to any item/situation within which the component elements are related to each other in a similar way.
Noting that abstractly similar relations are present within two different items/situations is a key component of analogical thought. This process, known as analogical mapping, requires more than simply identifying conventionalized semantic relations within each item/situation. Analogical mapping also involves an alignment process whereby the component elements of one item/situation are aligned one-to-one with corresponding elements of the other item/situation (Gentner, 1983, Holyoak and Thagard, 1997).
Recently, we have identified categorization as a potential mechanism underlying this alignment process in analogy (Green et al., in press). Specifically, using four-word stimuli, we provided evidence that mapping one item/situation onto another involves grouping component terms into categories. We have suggested that category relations facilitate appropriate one-to-one alignment of component terms between two items/situations. For example, in the solar system analogy, planets and electrons are grouped together in the category, satellites, and sun and nucleus are grouped together in the category orbited objects.
In addition to our recent empirical findings (Green et al., in press), connections between categorization and analogical reasoning have been suggested by previous models of categorical alignment (Bassok et al., 1998, Wisniewski and Bassok, 1999), and previous accounts of analogical mapping (Hess, 1966, Holyoak and Thagard, 1997, Sternberg, 1977). Bassok and colleagues (Bassok et al., 1998, Wisniewski and Bassok, 1999) have demonstrated that categorically related items such as apples and oranges can be readily compared because category co-membership makes them easier to mentally align with each other. Holyoak and Thagard (1997) have observed that “mapped elements… are typically similar but not identical” (p. 6). These authors provide the example of a military analogy wherein Saddam and Hitler are both members of the category, leaders, and invade and occupy are both members of the category, acts of war. Several other researchers have also suggested that analogies and categories may be importantly related (Bowdle and Gentner, 2005, Gentner, 1998, Gentner and Markman, 1997, Gick and Holyoak, 1983, Hummel and Holyoak, 2003). However, with the exception of Green et al. (in press), previous accounts have not addressed categorization as a means by which analogical mapping is accomplished, rather these accounts have addressed categorization only as a potential end result of analogical mapping. Thus, the role of categorization as a mechanism for accomplishing analogical mapping has not been clearly delineated.
Turning now to the neural underpinnings of analogy (a form of relational reasoning), convergent evidence has come from neuropsychological and neuroimaging research. Neuropsychological investigations of reasoning in patient populations have found specific deficits in relational integration of terms concurrent with damage to prefrontal cortex (Boroojerdi et al., 2001, Morrison et al., 2004, Waltz et al., 1999). In addition, recent neuroimaging research, using visuo-spatial stimuli, has specifically implicated left anterior prefrontal cortex in tasks demanding relational reasoning (Christoff et al., 2001, Kroger et al., 2002, Wharton et al., 2000). Taken together, these data clearly delineate the prominent role that frontal cortex, specifically left anterior prefrontal cortex, plays in the kind of relational thinking that underlies analogy.
Recently, Bunge et al. (2005) used four-word analogies to investigate the neural substrates of analogical thinking. Four-word analogies (also referred to as verbal proportional analogies) are word sets composed of two word-pairs such as ‘Hand:Glove + Foot:Sock’ (hand is to glove as foot is to sock). Bunge and colleagues presented four-word sets to a group of participants who were instructed to indicate whether each set constituted a true analogy. In the Bunge et al. (2005) experiment, subjects saw the first word-pair (e.g., Hand:Glove) and determined whether there was a conventionalized semantic relation between the two terms of this word-pair. The second word-pair (e.g., Foot:Sock) was presented afterwards. In the interim between the two word-pairs, subjects received one of two cues. Subjects were either cued to simply judge whether a conventionalized semantic relation was present within the second word-pair (no analogy condition), or to judge whether the two word-pairs represented abstractly similar relations (analogy condition). These authors were able to dissociate semantic retrieval of individual relations (no analogy condition) from subsequent processes of manipulating and integrating these relations (analogy condition). Retrieval of semantic information preferentially activated anterior left inferior PFC (aLIPC), whereas determining abstract similarity between the two word-pairs preferentially activated left frontopolar cortex.
The present investigation adopted this distinction between semantic retrieval and the subsequent processing that brings the pieces of an analogy together as a whole. The focus of the present investigation was on parsing this subsequent processing. Specifically, we were interested in distinguishing working memory processes (i.e., maintenance and manipulation of individual terms and relations in working memory) from abstract relational integration of multiple relations in order to form a single higher-order relation. By controlling for the relevant working memory demands, we sought to determine whether abstract relational integration could be isolated from the working memory processes that facilitate analogical reasoning. We also sought to determine whether cortical activity during analogical reasoning expresses a pattern consistent with our previous finding (Green et al., in press) that the manipulation of component terms in working memory during analogical reasoning involves categorizing these terms.
In the present investigation, we examined analogical reasoning using stimuli that were sets of four words. We varied the tasks that the subjects were required to perform with these words. In all conditions subjects responded True or False. Subjects were instructed to respond True if all the relations explicitly delineated for the condition were present, and to respond False if one or more of these relations was not present. All stimuli were previously determined True or False for their respective conditions with >90% agreement through pilot testing among a separate group of 27 participants. The different conditions are schematically represented in Fig. 1. In the Analogy condition (ANA), subjects saw four-word sets such as ‘Planet:Sun + Electron:Nucleus’ arranged in a rectangle as in Fig. 1. In True four-term analogies, a conventionalized semantic relation was present between the two terms of the word-pair on the left (planet revolves around sun) and an analogous conventionalized semantic relation was present between the two terms of the word-pair on the right (electron revolves around nucleus). Subjects responded True if the four-word set included two conventionalized semantic relations (left and right word-pairs), and if the left and right word-pairs, taken together, constituted an analogy.
In the ANA condition, no category relations were explicitly delineated. However, as noted above, previous research with four-term analogies (Green et al., in press) has suggested that categorization is a necessary mechanism sub-serving analogical mapping. Thus, we designed the ANA condition based on the premise that evaluating four-word analogies would involve grouping component terms into categories. For example, in Fig. 1, both planet and electron can be grouped into the category, satellites, and both sun and nucleus can be grouped into the category, orbited objects.
In order to control for the working memory demands of analogical reasoning, we devised a non-analogical control task, called CAT, to mimic the number and kind of relations present in an analogy. In the CAT condition, subjects saw four-word sets such as ‘Duck:Water + Cow:Milk.’ In this example, a conventionalized semantic relation exists between the terms of the word-pair on the left (duck swims in water) and a conventionalized semantic relation exists between the terms of the word-pair on the right (cow gives milk). At a categorical level, duck and cow are both farm animals, and milk and water are both liquids. Thus, as in four-term analogies, there are two conventionalized semantic relations and two categorical relations. However, unlike four-term analogies, there is no analogical relation (i.e., duck is not to water as cow is to milk). This is because the two conventionalized semantic relations do not have the requisite abstract similarity. Subjects were not instructed to look for analogical relations in the CAT condition. Subjects responded True if two conventionalized semantic relations (left and right word-pairs) and two categorical relations (top and bottom word-pairs) were present.
A third condition, called SEM, only involved conventionalized semantic relations. In this condition subjects were instructed to respond True if conventionalized semantic relations existed within the left and right word-pairs. This task neither involved assessing whether an overall analogy was present nor whether any category relations were present. Accordingly, this condition involved fewer working memory demands than the ANA and CAT conditions. The SEM condition was devised to facilitate our examination of categorization in analogical reasoning. Specifically, by comparing analogical reasoning (ANA) to two control tasks, one of which involved categorization (CAT) and one of which did not (SEM), we sought to test our hypothesis that categorization contributes to the working memory manipulations involved in analogical reasoning.
Based on the cognitive demands of each or our tasks, we made specific predictions regarding neural activity associated with these tasks. We predicted that the analogical reasoning task (ANA), which involved abstract relational integration, would engage left frontopolar cortex because of this region's apparent role in abstract relational integration (Boroojerdi et al., 2001, Bunge et al., 2005, Goel et al., 1997). Critically, if analogical reasoning involves abstract relational integration that is distinct from maintenance and manipulation of lower-order relations in working memory, then left frontopolar cortex should be preferentially recruited by the ANA task relative to the control task matched for working memory (CAT). Conversely, in brain regions other than frontopolar cortex, we hypothesized that activation elicited by the ANA task would be highly similar to that elicited by the CAT task. A comparison of the neural activation associated with these two tasks outside of frontopolar cortex was intended to provide an index of the extent to which the CAT task was successful in mimicking the demands (other than abstract relational integration) of the ANA task.
Because the CAT task involved both identifying conventionalized semantic relations and categorization of terms, while the SEM task only involved identifying conventionalized semantic relations, we predicted that the CAT task would preferentially recruit parieto-frontal regions associated with semantic categorization (Aizenstein et al., 2000, Elliott et al., 1999, Grossman et al., 2002, Koenig et al., 2005, Patalano et al., 2001) relative to the SEM task. Additionally, if analogical reasoning involves working memory manipulations for the purpose of categorizing terms, then the brain regions preferentially recruited by the CAT task relative to the SEM task should also be preferentially recruited by the ANA task relative to the SEM task.
Our specific predictions are consistent with the theoretical framework of functional organization in prefrontal cortex provided by Christoff and Gabrieli (2000). These authors have hypothesized a rostro-caudal hierarchy of prefrontal function such that rostral regions mediate abstract processing of the cognitive products of more caudal regions. In addition, Ramnani and Owen (2004) have specifically argued that frontopolar cortex mediates abstract relational integration of information produced at more caudal cortical regions. As noted above, research on abstract reasoning is consistent with this hypothesis as it has strongly implicated frontopolar cortex in tasks involving abstract reasoning (Boroojerdi et al., 2001, Bunge et al., 2003, Bunge et al., 2005, Goel et al., 1997, Strange et al., 2001, Wagner et al., 2001). Our prediction that abstract relational integration in the ANA task would recruit more rostral areas of prefrontal cortex (i.e., frontopolar cortex) than the CAT or SEM tasks is consistent with the proposed rostro-causal hierarchy of prefrontal function.
Section snippets
Results
The 14 participants whose data were retained for analysis performed at a mean response accuracy level of 92.86% for the three conditions (SEM 95.36%, CAT 92.02%, ANA 91.19%). Data analyses were restricted to trials for which the correct response was True. Effect size estimates for the behavioral analyses were computed using partial η2.
Selective recruitment of frontopolar cortex by abstract relational integration
Investigations into the functional underpinnings of complex human reasoning have only recently been undertaken and many questions about the nature of complex reasoning have not yet been answered. One fundamental question is whether the mental integration of abstractly related items and relations comprises cognitive processes beyond holding and manipulating items and relations in working memory. To address this question, we used fMRI to compare an analogical reasoning task (ANA) to another
Conclusion
By experimentally controlling for manipulation of terms in working memory, the present investigation sought to isolate the neural signature of abstract relational integration in an analogical reasoning task. We identified a region at the frontal pole of the brain that is selectively active for the abstract relational integration component of analogical reasoning. Both analogical reasoning and a task of explicit categorization engaged a parieto-frontal network of regions previously associated
Participants
Eighteen participants (9 females and 9 males, mean age = 22.5 years) took part in the fMRI study and were paid $20. Four participants (2 males, 2 females) were eliminated from subsequent analysis due to severe motion throughout the scanning session or poor behavioral task performance (<90%). All participants were right-handed, reported no significant abnormal neurological history and had normal or corrected-to-normal visual acuity. Informed written consent was obtained from all participants
Acknowledgments
This research was funded by a grant to Kevin Dunbar from Dartmouth College. The authors thank S. Grafton, J. Moran, and C. Bennett for thoughtful comments on earlier versions of the manuscript and technical expertise. Requests for reprints should be sent to Adam E. Green or Kevin N. Dunbar, Department of Psychological and Brain Sciences, 6207 Moore Hall, Dartmouth College, Hanover, New Hampshire, 03755 (E-mail: [email protected]; [email protected]).
References (82)
- et al.
A comparison of neural circuits underlying auditory and visual object categorization
NeuroImage
(2002) - et al.
Inhibition and the right inferior frontal cortex
Trends Cogn. Sci.
(2004) - et al.
Working memory
- et al.
Adding apples and oranges: alignment of semantic and formal knowledge
Cogn. Psychol.
(1998) - et al.
Functional–anatomical study of episodic retrieval using fMRI. II. Selective averaging of event-related fMRI trials to test the retrieval success hypothesis
NeuroImage
(1998) - et al.
Rostrolateral prefrontal cortex involvement in relational integration during reasoning
NeuroImage
(2001) - et al.
Persistent activity in the prefrontal cortex during working memory
Trends Cogn. Sci.
(2003) - et al.
The in vivo/in vitro approach to cognition: the case of analogy
Trends Cogn. Sci.
(2001) - et al.
Common regions of the human frontal lobe recruited by diverse cognitive demands
Trends Neurosci.
(2000) - et al.
Ventromedial prefrontal cortex mediates guessing
Neuropsychologia
(1999)