Author's personal copy
Reasoning and Problem Solving: Models 35
Reasoning and Problem Solving: Models
A K Barbey and L W Barsalou, Emory University, which in turn can cause an avalanche, supports the Atlanta, GA, USA prediction that an avalanche is possible during winter.
ã Finally, new predictions and explanations can be 2009 Elsevier Ltd. All rights reserved. generated by analogical reasoning, whereby relations from one domain are mapped onto another domain. For example, students can conceptualize particle Introduction motion in an atom through an analogy to the solar Neuroscience research has increasingly provided evi- system. dence that informs theories of reasoning. We begin by defining classic forms of reasoning; and summariz- Theoretical Perspectives on Reasoning ing psychological contemporary theories. We then in the Psychological Literature review empirical evidence from neuroscience that The psychological literature contains various theories bears on these theories. Finally, we summarize cur- rent challenges and identify promising areas for that address the cognitive architecture and symbolic future research. representations that underlie reasoning. After review- ing these theories, we turn to neuroscience evidence
that bears on them. Definition and Types of Reasoning
Cognitive Architecture Reasoning is a hallmark of human thought, support- ing the process of discovery that leads from what is One major distinction can be drawn between theories known or hypothesized, to what is unknown or that view the mind as containing specialized rea- implicit in one’s thinking. Reasoning can take the soning modules, versus theories that view the mind form ofdeductive inference, whereby the evidence as containing general-purpose reasoning systems. guarantees the truth of the conclusion. Alternatively, According to the modular view, the mind consists when reasoning depends on conditions of uncertainty, of specialized modules that are unavailable to con- it takes the form of inductive inference, whereby the scious awareness and deliberate control (cognitively evidence provides only limited support for the truth impenetrable), and that are only able to process of the conclusion. The deductive inference that a specific types of information (information encapsula- storm has emerged is supported by evidence that it is tion). Advocates of this view have proposed a diversity raining, whereas observing lightning clouds provides of modules that underlie reasoning, including modules only limited support for the inductive inference that for semantic inference, communicative pragmatics, it will rain. Many forms of reasoning typically depend social exchange, intuitive numbers, spatial relations, on conditions of uncertainty, including problem naive physics, and biomechanical motion. If cogni- solving, causal reasoning, and analogical inference. tive architecture emerges from an underlying neural Problem solving refers broadly to the inferential architecture, then strong modular views predict that steps that lead from a given state of affairs to a desired the neural systems for reasoning should be relatively goal state. For example, deciding who to vote for in a localized, implementing modules that are cognitively presidential election, diagnosing a patient on the basis impenetrable and informationally encapsulated. of observed symptoms, or preparing for a mountain Alternatively, dual-process theorists propose that climbing expedition all depend on problem solving. reasoning is based on two general-purpose systems: an Problem solving often requires the process of planning, associative system and a rule-based system. The asso- namely, formulating a method for attaining a desired ciative system uses basic cognitive operations such as goal state. Planning a mountain climbing expedition, association, similarity, and memory retrieval to pro- for example, requires scheduling the season of the duce primitive judgments quickly and unconsciously. expedition by predicting the consequences of taking The rule-based system reflects more evolutionarily the trip during different seasons. This is accomplished advanced mechanisms that implement reasoning pro- by modeling the situation and observing the conse- cedures deliberately and consciously. For example, quences of possible actions (e.g., the occurrence of inductive reasoning largely depends on the retrieval snow in winter introduces new challenges). and evaluation of world knowledge, whereas deductive
Prediction and explanation further depend on causal reasoning depends on rule-based, formal procedures. reasoning (i.e., the ability to infer causal relations). For The dual-process theory further motivates the example, knowledge that cold winter weather can cognitive demand hypothesis: when people have little cause an accumulation of snow in the mountains, time and limited processing resources, incentives,
Encyclopedia of Neuroscience (2009), vol. 8, pp. 35-43
Author's personal copy
36 Reasoning and Problem Solving: Models and/or external aids available for making a judgment, of a chair into amodal representations (e.g., feature they only use the associative system when reasoning. lists, semantic nets), processed by classic symbolic Conversely, when people have more time and greater mechanisms (e.g., predication, argument binding). processing resources, incentives, and/or external aids Thus, amodal theories predict that conceptual proces- available, they also use the rule-based system. The sing will recruit brain regions outside sensorimotor dual-process model predicts that reasoning recruits areas that underlie language and rules, including left different neural systems, depending on cognitive prefrontal and superior temporal regions implicated demand. When tasks are easy, the associative system in formal, rule-based operations. is sufficient for correct reasoning and primarily In contrast, embodied theories of knowledge pro- recruits left inferior frontal gyrus, especially Broca’s pose that knowledge and meaning are grounded in area, in addition to the temporal lobes and posterior modality-specific representations. Increasing empirical parietal association cortex. As tasks become more evidence from both the cognitive and the neuroscience difficult, the rule-based system becomes required for literatures suggests that modality-specific representa- correctreasoning and recruits the prefrontal cortex, tions underlie higher level cognition. According to this especially the ventrolateral subregion, which has been framework, concepts are represented by simulating implicated in rule maintenance. the modality-specific states that were initially acti- A second theory that provides a general-purpose vated during perception, action, and interoception. account of reasoning is grounded in a somewhat Embodied theories propose that simulations are different pair of systems: the linguistic system and organized at a higher level by simulators that inte- the conceptual system. According to this view, the grate information across a category’s instances. Over brain’s language system initially produces relatively time, for example, visual information about how superficial information about a reasoning problem, cakes look becomes integrated in a ‘cake’ simulator, such as word associates, syntactic structures, etc. As along with gustatory information about how cakes linguistic forms become generated, their meanings taste, somatosensory information about how they become increasingly represented in the conceptual feel, motor programs for interacting with them, emo- system. A key assumption of this approach is that tional responses to experiencing them, and so forth. superficial processing based on linguistic representa- The result is a distributed system throughout the tions may often be sufficient for adequate reasoning brain’s modality-specific areas that establishes con- performance. When it is not sufficient, conceptual ceptual content for the general category of ‘cake.’ representations must be generated to produce more Thus, embodied theories predict that conceptual sophisticated reasoning. Thus, consistent with the processing will recruit a broadly distributed system dual-process model, this view predicts that differ- of modality-specific brain regions. ences in the cognitive demands of a reasoning task Embodied theories further motivate the task speci- (i.e., superficial vs. deep processing) will differentially ficity hypothesis. According to this hypothesis, the engage neural systems. neural areas underlying a particular type of reasoning, In general, both dual coding frameworks predict such as deduction, may show little in common, as the that reasoning will recruit neural systems that sup- specific materials and tasks vary. Because different port two forms of coding. One important set of sys- materials and tasks produce different patterns of tems underlies language processing, including the left modality-specific activation, the same general form frontotemporal language system. A second set of of reasoning does not show a single, stable pattern, important systems underlies conceptual processing, and no common areas may emerge. In contrast to the mental simulation, and imagery, including bilateral ‘cognitive demand hypothesis,’ this view predicts that sensorimotor areas. significant differences in the task and/or materials – even those that do not affect cognitive demand – are
likely to dominate neural activation more than the Representation in Reasoning type of reasoning performed. A second major distinction can be drawn between theories of reasoning that are based on amodal versus Mental Models versus Mental Logic modality-specific representations. According to stan- dard theories, the modality-specific states that are A third major distinction can be drawn between the- active while perceiving an entity are redescribed into ories of reasoning that are based on visuospatial models amodal representations that bear no correspondence versus theories that are based on logical operations. to the neural systems producing them. For example, Mental-model theory proposes that deductive and the concept ‘chair’ is represented by redescribing the inductive reasoning depend on spatially organized modality-specific states that underlie the perception mental models. According to this view, an argument is
Encyclopedia of Neuroscience (2009), vol. 8, pp. 35-43
Author's personal copy
Reasoning and Problem Solving: Models 37 evaluated by generating alternative models of its C are B) activates a bilateral frontoparietal system, premises, where each model represents possible cir- including bilateral dorsal (BA 6) and inferior (BA 44) cumstances that render the premises true. If all the frontal lobes, bilateral superior and inferior parietal models constructed from the premises are consistent lobes (BA 7), and bilateral occipital lobes (BA 19). with the truth of the conclusion, then the argument This pattern of activation is also found during the is judged to be deductive or valid. If, however, only processing of spatial information, and is similar to the a limited number of models render the conclusion neural activity observed while people make transitive true, then the argument is judged to be inductive or inferences about geometrical shapes. probabilistic. Thus, deductive reasoning recruits left and right Given the proposed role of mental models in both prefrontal cortex asymmetrically as a function of deductive and inductive reasoning, this theory predicts familiarity. Across both familiar and unfamiliar that these forms of reasoning will recruit common deduction problems, left prefrontal cortex is generally neural systems. Specifically, mental-model theorists active, suggesting that this region is necessary for have predicted that deductive and inductive reasoning deductive inference. Conversely, right prefrontal cor- will both primarily recruit right-hemisphere regions. tex is engaged only when a problem involves unfamil- Furthermore, to the extent that mental models repre- iar semantic content or a conclusion that conflicts sent situations spatially, this view predicts that parietal with prior beliefs (e.g., No harmful substances are and occipital regions implicated in visuospatial proces- natural / All poisons are natural / Therefore, no poi- sing will be engaged. sons are harmful). Whereas language may often dom- In contrast, mental logic theory offers an account inate familiar reasoning, language and spatial/visual of deductive reasoning that is based on the applica- processing may often be central for unfamiliar tion of formal deductive rules according to formal reasoning. syntactic operations. Thus, rule theorists have pre- The brain systems that implement deductive dicted that reasoning is likely to recruit left prefrontal reasoning also depend on whether a reasoning prob- and superior temporal regions implicated in formal, lem produces correct versus incorrect conclusions. rule-based operations. Research on this issue has used inhibitory belief pro- blems, namely, problems whereby individuals must
inhibit a highly accessible belief that could interfere Empirical Evidence on Reasoning from with correct reasoning (e.g., No addictive things are the Neuroscience Literature inexpensive / Some cigarettes are expensive / There- fore, some cigarettes are not addictive). Drawing We next present empirical evidence from neuro- a correct conclusion on these problems requires that science that bears on the psychological theories just individuals (1) detect the conflict between their prior reviewed. We begin with research on deductive beliefs and the logical inference, (2) inhibit the prepo- reasoning, and also address research that compares tent response associated with their belief bias, and deductive and inductive reasoning. We then review (3) engage the appropriate reasoning mechanisms. research on problem solving, causal reasoning, and In contrast, drawing an incorrect conclusion on analogical reasoning. these problems results from failing to detect the con- flict between beliefs and logical inference, and/or fail- Deductive Reasoning ing to inhibit the prepotent response associated with The brain systems that implement deductive reasoning a belief bias. depend on whether the reasoning problem consists When people draw correct conclusions on inhibi- of familiar versus unfamiliar semantic content. When tory belief problems, right inferior prefrontal cortex reasoning about familiar semantic content (e.g., All dogs becomes active. When they draw incorrect conclu- are pets / All poodles are dogs / Therefore, all poodles sions, ventromedial prefrontal cortex is active instead. are pets), a left frontotemporal system is engaged, Activation in right inferior prefrontal cortex when including left inferior frontal cortex (BA 47), left drawing correct conclusions appears to reflect the middle/superior temporal cortex (BA 21/22), and detection and/or resolution of the conflict between left temporal pole (BA 21/38). Previous research belief and logic. Conversely, activation in ventro- has implicated this system, not only in deductive medial prefrontal cortex when drawing incorrect con- reasoning, but also in memory and language tasks clusions appears to reflect the role of nonlogical that employ familiar semantic content. In general, mechanisms, perhaps associated with greater affective linguistic processing appears central to all these tasks. processing. In contrast, reasoning about unfamiliar semantic In summary, the neural systems that underlie deduc- content (e.g., All P are B / All C are P / Therefore all tion vary considerably, depending on task factors and
Encyclopedia of Neuroscience (2009), vol. 8, pp. 35-43
Author's personal copy
38 Reasoning and Problem Solving: Models cognitive demand. Consistent with the task specificity tasks that employ visuospatial materials. Such activa- hypothesis, the areas that support deduction vary with tions suggest that people use visuospatial representa- the familiarity of the materials, and with whether tions, such as Venn diagrams or Euler circles, to belief violations occur and are detected. Consistent support deductive reasoning on categorical syllogisms. with the cognitive demand hypothesis, more neural In addition, activation was also found in right anterior areas are recruited for difficult unfamiliar problems cingulate (BA 24/32), implicating attention and execu- than for easier familiar ones. Consistent with the tive control in deductive reasoning. Inductive reasoning modularity view, left prefrontal cortex generally (induction minus deduction) revealed activation in left appears active across most deduction paradigms, dorsolateral frontal (BA 8 and 10) and right insular suggesting that it is essential for deductive inference. cortices. These regions are known to be involved in probabilistic reasoning tasks that require the estima- Deductive versus Inductive Reasoning tion of relative frequencies and other quantities (e.g., Many experiments have asked individuals to perform How fast do race horses gallop?). both deduction and induction in the same experi- A subsequent neuroimaging study extended the ment, so that the neural circuits underlying these methods and design of the study just described to a two types of reasoning could be distinguished. One deduction task that employed conditional statements neuroimaging experiment used a categorical syllo- instead of categorical syllogisms (e.g., If he is an gism task consisting of three conditions – deduction, electrician, then he spent two years in night school / induction, and baseline – to assess this issue. In the He is an electrician and owns a computer / Therefore, deduction condition, individuals received a valid or he spent two years in high school). In contrast to the invalid categorical syllogism (e.g., None of the bakers categorical syllogisms used in the previous study, play chess / Some of the chess players listen to opera / visuospatial processing did not seem relevant to Therefore, some of the opera listeners are not bakers). these conditional syllogisms. Thus, these researchers The task was to indicate whether the conclusion was predicted that the new task and materials would not valid (i.e., whether the conclusion was guaranteed by recruit brain regions that perform visuospatial pro- the truth of the premises). In the inductive reasoning cessing. Consistent with this prediction, deductive condition, individuals only received invalid categori- reasoning (deduction minus induction) revealed a cal syllogisms (e.g., Some of the computer program- major focus of activation in right inferior frontal mers play the piano / No one who plays the piano cortex (BA 44), in right anterior cingulate (BA 24), watches soccer matches / Therefore, some computer and in right middle temporal cortex (BA 21). These programmers watch soccer matches). The task was to researchers concluded that this frontotemporal system indicate whether the conclusion was probable or constitutes a logic-specific network in the right hemi- improbable. The deduction and induction conditions sphere that is comparable to the language-specific were each compared to a baseline condition, in which network in the left hemisphere. Specifically, these a categorical syllogism with anomalous semantic researchers argued that this right-hemisphere system content was presented (e.g., All the engineers own a implements a calculus of mental transformations that computer / None of the engineers has been to school / underlie formal deduction.
Therefore, all the people who own computers are In a further comparison, inductive reasoning married). Importantly, the categorical syllogisms (induction minus deduction) revealed large intense were fully counterbalanced across individuals so that activations in the left inferior frontal (BA 47) and the same materials occurred in every condition. left insular cortices, in addition to left posterior cin-
Deductive reasoning (deduction minus baseline) gulate (BA 31), parahippocampal (BA 36), left medial produced activation in the left dorsolateral frontal cor- temporal (BA 35), and superior and medial prefrontal tex (BA 6), broadly consistent with the left frontal cortex (BA 9). These areas are broadly consistent activation observed by other researchers for familiar with the left frontal (BA 8 and 10) and insular areas semantic content. When deduction was compared found in the previous study, known to be involved directly to induction, however, a different pattern in the recall and evaluation of familiar world knowl- emerged, namely, activation in bilateral posterior edge. Consistent with standard theories of induc- regions, with a right-hemisphere prevalence, including tive reasoning, the recall and evaluation of familiar associative visual cortex (e.g., cuneus, precuneus, mid- world knowledge appear to play central roles during dle and superior occipital gyri), as well as right superior induction. parietal lobe (BA 7) and thalamus. These areas have Another neuroimaging study developed a category been reported for visuospatial tasks that require form learning task to assess the component processes discrimination and imaginative operations. These of inductive reasoning. Of primary interest were the areas have also been reported for deductive reasoning neural systems that support rule application versus
Encyclopedia of Neuroscience (2009), vol. 8, pp. 35-43
Author's personal copy
Reasoning and Problem Solving: Models 39 rule inference during category learning. Stimuli con- methodological differences between the tasks and sisting of novel animals were presented to individuals, materials used across studies. Furthermore, the fact who were asked to judge whether all the animals in a that different materials are often used on deduction set were from the same category. In the rule applica- and induction tasks suggests that different activations tion condition, a rule was provided that specified the for these two forms of reasoning may often be driven criteria for category membership. In the rule inference by differences in materials rather than by differences condition, individuals had to infer the rule with in reasoning. no instruction. Each condition was further divided Consistent with the task specificity hypothesis, the into an easy and difficult condition based on the neural processes that underlie a given type of computational demands of the task. These research- reasoning (e.g., deduction) may depend more on the ers found that rule inference (rule inference minus tasks and materials used than on the type of reasoning rule application) preferentially recruited bilateral per se. Although left prefrontal cortex is often active hippocampus, an area in which activation is modu- on deduction tasks (as described in the previous sec- lated by stimulus novelty. In contrast, rule application tion), it was not active for some of the studies dis-
(rule application minus rule inference) revealed acti- cussed in this section. Perhaps the one consistent vation in the presupplementary motor area (BA 8). finding so far is that reasoning about familiar materi- This area is implicated in the anticipation of motor als tends to utilize left-hemisphere language and activity and likely reflects an anticipatory response to knowledge networks. Conversely, reasoning about category exemplars, a response that is absent when less familiar problems tends to utilize bilateral systems the categorization rule is unknown. that include right-hemisphere mechanisms. This pat- In addition to the episodic encoding of novel tern is consistent with theories that postulate two dif- stimuli, inductive reasoning requires the generation ferent reasoning systems, one that processes language, and testing of hypotheses. For example, inferring the and one that processes spatial/visual information. basis for category membership requires generating and testing possible rules, such as ‘‘has spots on the Problem Solving and Planning abdomen’’ or ‘‘has only two appendages’’ for a cate- gory of fictional animals. To assess the neural systems Depending on whether a problem-solving task is well that underlie hypothesis selection, these researchers structured or ill structured, different brain systems evaluated the task by difficulty interaction: {hard rule are engaged. On well-structured problems, such induction minus hard rule application} minus {easy as the Tower of Hanoi, the starting state, the goal rule induction minus easy rule application}. This state, and possible transformations are specified comparison assessed the effects of increased difficulty completely. For example, the starting state consists due to subtle variations in the stimulus features of of three pegs mounted on a platform and three disks animals that satisfied or violated the category mem- of varying sizes stacked in descending order on the bership rule. The result was an activation in right first peg. The goal state is to stack the disks in des- lateral orbital prefrontal cortex (BA 47 and BA 11), cending order on the third peg. The possible transfor- an area implicated in complex reasoning tasks such as mations are restricted to moving disks such that analogical and metaphorical transfer. Across different (1) only one can be moved at a time, (2) any disk types of reasoning, difficult problems often activate that is not removed must remain on a peg, and (3) a this area. larger disk cannot be placed on a smaller disk. In such In summary, current findings from neuroimaging tasks, well-structured planning typically recruits left studies of deduction and induction again suggest that prefrontal cortex, including frontopolar, dorsolateral, the neural bases of reasoning are highly sensitive to and ventrolateral regions. the particular tasks and materials employed. For In contrast, an ill-structured planning problem is example, deductive reasoning from conditional state- specified incompletely. In the Multiple Errands Task, ments engages a right frontotemporal system, thought for example, individuals are taken to an unfamiliar to support the application of formal deduction rules, neighborhood and asked to complete errands, such whereasdeductive reasoning from categorical syllo- as buying a loaf of bread, according to the follow- gisms activates a parieto-occipital system that sup- ing rules: you are to spend as little money as pos- ports visuospatial processing. Furthermore, the latter sible (within reason) and take as little time as result differs from the findings of other researchers, possible (without rushing excessively). You are not who have found that a left frontotemporal system to use anything not bought on the street (other than supports deductive inference in a categorical syllogism a watch) to assist you. You may perform task steps in task. The observed differences on the same fundamen- any order. In such tasks, ill-structured planning typi- tal reasoning process – deduction – probably reflect cally activates right dorsolateral prefrontal cortex
Encyclopedia of Neuroscience (2009), vol. 8, pp. 35-43
Author's personal copy
40 Reasoning and Problem Solving: Models
(DLPFC), in contrast to the left prefrontal regions classic Michotte launching event. In a launching that support well-structured planning. event, a ball travels horizontally across a computer The neural systems underlying planning are further screen and collides with a ball located in the center. differentiated into subsystems for plan formation ver- The collision results in the second ball ‘launching’ sus plan execution. One finding is that DLPFC sup- away from the first, horizontally, across the screen, ports sequential operations during plan formation, thereby eliciting the perception that the first ball whereas medial ventral prefrontal cortex (MVPFC) caused the second to move. One study compared the plays a motivational role in plan execution. More neural response produced by the launching event to specifically, DLPFC appears to support the genera- the response elicited by a control event in which the tion of hypotheses and the construction of plan steps, first ball passed below the second ball without a whereas the MVPFC appears to support the affective collision (noncausal condition). Of primary interest processing required for plan execution (e.g., initiative were the neural systems engaged when individuals and determination). judged either (1) the presence or absence of causation Lesion studies complement the neuroimaging stud- versus (2) the direction of the ball’s motion. A reliable ies just reviewed. One representative study evaluated increase in medial frontal activation occurred for judg- problem solving in patients with frontal lobe lesions ments of causality relative to judgments of ball move- (FLLs) on the Water Jug problem. Patients had to ment. Moreover, this increase occurred during both the construct unique action sequences that transferred causal and noncausal conditions, suggesting that specific quantities of water between jugs of different the signal increase was specifically associated with sizes (e.g., Jar A ¼ 8 units of water, Jar B ¼ 5 units of the process of making a causal judgment, not with the water, Jar C ¼ 3 units of water). Specifically, indivi- perception of actual causality. duals had to construct an action sequence that Another study evaluated whether causal perception achieved a goal state given by the experimenter and causal inference rely on common or distinct (e.g., Jars A and B must contain 4 units of water, hemispheric regions. Two callosotomy (split-brain) and Jar C must contain 0 units). The researchers patients and a group of neurologically intact patients found that patients with FLLs struggled to make were tested. Of primary interest was neural activity in counterintuitive moves that were required to solve the left versus right hemispheres during (1) the percep- the task but that appeared to deviate from the desired tion of causal events (i.e., the Michotte launching goal state. Furthermore, left and bilateral FLL event) and (2) causal inference tasks when the relation patients were more impaired than right FLL patients between a candidate cause and an observed effect had were, suggesting that poor performance in the Water to be inferred (rather than perceived directly). Percep-
Jug task was primarily linked to left DLPFC damage. tion of causality and causal inference depended on In summary, these findings demonstrate that prob- different hemispheres of the divided brain. Whereas lem solving and planning generally depend on pre- causal perception engaged the right hemisphere, frontalsystems. Because DLPFC appears important causal inference engaged the left-hemisphere. for most problem-solving tasks, modular views Another study assessed the brain systems for proces- receive support. Differences in task conditions modu- sing evidence that was either consistent or inconsistent late the active areas of the prefrontal cortex, however, with an individual’s existing causal beliefs. Individuals providing support for the task specificity hypothesis. received evidence on the effectiveness of drugs designed For example, well-structured versus ill-structured to relieve depressive symptoms. Two factors were problems rely on left versus right prefrontal regions, manipulated: the plausibility of the theory that respectively. This pattern is also consistent with the- explained the drug’s action, and the consistency ories that assume two different systems, such as dual- between theory and data. When individuals reasoned process and dual-code theories, support reasoning. with evidence that was consistent with existing causal
Finally, distributed theories receive support, given beliefs, a network of brain regions widely associated that multiple systems typically underlie problem solv- with learning and memory was engaged, including ing. For example, problem solving requires both the caudate and the parahippocampal gyrus. In con- planning and execution, and also both reasoning and trast, when they reasoned with inconsistent evidence, motivation, with different neural systems supporting a different pattern of activation occurred that is each component process. widely associated with error detection and conflict resolution, including the anterior cingulate cortex (BA Causal Reasoning 24/32), posterior cingulate, and precuneus (BA 7). The Several studies have evaluated the neural systems that researchers concluded that people’s beliefs and expec- support the perception of mechanical causation in the tations act as a filter during evidence evaluation. When
Encyclopedia of Neuroscience (2009), vol. 8, pp. 35-43
Author's personal copy
Reasoning and Problem Solving: Models 41 evidence is consistent with existing causal beliefs, the (BA 10), reflecting the internal generation of relations neural systems underlying those beliefs implement required to form complex analogies. Interestingly, this causal reasoning. When evidence is inconsistent with activation only occurred for 2-relational problems, existing beliefs, a different neural system detects this not for 1-relational problems, suggesting that fronto- inconsistency and triggers the construction of a novel polar cortex is important for processing complex rela- causal explanation. tional structures. Second, activations also occurred in The observed findings demonstrate that causal right DLPFC (BA 46), which reflected greater manip- reasoning is supported by a broadly distributed neural ulation of externally presented information in more system that is highly sensitive to the causal reasoning complex problems. Other studies using the Raven’s task (i.e., causal perception versus causal inference) Progressive Matrices task have found similar results, and the consistency of causal evidence with existing and have also reported bilateral posterior parietal beliefs. These findings provide support for theories activations (BA 7). of reasoning that are (1) based on distributed rather In another study, individuals received a source than localized representations (e.g., dual-process, dual- picture of colored geometric shapes, followed by a code, embodied theories), (2) incorporate the role of target picture of colored geometric shapes. Pictures existing knowledge rather than operating on the basis that did not share similar geometric shapes but that of purely logical representations (e.g., dual-process, did share the same system of abstract visuospatial dual-code, embodied, mental-model theories), and relations were also presented. Individuals judged (3) advocate the task specificity hypothesis (e.g., whether each source-target pairing was analogous dual-process, dual-code, embodied theories). (analogy condition) or identical (literal condition). Analogical reasoning (analogy minus literal) recruited
the dorsomedial frontal cortex (BA 8) and left-hemi- Analogical Reasoning sphere regions, including frontopolar (BA 10), infe- Several studies have found that analogical reasoning rior frontal (BA 44, BA 45, BA 46, and BA 47), engages frontopolar cortex. Furthermore, different and middle frontal (BA 6) cortices, and also inferior components of analogical reasoning appear to differ- parietal cortex (BA 40). These findings suggest that entially engage frontopolar versus dorsolateral pre- analogical reasoning is mediated by a predominantly frontal areas. Whereas dorsolateral areas are recruited left-hemisphere frontoparietal system. for processing externally generated information (e.g., Another study systematically evaluated the compo- the monitoring and manipulation of presented facts), nent processes of analogical reasoning. Specifically, frontopolar areas are recruited additionally for the this study assessed the neural systems that underlie evaluation and manipulation of internally generated (1) the storage of abstract relations in working information. memory and (2) the process of integrating abstract One study assessed the neural systems that support relations to form analogies. These researchers also analogical reasoning in the Raven’s Progressive Matri- found that analogical reasoning activated a left ces task. Study participants received a 3 3 matrix of frontoparietal system, with some regions of this cir- figures with the bottom right figure missing, and had cuit mediating working memory processes and others to infer the missing figure by selecting one of four mediating abstract relational integration. In particu- possible alternatives. Participants received three types lar, left frontopolar regions (BA 9/10) were again of problems that differed in their degree of relational central to the processing of relations that underlie complexity (0-relational, 1-relational, 2-relational). analogical reasoning.
The 0-relational problems involved no relation of In summary, the observed findings for analogical change and thus required no relational processing. reasoning are consistent with theories that advocate The 1-relational problems involved one relation of distributed rather than localized representations (e.g., change in either the horizontal or vertical dimension, dual-process, dual-code, embodied theories). As we and thus required relational reasoning. Finally, the saw, distributed frontal and posterior systems appear 2-relational problems involved two relations of to play a wide variety of roles as representations are change, in both the horizontal and vertical directions, retrieved, stored, and integrated. Findings on analog- and thus required even more relational reasoning. ical reasoning further support the cognitive demand A region-of-interest analysis was performed to assess and task specificity hypotheses (e.g., dual-process, the role of prefrontal cortex in processing multiple dual-code, embodied theories). Different kinds of relations simultaneously (i.e., relational integration). information recruited different prefrontal areas (e.g., This analysis produced two main findings. First, acti- external vs. internal information), and the harder vation occurred in frontopolar prefrontal cortex the reasoning, the more areas recruited (e.g., 1-relation
Encyclopedia of Neuroscience (2009), vol. 8, pp. 35-43
Author's personal copy
42 Reasoning and Problem Solving: Models vs. 2-relation problems). The general importance of In general, reasoning tends to recruit broadly frontopolar cortex for generating relations internally distributed and diverse neural systems. The difficulty suggests that this area is especially important for of establishing specific neural systems for a given type processing complex analogies. of reasoning (e.g., deduction) strikes us as one major
challenge for future research. Does a particular type of Summary, Conclusions, and Future reasoning consistently activate a specific neural circuit across wide variation in tasks and materials? If so, what Directions is this circuit? Another major challenge is that the mul- The findings discussed in this article support the tifaceted nature of the neurobiological evidence and following conclusions about the neural bases of cognitive theories results in a many-to-many mapping: neural regions often serve multiple cognitive functions reasoning. First, deductive reasoning does not appear to recruit a unitary neural system but instead engages that can be mapped onto multiple cognitive theories. different brain regions based on the particular Sorting out the functional roles of particular brain areas reasoning task and materials employed. Second, and their roles in psychological theories of reasoning inductive inferences drawn from familiar categorical should be another major goal in this research area. syllogisms and conditional statements engage a left- Many other challenges also await future research hemisphere language and knowledge network imple- in this area. For example, do the results from the mented in frontal and temporal regions. Third, laboratory tasks reviewed here generalize to everyday problem solving generally recruits prefrontal regions, reasoning tasks? Rather than occurring in a vacuum, including bilateral DLPFC and MVPFC. Fourth, causal everyday reasoning often occurs in social situations and is associated with emotional affect. What are the reasoning does not engage a single neural system but instead recruits different systems based on the causal neural bases of reasoning under these conditions? reasoning task and the consistency of causal evidence How do social and emotional processes modulate with background beliefs. Finally, analogical inference reasoning? Future research should also address the selectively recruits frontal and parietal regions, with important distinction between reasoning to a conclu- frontopolar cortex becoming increasingly important sion versus recognizing a conclusion. Because many as task complexity increases. real-world situations require that people generate
The neuroscience evidence reviewed helps evaluate valid conclusions (not just recognize them), future current psychological theories. First, reasoning typi- neuroscience research should assess the neural bases cally recruits broadly distributed neural systems, of this process. Future research should also continue to address the role of task difficulty in reasoning. providing evidence inconsistent with the relatively localized predictions of modularity theory. Possible Does task difficulty result in the recruitment of more exceptions include the importance of certain frontal brain regions, or are the same regions activated regions for problem solving (left DLPFC) and analogi- more intensely? Future research should also address cal reasoning (frontopolar cortex). Second, the results the effect of learning on the neural systems that demonstrate that reasoning does not solely recruit left- underlie reasoning. Do novices and experts engage hemisphere regions for language and rule-based opera- similar neural systems during reasoning? tions. Instead, reasoning often engages bilateral and Finally, researchers should develop psychological the- posterior regions beyond those implicated by amodal ories that motivate fine-grained neurobiological predic- and mental logic theories of reasoning. Third, the pre- tions, and should design experiments that distinguish between theories, rather than simply attempting to con- dominantly right-hemisphere system predicted by the mental-models theory across both deduction and firm one. Although much progress has been made in induction is inconsistent with many patterns of left- framing the issues and establishing preliminary evi- hemisphere and bilateral activation observed across dence, neuroscience research on reasoning is still in its studies. Fourth, dual-process, dual-code, and embod- infancy. We anticipate major advances in coming years. ied theories are generally consistent with the reviewed findings. These theories receive support because they Acknowledgments predict the presence of (1) broadly distributed neural systems and (2) multiple reasoning systems (e.g., asso- This work was supported by National Science Foun- ciative vs. rule, language vs. knowledge). These the- dation Grants DGE-0536941 and DGE-0231900 to A K Barbey, and by National Science Foundation ories also receive support from the many studies that exhibit effects of task specificity and cognitive demand. Grant BCS-0212134 and Defense Advanced Research As task conditions change, so do the neural systems Programs Agency Contract FA8650–05-C-7256 to that represent and process the relevant information. L W Barsalou.
Encyclopedia of Neuroscience (2009), vol. 8, pp. 35-43
Author's personal copy
Reasoning and Problem Solving: Models 43
See also: Decision-Making and Vision; Executive Function Goel V (2005) Cognitive neuroscience of deductive reasoning. In: and Higher-Order Cognition: Assessment in Animals; Holyoak K and Morrison R (eds.) Cambridge Handbook of Thinking and Reasoning, pp. 475–492. Cambridge Univer- Games in Monkeys: Neurophysiology and Motor Deci- sity Press: New York. sion-Making; Memory Representation; Prefrontal Cortex: Structure and Anatomy; Prefrontal Cortex; Referentiality Goel V, Buchel C, Frith C, et al. (2000) Dissociation of mechanisms underlying syllogistic reasoning. NeuroImage 12: 504–514. and Concepts in Animal Cognition; Reward Decision- Goel V and Dolan RJ (2003) Explaining modulation of reasoning Making. by belief. Cognition 87: B11–B22. Green AE, Fugelsang JA, Kraemer DJM, et al. (2006) Frontopolar cortex mediates abstract integration in analogy. Brain Research Further Reading 1096: 125–137. Holyoak KJ (2005) Analogy. In: Holyoak KJ and Morrison RG Barbey AK and Sloman SA (in press) Base-rate respect: From (eds.) Cambridge Handbook of Thinking and Reasoning, pp. ecological rationality to dual processes. Behavioral and Brain 117–142. Cambridge, UK: Cambridge University Press. Sciences. Johnson-Laird PN (1983) Mental Models: Towards a Cognitive
Barsalou LW (1999) Perceptual symbol systems. Behavioral and Science of Language, Inference, and Consciousness. Cambridge,
Brain Sciences 22: 577–660. UK: Cambridge University Press. Braine MDS (1990) The ‘natural logic’ approach to reasoning. In: Novick LR and Bassok M (2005) Problem solving. In: Holyoak KJ Overton WF (ed.) Reasoning, Necessity, and Logic: Develop- and Morrison RG (eds.) Cambridge Handbook of Thinking and mental Perspectives , pp. 135–158. Mahwah, NJ: Erlbaum Press. Reasoning, pp. 321–349. New York: Cambridge University Christoff K and Gabrieli JDE (2000) The frontopolar cortex and Press. human cognition: Evidence for a rostrocaudal hierarchical orga- Osherson D, Perani D, Cappa S, et al. (1998) Distinct brain loci in nization within the human prefrontal cortex. Psychobiology 28: deductive versus probabilistic reasoning. Neuropsychologia 36: 168–186. 369–376. Christoff K, Prabhakaran V, Dorfman J, et al. (2001) Rostrolateral Over DE (2003) From massive modularity to meta-representation: prefrontal cortex involvement in relational integration during The evolution of higher cognition. In: Over DE (ed.) Evolution reasoning. NeuroImage 14: 1136–1149. and the Psychology of Thinking: The Debate, pp. 121–144.
Colvin MK, Dunbar K, and Grafman J (2001) The effects of frontal New York: Psychology Press.
lobe lesions on goal achievement in the water jug task. Journal Paivio A (1986) Mental Representations. New York: Oxford of Cognitive Neuroscience 13: 1129–1147. University Press. Flombaum JI, Santos LR, and Hauser MD (2002) Neuroecology Parsons LM and Osherson DN (2001) New evidence for distinct and psychological modularity. Trends in Cognitive Sciences 6: right and left brain systems for deductive versus probabilistic 106–108. reasoning. Cerebral Cortex 11: 954–965. Fonlupt P (2003) Perception and judgment of physical causality Roger ME, Fugelsang JA, Dunbar KN, et al. (2005) Dissociating involve different brain structures. Cognitive Brain Research 17: processes supporting causal perception and causal inference in 248–254. the brain. Neuropsychology 19: 591–602. Fugelsang J and Dunbar K (2004) A cognitive neuroscience frame- Wharton CM, Grafman J, Flitman SS, et al. (2000) Toward neuro- work for understanding causal reasoning and the law. Philo- anatomical models of analogy: A positron emission tomography sophical Transactions of the Royal Society of London. Series B study of analogical mapping. Cognitive Psychology 40:
359: 1749–1754. 173–197.
Goel V (2002) Planning: Neural and psychological. In: Nadel L Whitaker H, Savary F, Markovits H, et al. (1991) Inference deficits (ed.) Encyclopedia of Cognitive Science, vol. 3, pp. 697–703. after brain damage. Paper presented at the International New York: Macmillan. Neuropsychological Meeting, San Antonio, 1991.
Encyclopedia of Neuroscience (2009), vol. 8, pp. 35-43