Reasoning and Thinking, Neural Basis of

875

Reasoning and Thinking, Neural Basis of

Intermediate article

Jordan Grafman, National Institute of Neurological Disorders and Stroke, Bethesda, Maryland, USA Vinod Goel, York University, Toronto, Ontario, Canada
CONTENTS
Introduction Reasoning and the brain Assessment of the Studies Future directions

Reasoning is the cognitive activity of drawing inferences from given information. All arguments in reasoning involve the claim that one or more propositions (the premises) provide some grounds for accepting another proposition (the conclusion).

INTRODUCTION
Reasoning is the cognitive activity of drawing inferences from given information. All arguments in reasoning involve the claim that one or more propositions (the premises) provide some grounds for accepting another proposition (the conclusion). Consider the following example. George returns from the grocery store accompanied by his young daughter and tells her that kiwi are perishable fruit, and that perishable fruit are placed in the refrigerator. Even though she has never seen kiwi before, she promptly places the kiwi in the refrigerator. How did she know that she should put them in the refrigerator? Furthermore, after seeing the purchased kiwi, she formed the belief that all kiwi have a brown, furry skin. Notice that she was not explicitly told that kiwi should be placed in the refrigerator, and she saw only the three purchased kiwi, yet we are not surprised by her actions. Her behavior is not a mystery. It is just an example of the reasoning brain at work. The former inference is an example of deductive reasoning, while the latter is an example of inductive reasoning. A key feature of deductive arguments is that conclusions are contained within the premises and are independent of the content of the sentences. They can be evaluated for validity, a relationship between premises and conclusion involving the claim that the premises provide absolute grounds for accepting the conclusion. The reader is referred to Johnson-Laird's article on deduction for

background information on the cognitive theories of deductive reasoning. In inductive reasoning the conclusion reaches beyond the original set of premises, allowing for the possibility of creating new knowledge. Induction has long been a concern of philosophers, and it has been empirically studied by cognitive scientists for the past 25 years. Cognitive/computational models of induction typically view it as a form of hypothesis generation and selection, where one must search a large database and determine which items of information are relevant and how they are to be mapped on to the present situation. The determination of hypothesis relevance is the crucial component of induction. In the above example, the inference that `all kiwi have a brown, furry skin', which was drawn after seeing three kiwi, seems to be plausible. However, each of the three kiwi also had a small sticker attached stating the name of the store at which they were purchased (`IGA'). Yet from seeing this sticker on three kiwi fruits we do not conclude that all kiwi have an IGA sticker on them. The former inference is plausible, but the latter is not. The logic of the two inferences is identical. However, we recognize that the property of having `brown furry skin' is a relevant property for generalization across members of a species, but that the property of having an `IGA' sticker is a matter of individual accident. The puzzle of induction is, to a large extent, the question of how we make these judgments of relevance. Given that almost nothing is known about the cognitive, computational, and neural basis of inductive inference, we shall restrict ourselves to deductive reasoning in this article.

REASONING AND THE BRAIN

There are two main ways in which cognitive neuroscientists investigate the neural mechanisms that

876

Reasoning and Thinking, Neural Basis of

underlie reasoning processes. One approach is to study patients who have damage to specific brain regions and to observe whether they have difficulty in solving certain types of reasoning problems and, if so, which specific cognitive impairments can explain their reasoning deficits. The second approach is to study normal subjects and patients using functional neuroimaging techniques such as positron emission tomography (PET) and functional magnetic resonance imaging (fMRI), which allow the investigator to view which brain areas become selectively activated by specific reasoning processes and tasks.

Patient studies
Patient studies of logical reasoning have been few and infrequent. Gazzaniga and colleagues have administered simple reasoning tasks to split-brain patients and concluded that reasoning is a lefthemisphere phenomenon (Gazzaniga and Smylie, 1984). For example, they report that the left hemisphere will readily infer `boiling water' when presented with `water' and `pan', while the right hemisphere seems to be incapable of such inferences. Gazzaniga goes on to postulate a `left-brain interpreter' that is, a mechanism which continuously elaborates and interprets information presented to it, and which readily draws inferences. Caramazza et al. (1976) administered to braindamaged patients two-term problems such as the following: `Mike is taller than George. Who is taller?' They reported that left-hemisphere patients showed impairment for all forms of the problem, but right-hemisphere patients were impaired only when the form of the question was incongruent with the premise (e.g. `Who is shorter?'). Read (1981) tested temporal-lobectomy patients by presenting them with three-term relational problems with semantic content (e.g. `George is taller than Mary. Mary is taller than Carol. Who is tallest?'). The subjects were told that using a mental imagery strategy would help them to solve these problems. Read reported that the performance of left-temporal-lobectomy patients was more impaired than that of right-temporal-lobectomy patients. Left-temporal-lobectomy patients reported less use of mental imagery to solve the problems, and they also did not show deficits on standard neurological tests of verbal comprehension. Read concluded that `The form of imagery utilized in solving deductiverecent problems is based upon verbal symbolic affirmation, and as such is mediated by the left hemisphere' (Read, 1981).

Golding (1981) evaluated responses to the Wason Four-Card Selection Task (Wason, 1966), which is perhaps the most widely used task for exploring the role of content in reasoning. In this task subjects are shown four cards (Figure 1). They can see what is on one side of each card, but not what is on the other side. They are given a rule of the form if p then q (`If a card has a vowel on one side, it has an even number on the other side') and asked which cards they would turn over in order to verify the rule. The visible values on the cards correspond to the p, not-p, q, and not-q cases of the rule. According to standard propositional logic, the correct choices are p (to verify that q is on the other side) and not-q (to verify that p is not on the other side). Although no control subjects and only one left-hemispheredamaged subject selected the p card and not-q card (the logically correct answer), 50% of the righthemisphere-damaged subjects chose both cards. Because control subjects tended to select those cards that matched the items described in specific test sentences, Golding proposed that the perceptual aspects of the task interfered with control subjects' verbal reasoning (e.g. when given the sentence `Whenever there is a circle on one half of the card, there is yellow on the other half of the card,' control subjects selected the circle card or the circle and yellow cards). According to Golding, right-hemisphere-damaged patients with impaired visual processing showed superior verbal reasoning skills owing to a lack of visual perceptual interference. Adolphs et al. (1996) administered a Wason selection task, using both familiar and unfamiliar stories, to patients with dorsolateral frontal lesions or ventromedial frontal lesions, and to normal controls. When given an arbitrary rule (as in Figure 1), typically fewer than 15% of normal subjects will turn over both the p card (A) and the not-q card (3). The introduction of meaningful content in a rule (e.g., `If anyone is drinking beer, then that

not-p

not-q

Figure 1. The Wason Four-Card Selection Task. Each card has a letter on one side and a number on the other side. The task is to determine which cards need to be turned over in order to verify the rule `If a card has a vowel on one side, it has an even number on the other side.'

Reasoning and Thinking, Neural Basis of

877

person must be over 19 years old') greatly facilitates performance. They reported that subjects in all three groups performed equally poorly when given unfamiliar stories (arbitrary rules). However, both dorsolateral patients and normal subjects chose the p and not-q cards (the logically correct selections) when the story involved familiar material. In contrast, the ventromedial patients, who had damage to the medial orbitofrontal cortex (five out of six cases had bilateral damage), did not show this facilitatory effect with familiar material. The authors concluded that in general people may reason by analogy, retrieving past experiences (including the emotion experienced) when confronted with a familiar situation. Ventromedial prefrontal patients may fail to retrieve or appropriately use past experiences when reasoning. Waltz et al. (1999) found that compared with patients with damage to the temporal cortex, patients with frontal-lobe dementia were dramatically impaired in their ability to make inferences with regard to the integration of multiple relational representations. If problems varied on only one dimension, there were no between-group differences (including controls). If a problem only required judgments about canonical ordering, then patients with frontal-lobe dementia could solve the problem, but if they had to reorder the relationships to make the judgment, they failed to solve it. In summary, studies of patients with focal lesions indicate that the left hemisphere helps to interpret stimuli as linguistic elements for further processing and inferential processes. Damage to the left hemisphere interferes with deductive reasoning. The right hemisphere may become more involved when distant relationships are being processed, and thus right-hemisphere lesions affect the processing of apparently incongruent stimuli. It is likely that left-hemisphere lesions lead to an inability to develop certain forms of mental imagery that are used when developing mental models of relationships. If a subject needs to use past experience to solve a problem, then the ventromedial prefrontal cortex may be important.

Neuroimaging studies
Goel et al. (1997) conducted an imaging study of deductive reasoning. Ten normal volunteers performed reasoning tasks while their regional cerebral blood flow pattern was recorded using the [15O] H2O positron emission tomography (PET) blood flow technique. Subjects were presented with arguments such as `All apples are red; all red fruit are sweet; therefore all apples are sweet', and

were then asked to make judgments about the validity of the conclusion. The deduction condition (versus a semantic baseline) resulted in activation of the left inferior frontal gyrus (Brodmann areas 45 and 47) and a region of the left superior occipital gyrus (Brodmann area 19). These brain areas are known to mediate aspects of language processing, and as such the results provide support for sentential theories of reasoning, at the expense of spatial models. However, given that many people have the phenomenological experience of mapping syllogisms on to Venn-like diagrams, it is difficult to accept that the visuospatial system plays no role in deductive reasoning. Perhaps the absence of performance-associated brain activation in regions that are known to be involved in spatial processing (i.e. right hemisphere and parietal regions) was just a function of the fact that the deductive reasoning items which were used (a combination of categorical syllogisms, implications, disjunctions, and conjunctions) did not involve overt spatial relationships (e.g. John is standing behind Mary). Perhaps if the arguments explicitly required spatial encoding, then we might find right-hemisphere and parietal activation. Goel et al. (1998) conducted another imaging experiment, this time using explicitly spatial arguments. Again they found unilateral (lefthemisphere) activation confined to the dorsolateral prefrontal cortex, the anterior cingulate, and the middle and superior temporal lobes. They did not find any significant activation in cortical regions known to be associated with spatial encoding and spatial working memory. These results were consistent with the first study (and a number of other patient studies), and demonstrated a lack of involvement of classic spatial encoding regions in reasoning. However, the results of another study by Goel et al. (2000) may go some way towards resolving these counterintuitive results. In an fMRI study of deductive reasoning, using sentences with semantic content (e.g. `All apples are red; all red fruit are sweet; therefore all apples are sweet') and without semantic content (e.g. `All A are B; all B are C; therefore all A are C), they found evidence for the engagement of both linguistic and spatial systems, but under circumstances not predicted by cognitive theories. During content-based reasoning, a lefthemisphere temporal system was recruited. By contrast, a formally identical reasoning task that lacked semantic content activated a bilateral parietal system. This dissociation is remarkable because the logically relevant information is identical in

878

Reasoning and Thinking, Neural Basis of

both conditions. Based on brain activation profiles, the two systems shared common neural components in bilateral basal ganglion nuclei, right cerebellum, bilateral fusiform gyri, and left prefrontal cortex. Goel et al. concluded that syllogistic reasoning is implemented in two distinct systems whose engagement is primarily a function of the presence or absence of semantic content. Furthermore, when a logical argument results in a belief logic conflict, the nature of the reasoning process is changed by recruitment of the right prefrontal cortex. The involvement of a parietal visualspatial system in the abstract syllogism condition raises the question of whether argument forms involving three-term relational items (e.g. `The apples are in the barrel; the barrel is in the barn; therefore the apples are in the barn' and `Apples are more expensive than pears; pears are more expensive than oranges; therefore apples are more expensive than oranges') are sufficient to engage the parietal system. Goel and Dolan (2001) addressed this question in an event-related fMRI study of three-term relational reasoning, using sentences with concrete content (e.g. `The apples are in the barrel; the barrel is in the barn; therefore the apples are in the barn') and abstract content (e.g. `A are in B; B is in C; therefore A is in C'). They reported that both concrete and abstract three-term relational arguments activate a similar bilateral occipitalparietalfrontal network. However, the abstract reasoning condition engendered greater parietal activation than the concrete reasoning condition. It appears that arguments involving relationships that can be easily mapped on to explicit spatial relationships engage a visuospatial system, irrespective of concrete or abstract content. The only other imaging study of deductive reasoning to date is that by Osherson et al. (1998), who performed a [15O] H2O PET study to compare deductive reasoning with probabilistic reasoning and language-comprehension tasks involving the same stimuli. They reported that deductive reasoning compared with probabilistic reasoning activated occipital and parietal regions, with a right-hemispheric prevalence. The comparison most relevant to us, namely deductive reasoning (using arguments such as `None of the bakers play chess; some of the chess players listen to opera; some of the opera listeners are not bakers') versus language comprehension, resulted in activation in the left dorsal frontal gyrus (BA6), left cuneus (BA18), thalamus, caudate, and cerebellum. In summary, the functional neuroimaging results demonstrate that the left hemisphere, and in

particular the left prefrontal cortex, are very important for deductive reasoning. The more anterior the activation, the more likely it is that there is a semantic content requirement for reasoning to solve the problem.

ASSESSMENT OF THE STUDIES

The neuroimaging and patient studies to date support some form of dual-mechanism theory of deductive reasoning. The presence of semantic content engages the language system in the reasoning process. The absence of semantic content engages the visuospatial system in the identical reasoning task. At an anatomical level, these data predict that the left hemisphere is necessary and often sufficient for logical reasoning, whereas the right hemisphere is sometimes necessary but never sufficient. There is considerable consistency between the neuroimaging findings and the lesion data reported above. As described above, both brain-lesion patients and functional neuroimaging studies suggest that each hemisphere and specific brain regions may contribute different cognitive processes to human reasoning. One interpretation of the results of these studies is that the right hemisphere may be involved in formal reasoning independent of the content of the task. Therefore the right hemisphere would be more adept at abstract reasoning. In contrast, the left hemisphere may use past experiences (factual, emotional), and would thus be more adept when reasoning involves familiar scenarios. The ventromedial prefrontal neural network would play a role when the behavioral relevance of possible responses could aid selection of the appropriate action. The left hemisphere may also store in memory emotional states associated with personally experienced events. When an individual encounters a familiar scenario, representations of related past emotional experiences are retrieved by the left hemisphere and incorporated into the reasoning process. In the absence of or failure to activate such representations, the right hemisphere is engaged in the reasoning process. Asymmetrical advantages of processing based on receptive-field size offer another explanation for hemispheric differences in reasoning skills. Beeman (1993) provided evidence that large semantic receptive fields account for the right hemisphere's role in understanding discourse and metaphor. A similar explanation may underlie hemispheric differences in reasoning. Large receptive fields in the right hemisphere would permit individuals to

Reasoning and Thinking, Neural Basis of

879

activate all possible relationships, local and distant, between the items in the problem to be solved. Overlap between relational features from the activation of multiple relationships would give the right hemisphere an adaptive advantage over the left hemisphere for reasoning involving unfamiliar situations. By comparison, the left hemisphere's fine coding would allow individuals to focus on the main feature or event and the local relationships between the items in the problem. As a result, the left hemisphere would have an adaptive advantage over the right hemisphere in logical reasoning involving familiar content. One clear conclusion to be drawn from the cognitive neuroscience investigation of reasoning is that there is no reasoning module, and that several different cognitive-processing components are required to reason, depending on the type of task with which the subject is faced and the particular strategies with which the subject prefers to reason.

FUTURE DIRECTIONS
The cognitive neuroscience investigation of reasoning is still in its infancy. It is clear from this article that we are only crudely able to map reasoning processes primarily to the left hemisphere and to the prefrontal cortex. More posterior cortices may contribute to reasoning when visuospatial representations are used by subjects to think about the elements of the reasoning task they are performing. The semantic content of the elements of the problem will require temporal-lobe processing, and more personal or familiar content will utilize the functions of the prefrontal cortex. The latter does not fully mature until the end of the teenage years in humans, and it is of interest to note that reasoning ability matures over the same time frame. Reasoning does not stand alone as a cognitive process. Humans frequently reason during problem-solving tasks, when forming metaphors, and when developing a plan to achieve a goal. People also reason in social and non-social situations, and it may be that overlapping but different cognitive systems and neural regions are used for each type of reasoning. Future cognitive neuroscience studies that incorporate these factors into the design of reasoning tasks will enrich our understanding of the neural basis of reasoning processes. References
Adolphs R, Tranel D, Bechara A, Damasio H and Damasio AR (1996) Neuropsychological approaches to

reasoning and decision-making. In: Damasio AR (ed.) Neurobiology of Decision-Making, pp. 157180. Berlin, Germany: Springer-Verlag. Beeman M (1993) Semantic processing in the right hemisphere may contribute to drawing inferences from discourse. Brain and Language 44: 80120. Caramazza A, Gordon J, Zurif EB and DeLuca D (1976) Right-hemispheric damage and verbal problem-solving behavior. Brain and Language 3: 4146. Gazzaniga MS and Smylie CS (1984) Dissociation of language and cognition. Brain 107: 145153. Goel V and Dolan RJ (2001) Functional neuroanatomy of three-term relational reasoning. Neuropsychologia 39: 901909. Goel V, Gold B, Kapur S and Houle S (1997) The seats of reason: a localization study of deductive and inductive reasoning using PET (O15) blood flow technique. Neuroreport 8: 13051310. Goel V, Gold B, Kapur S and Houle S (1998) Neuroanatomical correlates of human reasoning. Journal of Cognitive Neuroscience 10: 293302. Goel V, Buchel C, Frith C and Dolan RJ (2000) Dissociation of mechanisms underlying syllogistic reasoning. Neuroimage 12: 504514. Golding E (1981) The effect of unilateral brain lesion on reasoning. Cortex 17: 3140. Osherson D, Perani D, Cappa S et al. (1998) Distinct brain loci in deductive versus probabilistic reasoning. Neuropsychologia 36: 369376. Read DE (1981) Solving deductive-reasoning problems after unilateral temporal lobectomy. Brain and Language 12: 116127. Waltz JA, Knowlton BJ, Holyoak KJ et al. (1999) A system for relational reasoning in human prefrontal cortex. Psychological Science 10: 199125. Wason PC (1966) Reasoning. In: Foss B (ed.) New Horizons in Psychology, pp. 135151. Harmondsworth, UK: Penguin.

Further Reading
Evans JSBT, Newstead SE and Byrne RMJ (1993) Human Reasoning: the Psychology of Deduction. Hillsdale, NJ: Lawrence Erlbaum. Garnham A and Oakhill J (1994) Thinking and Reasoning. Oxford, UK: Blackwell Science. Goel V and Dolan RJ (2000) Anatomical segregation of component processes in an inductive inference task. Journal of Cognitive Neuroscience 12: 110. Goodman N (1955) Fact, Fiction and Forecast. Cambridge, MA: Harvard University Press. Henle M (1962) On the relation between logic and thinking. Psychological Review 69: 366378. Johnson-Laird PN (1994) Mental models, deductive reasoning, and the brain. In: Gazzaniga MS (ed.) The Cognitive Neurosciences, pp. 9991008. Cambridge, MA: MIT Press. McCarthy RA and Warrington EK (1990) Cognitive Neuropsychology: a Clinical Introduction. New York, NY: Academic Press.

880

Reasoning and Thinking, Neural Basis of Rips LJ (1994) The Psychology of Proof: Deductive Reasoning in Human Thinking. Cambridge, MA: MIT Press. Sloman SA (1996) The empirical case for two systems of reasoning. Psychological Bulletin 119: 322.

Newell A (1980) Reasoning, problem solving and decision processes: the problem space as a fundamental category. In: Nickerson RS (ed.) Attention and Performance VIII, pp. 693718. Hillsdale, NJ: Lawrence Erlbaum. Rescher N (1980) Induction: an Essay on the Justification of Inductive Reasoning. Pittsburgh, CA: University of Pittsburgh Press.

Reasoning under Uncertainty

CONTENTS

Intermediate article

Francisco J Dez, UNED, Madrid, Spain Marek J Druzdzel, University of Pittsburgh, Pittsburgh, Pennsylvania, USA
Introduction Naive Bayes MYCIN's certainty factors PROSPECTOR's Bayesian model DempsterShafer theory Bayesian networks Influence diagrams Fuzzy logic and fuzzy sets Rough sets Non-monotonic logics Conclusion

Most artificial intelligence applications, especially expert systems, have to reason and make decisions based on uncertain data and uncertain models. For this reason, several methods have been proposed for reasoning with different kinds of uncertainty.

This article reviews some of the uncertain reasoning methods that have been proposed in the field of artificial intelligence.

INTRODUCTION
We often have to make decisions based on uncertain knowledge, not only in our private lives (which job to take, which house to buy, where to invest our money) but also in professional activities, such as medicine, economics, politics, engineering, and education. Therefore, any reasoning method that tries to replicate human reasoning must be able to draw conclusions from uncertain models and uncertain data. Models may be uncertain because of indeterminism in the real world or because of our lack of knowledge. Furthermore, data may be incomplete (pieces of information may be not available in a diagnostic case), ambiguous (a pronoun in a sentence may refer to different subjects), erroneous (patients may lie to their doctors, or sensors may be faulty), or imprecise (because of the limited precision of measuring devices, subjective estimations, or natural language).

NAIVE BAYES
The oldest method applied in uncertain reasoning is probability theory. Probabilistic reasoning concentrates basically on computing the posterior probability of the variables of interest given the available evidence. In medicine, for example, the evidence consists of symptoms, signs, clinical history, and laboratory tests. A diagnostician may be interested in the probability that a patient suffers from a certain disease. In mineral prospecting, we may wish to know the posterior probability of the presence of a certain deposit given a set of geological findings. In computer vision, we might be interested in the probability that a certain object is present in an image given observation of certain shapes or shadows. The probability of the diagnoses given the available evidence can be computed by the generalization of Bayes' theorem to several variables. However, the direct application of this method would need a prohibitive number of parameters