Report

A Neural Basis for General Intelligence

See allHide authors and affiliations

Science  21 Jul 2000:
Vol. 289, Issue 5478, pp. 457-460
DOI: 10.1126/science.289.5478.457

Abstract

Universal positive correlations between different cognitive tests motivate the concept of “general intelligence” or Spearman'sg. Here the neural basis for g is investigated by means of positron emission tomography. Spatial, verbal, and perceptuo-motor tasks with high-g involvement are compared with matched low-g control tasks. In contrast to the common view that g reflects a broad sample of major cognitive functions, high-g tasks do not show diffuse recruitment of multiple brain regions. Instead they are associated with selective recruitment of lateral frontal cortex in one or both hemispheres. Despite very different task content in the three high-g–low-g contrasts, lateral frontal recruitment is markedly similar in each case. Many previous experiments have shown these same frontal regions to be recruited by a broad range of different cognitive demands. The results suggest that “general intelligence” derives from a specific frontal system important in the control of diverse forms of behavior.

As discovered by Spearman (1) early in the last century, measures of performance or success in diverse cognitive tests show a pattern of almost universal positive correlation: To some extent at least, the same people tend to perform well in very different tasks. To explain this result, Spearman put forward the hypothesis of a general or g factor making some contribution to success in diverse forms of cognitive activity. People with high g scores will be those usually performing well, leading to the interpretation of g as “general intelligence.” Factor analysis can be used to show which tasks are most correlated with g and are thus the best general intelligence measures; often, these turn out to be tests of novel problem solving such as Raven's Progressive Matrices (2). An alternative hypothesis, originally proposed by Thomson (3), has also received detailed consideration. According to this hypothesis, any task receives contributions from a large set of component factors or information-processing functions. Universal positive correlation arises not for any common reason, but simply because any two tasks are likely to share at least some components. The “g factor” measured by standard intelligence tests is now interpreted as the average efficiency of the total set of cognitive functions (4); as Thomson showed, tasks with high apparent g correlations will be those sampling the total set of cognitive functions most broadly. The indeterminacy of factor analysis has made it impossible to distinguish these alternative hypotheses from correlational data, and after almost a century of debate, both are still vigorously defended (5). Here we use positron emission tomography (PET) to investigate the neural basis forg and the light this casts on Spearman's and Thomson's interpretations.

One possibility—more closely allied to the Spearman view—is that g may reflect some relatively confined set of neural functions, broadly contributing to success in diverse cognitive tests. In recent years, in particular, similarities have been noted between some effects of frontal lobe lesions and the normal behavior of people from the lower part of the g distribution, suggesting that frontal functions may be particularly central to g(6). Though frontal functions are not well understood—as reflected in rather general information-processing concepts such as executive control, strategy formation, or monitoring the contents of working memory (7)—certainly they are important in a wide diversity of behavior, as a major role in g would imply. According to this hypothesis, tasks with high gcorrelations should be characterized by specific recruitment of prefrontal cortex. The alternative—directly implied by the Thomson hypothesis—is that increasing g correlations should be associated with an increasingly diverse pattern of neural activation, reflecting increasingly broad sampling of all major cognitive functions.

Our method took advantage of the psychometric finding that tasks with very different surface content can share the property of highg correlation (8). Thus, converging investigations of several high-g tasks can be used to ask what property these tasks share at the level of neural activity. In our first test, we used problem-solving tasks based on spatial and verbal materials. In each case, we began by adapting standard psychometric tests whose g correlations were known to be high. Example problems from spatial and verbal tasks are shown in Fig. 1, A and B. In two large-Nbehavioral studies, we confirmed high g correlations in the adapted task versions (9) (Table 1). For each task, we developed a corresponding low-g control, based on similar materials but without the problem-solving element. Again, behavioral pretesting confirmed the lower gcorrelations of these newly developed control tasks. In all cases, tasks were designed to keep the participant continuously active, despite wide variations in the time taken to respond to individual test items. We then used PET to compare each high-gproblem-solving task with its corresponding low-g control (10).

Figure 1

Example test items for each task. Each item consisted of four display elements (drawings or letter sets), and the task was to identify the element that in some sense mismatched or differed from the others. In each task, participants completed as many items as possible in a fixed period of 4 (behavioral sessions) or 2 (PET sessions) min, after 0.5 min of practice. (A) Materials for the high-g spatial task were adapted with permission from a standard nonverbal test of g, Cattell's Culture Fair, Scale 2 Form A and Scale 3 (17). Display elements were four panels, each containing one or more shapes, symbols, or drawings. One panel differed in some respect from the others; extensive problem solving was required to identify this panel because the difference could concern any property, often abstract and/or complex. In the example shown the relevant property is symmetry; the mismatching panel is the third in the row. In the low-g spatial control task, in contrast, there was minimal problem solving. In each display, the four panels each contained a single geometrical shape, three of which were physically identical whereas the fourth differed in some visually obvious respect (shape, texture, size, orientation, or a combination of these). (B) Materials for the high-g verbal task were adapted with permission from a standard letter-based problem-solving task, Letter Sets from the ETS Kit of Factor-Referenced Tests (24). The high gloading of the original test was established by analysis of a large preexisting data set (25). Display elements were four sets of four letters each. One set differed in some respect from the others; again, the task required extensive problem solving because a variety of alphabetic and other rules could distinguish the mismatching letter set in any given test item. In the example the mismatching set is the third, whose letters are equally spaced in the alphabet. In the low-g verbal control task, the task was simply to find the one set in each display whose letters were not in strict alphabetical order. (C) Displays in the circles task were based with permission on two drawings taken from a single item of the Cattell Culture Fair (17). As illustrated, in one drawing the smallest circle was toward the center of the overall figure, whereas in the other it was toward the periphery. Again there were three identical panels (all small circles central, or all peripheral), and the task was to identify the single panel with the alternative arrangement. For all tasks, displays were presented on an Apple Macintosh monitor; horizontal display extent was approximately 12° for spatial tasks, 19° for verbal tasks. The position of the mismatching element was indicated by pressing the corresponding key on a four-choice keyboard, operated with middle and index fingers of the two hands. The screen cleared when a response was made, and a new test item was presented after a pause of 500 ms. In problem-solving tasks, participants were encouraged not to guess but to work on each problem until they were confident of their answer. These arrangements ensured that participants worked continuously throughout the period of each task, despite long solution times for problem-solving items but much shorter times for control items.

Table 1

Behavioral data for all tasks. Correlations with standard measures of g derive from one (spatial and circles tasks) or two (verbal tasks) behavioral studies conducted before the PET experiment (9). Correlations for verbal tasks are averages from the two studies calculated by Fisher'sz-transform. N denotes number of participants contributing to each correlation, excluding cases with missing data.

View this table:

Regions of significantly greater blood flow (P < 0.05, corrected for multiple comparisons) in high-g compared with low-g tasks are shown in Fig. 2, A (spatial comparison) and B (verbal comparison) (11). Corresponding peak activations appear inTable 2. In the spatial comparison, the strongest high-g activations occurred bilaterally in the lateral prefrontal cortex, and in a discrete region of the medial frontal gyrus/anterior cingulate. Elsewhere in the brain, activations were restricted to the posterior visual system, presumably reflecting more extensive visual analysis and/or the effects of eye movements, and to discrete regions of parietal and premotor cortex, recruited in a wide range of visuospatial tasks (12). These results resemble those previously seen in comparisons of Raven's Progressive Matrices with simple sensorimotor controls (13). For the verbal comparison, the only significant high-g activation occurred in the lateral frontal cortex of the left hemisphere, closely corresponding to the similar activation in the spatial comparison.

Figure 2

Significant activations for three contrasts, rendered onto canonical T1-weighted brain image of SPM99. (A) Spatial high-g minus spatial low-g(P < 0.05 corrected for multiple comparisons). (B) Verbal high-g minus verbal low-g(P < 0.05 corrected). (C) Circles minus spatial low-g (P < 0.001 uncorrected).

Table 2

Peak activations for each high-g minus low-g task contrast. Coordinates [x,y, z in space of Montreal Neurological Institute (MNI) template] and selection of foci are according to the conventions of SPM99. Brain regions (approximate Brodmann areas) are estimated from Talairach and Tournoux (23), after adjustment (www.mrc-cbu.cam.ac.uk/Imaging/mnispace.html) for differences between MNI and Talairach coordinates.

View this table:

Such results argue strongly against the possibility that high-g tasks are associated with diffuse neural recruitment, as predicted by broad sampling of the brain's major cognitive functions. Examination of the data at a less conservative significance threshold (P < 0.001 uncorrected) did not change this conclusion; in the spatial comparison, there was simply a strengthening of the major activation foci shown in Fig. 2, whereas in the verbal comparison, frontal activation was accompanied by weak occipital activations resembling those shown for the spatial contrast. Neither did global blood-flow measurements suggest greater diffuse activity in high-g tasks (14). Instead, the data strongly favor the hypothesis that lateral frontal functions are selectively recruited by high-g tasks.

Our behavioral development work suggested a converging test of this conclusion. There was considerable variation ing correlations among various candidates we investigated for the role of low-g spatial control task. Although the reasons for this variation are unknown, it allowed us to select a further task for inclusion in the PET experiment whose g correlation was high despite a simple physical-match format and no strong element of problem solving [Fig. 1C; for behavioral data see Table 1 and (15)]. The new task was compared with our standard low-g spatial control, to which it was identical except for the exact shapes used in the stimulus array. At the threshold ofP < 0.05 corrected for multiple comparisons, there were no frontal differences between these two tasks. At the less conservative threshold of P < 0.001 uncorrected, however, the higher g task was associated with lateral frontal activation in the right hemisphere (Fig. 2C and Table 2), closely corresponding to the similar activation in our main spatial contrast.

Evidently, a neural system associated with Spearman'sg should be recruited by many different forms of cognitive demand. In a recent analysis of imaging findings, indeed, we have shown that diverse forms of demand, including task novelty, response competition, working memory load, and perceptual difficulty, produce broadly similar lateral frontal activations covering a region closely similar to the frontal activations seen here (16). On the medial surface, all these demands are also associated with specific recruitment of the dorsal part of the anterior cingulate, close to the medial frontal activity seen here only for the spatial problem-solving task. If future work shows this medial activity also to be generally associated with different high-g tasks, this will suggest an interpretation of g in terms of a specific frontal network important in the brain's response to diverse cognitive challenges.

To show that g is associated with a relatively restricted neural system is not, of course, to show that it cannot be divided into finer functional components. For the future, indeed, a central problem will be development of more detailed models ofg in terms of component frontal functions and their interactions. Meanwhile, the almost century-long debate between rival theories of g reflects the interpretational limitations of correlational data. The present data offer hope that the neural basis for g may prove a more tractable problem. They suggest thatg reflects the function of a specific neural system, including as one major part a specific region of the lateral frontal cortex.

  • * To whom correspondence should be addressed. E-mail: john.duncan{at}mrc-cbu.cam.ac.uk

REFERENCES AND NOTES

View Abstract

Navigate This Article