In their target article, Jung & Haier (J&H) propound “the P-FIT model” implicating a variety of cortical nodes as the network responsible for supporting general intelligence and reasoning. This model, however, appears insufficient to explain the neural basis for general intelligence, although it could successfully underpin fluid reasoning function or working memory.

As J&H mention, “general intelligence” refers to intellectual ability in general, which is conceptually somewhat different from the general factor g or fluid reasoning ability. In the psychometric tradition, it is widely accepted that there are two related but distinct components of general intelligence, referred to as “fluid (gF) and crystallized (gC) general intelligence” (Cattell Reference Cattell1963; Reference Cattell1987). gF generally refers to reasoning and novel problem-solving ability, to be able to see relationships, as in analogies and letter and number series, which is independent from prior experience and learned knowledge. In contrast, gC is cognitive functioning based on previously acquired knowledge available in long-term store, including semantic knowledge and episodic memory. Indeed, most psychometric batteries for general intelligence (i.e., IQ tests; for example, the Wechsler-derived batteries, the Thorndike test, the Kaufman test) contain diverse verbal subtests to assess the cognitive skills belonging to gC, whereas pictorial, matrix-based reasoning tests such as the Raven's tests assess largely gF. Notably, factor analysis suggests that the Wechsler-derived batteries are biased toward crystallized content and even that the “verbal,” or gC, factor is the most valid intelligence factor (Ashton et al. Reference Ashton, Lee and Vernon2001; Robinson Reference Robinson1999; Reference Robinson2005).

More importantly, evidence on the two-factor theory has been provided by numerous longitudinal behavioral studies demonstrating that diverse cognitive skills and functions belonging to gC and verbal ability (such as performance on the WAIS verbal subtests), persist and even improve for decades after adolescence. In contrast, cognitive fluid reasoning and working memory peak in the third decade and then decline (Botwinick Reference Botwinick, Birren and Schaie1977; Dixon et al. Reference Dixon, Kramer, Baltes and Wolman1985; Kaufman et al. Reference Kaufman, Kaufman-Packer, McLean and Reynolds1991). This is thought to be a reason why scientists especially in quantitative disciplines, who need fluid intelligence, mainly produce their best work in their 20s and 30s, whereas those in the field of history and philosophy produce their best work in their 40s, 50s, and beyond as they have accumulated more knowledge.

More causal evidence for the gC/gF distinction has come from studies of patients with brain damage. Patients with prefrontal damage showed profound deficit in resolving many reasoning tasks, whereas those with anterior temporal lobe damage revealed bad performance on the tests of declarative knowledge (Waltz et al. Reference Waltz, Knowlton, Holyoak, Boone, Mishkin, Santos, Thomas and Miller1999). Other lesion studies emphasized that the frontal lobe plays a crucial role in abstract reasoning but not in general intelligence as assessed by WAIS (Duncan et al. Reference Duncan, Burgess and Emslie1995; Reference Duncan, Emslie, Williams, Johnson and Freer1996). All these findings strongly suggest that the neural bases of fluid reasoning and crystallized knowledge are dissociable. Thus, to elucidate the nature of the biology of intelligence, at least two distinguishable mechanisms should be addressed.

Over the last decade, neuroimaging studies using various techniques, including anatomical MRI, fMRI, PET, and MRS, have rapidly unveiled the neurobiological bases of diverse cognitive functions such as fluid reasoning, working memory, and problem-solving ability (Gray & Thompson Reference Gray and Thompson2004). On the basis of commonality of the results from these studies, J&H have developed the integrated network model emphasizing functional connectivity to explicate the neural basis of general intelligence. However, this approach appears to have some intrinsic limitations to differentiate the neural basis of gC from gF or the unitary factor g.

First, almost all imaging data, regardless of both task modality and imaging method, is correlational rather than causal evidence. Commonality of these data is also correlational and therefore unable to articulate specific neural correlates of the diverse cognitive skills comprising general intelligence, although this concern could be tempered using a multiple regression approach. Second, individual differences in gF and gC exhibit robust intercorrelation in the normal cohort (r=.7 to .8; see Jensen Reference Jensen1998; Kaufman & Horn Reference Kaufman and Horn1996). Their relation could be explained by the notion that gF plays a substantial role in encoding and retrieving information in long-term store and thereby in facilitating the accumulation and expression of gC, although there are distinct neural bases for these two functional domains of intelligence. Third, the typical crystallized knowledge content of WAIS subtests “Information” and “Vocabulary” reveals high g-loadings (r=.6 to .7) despite low reliance upon fluid reasoning ability and working memory capacity (Colom et al. Reference Colom, Jung and Haier2006a; Lee et al. Reference Lee, Choi, Gray, Cho, Chae, Lee and Kim2006). Therefore, to dissect the neural mechanism specific for crystallized knowledge, more sophisticated experimental paradigms and methods are required. Such an effort would allow us to formulate a combined model of gF and gC that accounts for dissociation of gC and gF, and further, provide better prediction for individual differences in general intelligence.

Where and how is crystallized knowledge organized in the human brain? Neurobiological studies on “learning and memory” in animal models have begun to shed light on closely related questions using diverse technical approaches based on genetics, electrophysiology, pharmacology, and anatomy (Frankland & Bontempi Reference Frankland and Bontempi2005; Miyashita Reference Miyashita2004). Long-term memory, often referred to as “remote” memory in neurobiology, is fundamental for maintaining crystallized knowledge. Several lines of evidence have demonstrated that the hippocampus functions as a temporary store for new information, whereas more permanent storage depends on a distributed cortical network including the anterior cingulate, the lateral prefrontal, and the temporal cortices. In monkeys, memory traces representing repeated associations (likely the basis of semantic-like memory) are consolidated in the domain-specific regions in the temporal cortex (Sakai & Miyashita Reference Sakai and Miyashita1991; Yoshida et al. Reference Yoshida, Naya and Miyashita2003). During memory consolidation, structural reorganization of cortical circuits occurs, which includes the addition/elimination of synapses and modulation of axonal dendritic growth through a cellular program of gene expression that eventually may lead to changes of cortical gray matter density and thickness (Chklovskii et al. Reference Chklovskii, Mel and Svoboda2004; Tokuyama et al. Reference Tokuyama, Okuno, Hashimoto, Xin Li and Miyashita2000).

Consistent with this account based on animal models, human functional neuroimaging studies suggest that the prefrontal and temporal lobes, particularly in the left hemisphere, may be involved differentially in semantic memory representation (Martin & Chao Reference Martin and Chao2001; McClelland & Rogers Reference McClelland and Rogers2003). The left inferior prefrontal cortex generally plays a crucial role in retrieving and manipulating lexical and semantic information stored elsewhere, whereas the temporal lobe integrates semantic information with increasing convergence along its posterior-to-anterior axis. Furthermore, evidence from studies on patients with semantic dementia emphasizes a pivotal role of the anterior temporal lobe in semantic working memory as well as memory storage (Gainotti Reference Gainotti2006; Hodges et al. Reference Hodges, Patterson, Oxbury and Funnell1992; Mummery et al. Reference Mummery, Patterson, Wise, Vandenbergh, Price and Hodges1999). Thus, individual differences in gC may depend on declarative knowledge stored in the temporal lobe and association areas, leading to structural differences across individuals. If this is correct, however, it is puzzling that the structural correlates of intelligence did not tend to implicate temporal regions, and that J&H interpreted the results to mean: “temporal and occipital lobe relationships to intelligence may be functional and ‘task dependent’ on the sensory modality employed” (sect. 5.3, para. 1).

A possible explanation is technical limitation of the voxel-based morphometry (VBM) method. As commented in the target article, there has been some dispute over the application of VBM, especially to characterizing group differences (Bookstein Reference Bookstein2001; Davatzikos Reference Davatzikos2004; Mehta et al. Reference Mehta, Grabowski, Trivedi and Damasio2003). Within individuals, this method would convincingly demonstrate longitudinal changes of cortical structure through scanning on multiple occasions before and after training (Draganski et al. Reference Draganski, Gaser, Busch, Schuierer, Bogdahn and May2004; Reference Draganski, Gaser, Kempermann, Kuhn, Winkler, Buchel and May2006; Golestani et al. Reference Golestani, Paus and Zatorre2002). However, when used for group analysis, the results are vulnerable to systematic misregistration errors of the spatial normalization that are caused by simple cortical geometric differences across individuals. It, therefore, is unclear whether structural correlates based on VBM are relevant to changes of cortical gray matter density or to simple shape differences. In addition, the spatial normalization process based on the Talairach coordinate system provides just a rough registration without considering gyral and sulcal pattern variation. Even the standard Montreal Neurological Institute (MNI) template is not exactly the same size or shape as the model brain: In particular, the temporal lobe's discrepancy between the MNI template and the Talairach brain extends to about 10 mm (Brett et al. Reference Brett, Johnsrude and Owen2002; Westbury et al. Reference Westbury, Zatorre and Evans1999). This possibly contributed to the weaker implication of the temporal lobe regions in the structure correlates of intelligence as compared to other cortical regions.

Another potential problem could be the spatial resolution of anatomical images. Most structural studies of intelligence have employed T1-weighted images resolved into maximally 256×256 matrixes from 1.5-T MR scanners, and therefore a single voxel size is at least 1 mm³. On the other hand, cortical gray matter thickness varies between 1 mm and 5 mm depending on the cortical regions (average thickness=2.5 mm) (Fischl & Dale Reference Fischl and Dale2000). Notably, individual differences in cortical thickness are on the order of a few percentage points. Current resolution of anatomical imagery has been an obstacle in the way of detecting subtle structure changes in the cortex, especially the changes smaller than the voxel size.

In conclusion, methodological advances in both anatomical imaging and structural analysis would extend our understanding on the neurobiology of intelligence – including distinguishing among g, gF, and gC. For example, a high-Tesla scanner such as 7-T MRI, although still in a developing stage as regards human applications, could provide more detailed information about the structural correlates of gC and even about changes within the laminar structure of cortical gray matter; these already have been demonstrated using molecular and cellular imaging in experimental animals after semantic-like memory tasks or special working memory tasks (Maviel et al. Reference Maviel, Durkin, Menzaghi and Bontempi2004; Tokuyama et al. Reference Tokuyama, Okuno, Hashimoto, Xin Li and Miyashita2000). These future findings would open exciting avenues not only for establishing neurobiological model of general intelligence, but also for developing neurometric intelligence that explicates individual differences comparable to psychometric intelligence.