Analogy involves a structured comparison, or mapping, between one situation (source) and another (target). For instance, a reasoner may be given a problem such as:

and be asked which word, CAVE or HONEY, completes the analogy. To choose CAVE, the participant would need to realize that birds live in nests as bears live in caves while not being distracted by the fact that bears eat honey. Using several priming tasks, Spellman et al. (Reference Spellman, Holyoak and Morrison2001) investigated whether analogy might just be a consequence of the organization of concepts in semantic memory. They found that unlike traditional semantic priming, “analogical” priming was not automatic and instead required the participant to direct attention to relations between word pairs. This suggested that controlled retrieval of a bound relation into working memory (WM) may be a necessary process for analogical reasoning. Subsequent experiments demonstrated that WM was indeed important for analogical mapping (e.g., Morrison et al. Reference Morrison, Holyoak, Truong, Schunn and Gray2001), as well as relational binding (see Morrison Reference Morrison, Holyoak and Morrison2005), a finding confirmed using functional magnetic resonance imaging (fMRI; Bunge et al. Reference Bunge, Wendelken, Badre and Wagner2005).

WM is also important for suppressing distracting information, such as irrelevant semantic associates or featural similarities likely to enter WM during analogical retrieval and mapping. Waltz et al. (Reference Waltz, Lau, Grewal and Holyoak2000) demonstrated that adults performing a semantically rich scene-analogy task shifted from preferring analogical to featural mappings under WM dual-tasks. Using the same task, Morrison et al. (Reference Morrison, Krawczyk, Holyoak, Hummel, Chow, Miller and Knowlton2004) found that frontal patients with damage to WM areas showed a similar pattern. Morrison et al. also developed an A:B::C:D or D′ verbal analogy task that required participants to choose between D (analogically correct choice) and D′ (foil), which were both semantically related to the C term of the analogy. When the foil was more semantically associated to the C term than was the correct choice, frontal patients performed near chance. In contrast, semantic dementia patients who exhibited profound decrements in relational knowledge performed poorly on all of the verbal analogies regardless of the degree of semantic association between C:D and C:D′. Using the same task, Cho et al. (Reference Cho, Moody, Cannon, Poldrack, Knowlton and Holyoak2007b) found that individuals who scored higher on the Raven's Progressive Matrices (RPM) showed greater fMRI activation increase in neural areas, including the prefrontal and visual cortices, on trials in which reasoners had to reject foils that were highly associated with the C term. This finding suggests that there are neural regions whose level of activation for interference resolution during analogical reasoning relates to individual differences in fluid intellectual capacity.

Many real-world analogies, as well as reasoning tasks developed for psychometric purposes such as the RPM and People Pieces Analogy task (PPA; Sternberg Reference Sternberg1977b), require integration of multiple relations to map more relationally complex analogies. Numerous fMRI studies (e.g., Christoff et al. Reference Christoff, Prabhakaran, Dorfman, Zhao, Kroger, Holyoak and Gabrieli2001; Kroger et al. Reference Kroger, Sabb, Fales, Bookheimer, Cohen and Holyoak2002) have shown increasing levels of activation in anterior prefrontal cortex for more relationally complex RPM problems, a finding consistent with a neuropsychological study with frontal patients (Waltz et al. Reference Waltz, Knowlton, Holyoak, Boone, Mishkin, de Menezes Santos, Thomas and Miller1999). Using an adaptation of the PPA task, Viskontas et al. (Reference Viskontas, Morrison, Holyoak, Hummel and Knowlton2004) found that older adults showed decrements in both relational integration and relational distraction. Using this same task, Cho et al. (Reference Cho, Holyoak and Cannon2007a) found that executive resources are shared between relational integration and interference resolution during analogical reasoning. In an fMRI follow-up study, Cho et al. (Reference Cho, Moody, Poldrack, Cannon, Knowlton and Holyoak2007c) found partially overlapping but distinct regions within inferior frontal gyri (IFG) showing sensitivity to each component process of analogical reasoning. Separate regions that showed exclusive sensitivity to each component process were also identified within IFG. In addition, the degree of activation increase in the right ventral IFG during trials in which participants had to integrate three relations (compared to one) was greater for individuals whose performance accuracy was higher.

Although the above studies do not directly deal with the development of analogy during childhood, they do clearly demonstrate several component processes involved in analogical reasoning that are dependent on prefrontal cortex, an area of the brain that actively develops throughout childhood (Diamond Reference Diamond, Stuss and Knight2002). In an effort to explore these processes directly in children, Richland et al. (Reference Richland, Morrison and Holyoak2006) developed a scene-analogy task manipulating both relational complexity and featural distraction. Even 3-year-olds could solve simple (one-relation, no-distraction) problems, but they had difficulty if the problem required integration of multiple relations or ignoring a featurally similar object. Similarly, Wright et al. (2007) performed an fMRI study with children using another semantically rich visual analogy task, and found that brain activation in areas associated with relational integration was the best predictor of analogy performance. Wright et al. also found that these areas, which are not associated with semantic retrieval (Bunge et al. Reference Bunge, Wendelken, Badre and Wagner2005), become more and more engaged over the same time period in which children dramatically improve in their ability to solve more relationally complex problems (Richland et al. Reference Richland, Morrison and Holyoak2006).

We are highly sympathetic with the target article's efforts to computationally model the development of analogy, and we certainly don't dispute the importance of relational knowledge in development. However, we believe that a successful model of development must (1) explain how knowledge representation and process constraints interact to produce the changes in analogy observed in children, including increases in ability to perform relational integration and resist featural distraction; and (2) explain how an architecture consistent with the demands of adult analogical reasoning develops. Unfortunately, the connectionist model described in the target article does not meet these requirements. In contrast, Morrison and collaborators have used LISA (Learning and Inference with Schemas and Analogies; Hummel & Holyoak Reference Hummel and Holyoak1997; Reference Hummel and Holyoak2003), a neurally plausible model of analogical reasoning, to successfully simulate many of the developmental and neuropsychological results discussed in this commentary (e.g., Morrison et al. Reference Morrison, Krawczyk, Holyoak, Hummel, Chow, Miller and Knowlton2004; Reference Morrison, Doumas, Richland, McNamara and Trafton2006; Viskontas et al. Reference Viskontas, Morrison, Holyoak, Hummel and Knowlton2004).

We believe that the development of analogical reasoning is best conceptualized as an equilibrium between children's relational knowledge and their current processing ability. As children mature, their prefrontal cortices more efficiently implement WM and thereby can process more complex analogies. However, more efficient relational representations can impose fewer processing demands at any given age, which is why a child who becomes an expert in a given domain can show rapid progress even though the child's WM system has not improved (Morrison et al. Reference Morrison, Doumas and Richland2007). This framework can account for the observed changes in children's analogical reasoning, as well as subsequent changes in analogy during normal and abnormal human aging. It can also be simulated in symbolic-connectionist models of relational learning and reasoning (e.g., Doumas et al. Reference Doumas, Hummel and Sandhofer2008; Hummel & Holyoak Reference Hummel and Holyoak1997; Reference Hummel and Holyoak2003).