Representing Categorical Knowledge: An fMRI Study

This paper investigated the nature of mental representation of categorical knowledge by examining the neural substrates that support classification and feature inferences. The results from one experiment suggest that frontal lobe regions were activated when an inference task was given, while temporal lobe regions, especially the fusiform gyrus, became active when a classification task was given. These results are consistent with the view that categorical knowledge is mediated by two separate neural structures, and further indicate that two interrelated functions of categories – classification and feature inference – are processed by distinct brain regions. Human knowledge is organized largely by categories that we form in everyday situations. Many of man-made names, such as “terrorists,” “fundamentalists” or names of natural objects such as dogs and cats are all categorically arranged. Perhaps, because of this generic characteristic of knowledge representation, we are able to carry out many complex and abstract tasks such as communication, reasoning, and solving math problems relatively easily. How do we represent categorical knowledge in our brain? In this paper, we examined the neural structures that support the representation of categorical knowledge. We addressed this problem by comparing two basic functions of categories – classification and feature inference. Classification and inference are two of the most important functions of categories (Smith, 1994). The very reason why human concepts are organized categorically is said to license inferential predictions (Anderson, 1990; Rosch, 1978). In this regard, we consider that scrutinizing the underlying neural structures of these two functions would provide a crucial clue to understand the nature of mental representations of categorical knowledge. Despite the close relationship between classification and feature inference, much research in concept formation in the last 40 years has focused on the investigation of classification (see Murphy 2002 for review). These studies suggest that subjects acquire some form of “category-level knowledge” after extended training of classification (see Medin & Schaffer, 1978; Posner & Keele, 1968). That is, (1) subjects are able to classify new items accurately even though they have never seen exactly the same items during a learning phase; (2) their accuracy of classification transfer depends on the degree of feature overlap between a transfer item and studied items; (3) subjects generally classify non-studied prototypes (average instances) as accurately as studied items. A dominant theory of category learning explains that the aforementioned effects of category learning arise because participants in an experiment make classifications by measuring similarities between a new target item and old exemplars that were stored in memory. In this theory, the mental representation of category-level knowledge consists of the episodic memory trace of the instances that were experienced during learning (Medin & Schaffer, 1978; Nosofsky, 1986; Kruschke, 1992). Although this theory has received an impressive amount of support from a variety of studies, this view has met serious challenge from two sources: (A) Patients who have severe damage to hippocampus and the medial temporal lobe learn to classify new items as accurately as normal control subjects, while their recognition memory for training items is severely impaired (Knowlton & Squire, 1993; Squire & Knowlton, 1995; but see also Nosofsky & Zaki, 1998; Palmeri & Flanery, 1999; for an alternative explanation); (B) The extent to which specific exemplar information affects category-level representations depends on the task and the structure of categories that are employed in a category learning task (Smith & Minda, 1998; Maddox, et al., 2003; Markman & Ross, 2003; Ross, 1997; Smith, Patalano, & Jonides. 1998; Strange et al. 2001; Yamauchi & Markman, 1998). For example, recent neuroimaging studies have reported increased activities in the vicinity of the frontal lobe during categorization tasks that require an application of abstract rules (Smith et al. 1998; Strange et al. 2001), while the temporal lobe is responsible for object recognition and classification (e.g., Haxby et al. 2001) These findings indicate that the mental representation of categorical knowledge is likely to consist of multiple neural substrates – one that processes similarity-based associative information, and the other that assesses rule-based abstract information (Ashby, et al., 1998; Erickson & Kruschke, 1998; Smith et al., 1998). Consistent with this suggestion, our studies have indicated that different types of information are used for classification and feature inference: Subjects often rely on the information about concrete instances to make classifications, while they use abstract rule-like information to make feature inferences (Yamauchi & Markman, 2000). On the basis of these findings, we reasoned that if categorical knowledge in general is represented by multiple neural substrates, then the category information acquired by a particular single learning task should generate a mental representation that encompasses separate brain regions; furthermore, such a representation should be selectively accessed by a classification test and an inference test. We tested this hypothesis in the following experiment. In our experiments, participants learned two arbitrary categories incrementally either in a classification task or in a feature inference task (Nosofsky, 1986; Medin & Schaffer, 1978; Yamauchi & Markman, 1998). Stimuli were geometric figures similar to those used in the Medin and Schaffer study (1978) (Figure 1). In a classification task, participants predicted the category membership of a stimulus on the basis of the information about other four features; in an inference task, participants predicted an unknown value of a feature on the basis of the information about the other three features and the category membership of the stimulus. After each trial, participants received feedback. Our main focus of investigation was concerned with the characteristics of the mental representation thus formed after classification learning or inference learning. Given the hypothesis that category representation incorporates two types of information – similarity-based exemplar information and rule-based abstract information, we hypothesized that two different brain regions should become primarily active in the same subjects. More specifically, we predicted that, irrespective of the format of learning, the frontal lobe region that supports rule-based abstract information becomes active when an inference transfer task is given; in contrast other cortical areas such as the medial temporal lobe should become primarily active when a classification task is given.