From instantiation to information: Explicating the relations of instantiated and informational features
When making categorization decisions, people can categorize new items that bear little resemblance at the level of features to items they were trained on (e.g., Yamauchi & Markman, 2000). Such transfer of the informational content of study items to novel situations has long been taken as central evidence for the capacity for abstract thought. In order to recognize new category exemplars that bear no similarity at a featural level to previously encountered exemplars, people must be capable of encoding the abstract information that the studied features instantiate. People must be capable, that is, of extracting and using informational representations of features, or informational features (Brooks & Hannah, 2006).
However, as Brooks and Hannah (2006) showed, people are not only sensitive to the particular details of a given feature’s instantiation, they will often base their categorizations on the presence of a single familiar feature, even if that feature is contradicted by more numerous features, if those more numerous rivals take an unfamiliar instantiation. Under some conditions, a familiar instantiation is taken more seriously than familiar information alone. People must be capable, that is, of extracting and using instantiated representations of features, or instantiated features, when using conceptual information to make decisions.
Brooks and Hannah argued (2006) that in the course of learning about categories, people acquire both informational and instantiated features, as each has a different set of properties that prove critical under different types of tasks. As illustrated in Figure 1, a feature representation consisting of a particular instantiation (bottom left) is necessarily more selective than a more generic feature representation (denoted paw, bottom right). The narrower selectivity of instantiated features relative to informational features allows for rapid categorization even under impoverished viewing conditions for many natural categories. Knowing what cats’ paws look like can be enough to recognize a cat from just a flash of paw darting out from under a bed. In contrast, the greater generality of informational features makes them useful for reasoning across category bounds, and for efficiently summarizing commonalities that persist across variation. Informational features are thus useful for communicating, especially when teaching learners about a category, and for reasoning. An instantiated feature may allow recognition of a cat as a cat, but informational features allow recognition of a cat as a mammal.
Understanding how informational and instantiated features are related can help us understand the kind of mechanisms by which they arise, and the relations among the decision processes they support. As the difference between the two types of feature representations is the degree of abstraction, understanding how these two types of feature representations relate to one another can help understand how abstract and concrete forms of thought in general relate to one another, and how abstract thought may arise out of concrete experience.
Pothos (2005) has suggested that abstract thought in the form of rule use differs from reasoning by similarity among concrete representations in the sparsity of the mental representations underlying these decision processes. Extending this to the distinction between informational and instantiated features would lead to the proposal that informational features are simply sparser versions of instantiated features, perhaps derived from some automatic averaging or aggregation process like Hintzman’s (1986) Minerva-2 model. The generation of schematic representations from particular instances was a major motivation for Minerva-2 development. A Minerva-like process would passively generate context-specific feature abstractions by activating all memory traces of particular features, weighted by similarity to some current feature, and then summing the weighted traces. In a resonance-like fashion, similarities across traces would reinforce one another, while differences would cancel out. The more varied the activated traces, the sparser the constructed informational feature; the more varied the instantiations a person has experienced, the more abstract their abstract thought becomes.
Rather than a passive process of aggregation across instantiations, informational features could be derived by actively comparing aligned instantiations. This would enable the identification of higher-order relational properties (e.g., thin neck) or to identify smaller-scale features within the macrofeatures that occur across exemplars (e.g., number of legs), or both – relational properties based on smaller scale features occurring within the macrofeatures. Gentner and Medina (1998) described rules as emerging from particular instances though just such a process of aligned abstraction. In contrast to the mere sparsity proposal outlined above, informational features would not be merely sparser instantiations, but would embody new information, or would make explicit what had been tacit.
Although the mere sparsity and aligned abstraction accounts postulate different relations between informational and instantiated features, both treat informational features as being derived from instantiated features. We could imagine, however, that they are generated in separate, parallel processes. For example, Ashby, Alfonso-Reese, Turken & Waldren (1998) have outlined a model of conceptual organization and processes, COVIS, based on the idea that there are independent verbal and perceptual conceptual systems. This suggests that informational features may be products of an explicit verbal system, while instantiated features are produced by an implicit perceptual system. Although correlated through both being extracted from the same stimuli, the two forms would be causally independent. Activating informational features, therefore, would not necessarily activate instantiated features. Applying perfectly reliable verbal rules should not be influenced by the similarity of features.
Strong versions this separate representations account seem refuted by the findings of Thibaut and Gelaes (2006, Experiment 3A & 3B). In replicating work by Regher and Brooks (1993), they gave participants test items consisting of near reproductions of training items, with one feature given a new value, and with the features spatially separated to eliminate any contributions of holistic representations. In training, participants were given the perfectly reliable rule defining the set. Half of the test items were good transfer items—similar to, and in the same category, as soem training item—and half were bad transfer items—similar to, yet in the opposite category, as some training item. The similarity of the bad transfer items to training rivals resulted in slower reaction times and lower levels of accuracy in a speeded classification task. Thus, even the use of an explicit, perfect rule is influenced by the presence of familiar features. While this could be accommodated by weaker versions of a separate systems account that allows for associative links between instantiated and informational features to arise due to the temporal contiguity of their origin and use, it is not clear how to distinguish this from the aligned abstraction account. Further, if the links are pervasive and common, it is unclear how the supposed separate systems can be held to be separate.
Our goal in this paper is to explore how the contents of informational and instantiated features are related, whether they differ quantitatively, as in the mere sparsity account, or whether they differ qualitatively, as in the aligned abstraction account. In two experiments, participants are trained to discriminate between members of two species of imaginary animals that are nearly identical to those used in Brooks and Hannah (2006). After training, participants are given novel test items to categorize, all of which follow the same rule used to create the training items. We are interested in contrasting the response to familiar features of people who rely on informational features to make their decision versus those who rely on instantiated features to make their decisions, measured with both behavioral and ERP measures.
If informational features are just sparse instantiations, we would to see similar, if attenuated, behavioral and ERP responses. If informational features are qualitatively different, but derived from instantiated features and still grounded in particular instantiations, then at least ERP responses should differ, although we may expect to see commonalities in behavioral measures. If people reliant on informational features show no differential responsiveness to familiar and novel features, then the strong version of separate representation account has to be reconsidered.
This project thus requires we have some way of encouraging a reliance on either informational or instantiated features when making categorization decisions. Brooks and Hannah (2006) suggested that everyday rules function as pointers to which instantiations are important. In Hannah and Brooks (2006) we expanded on this argument, noting that for most everyday categories, instantiations, at least at the basic level of category, rarely overlap across categories, as depicted in Figure 1. Thus, the categorization rules generated by people reliant on instantiated feature need not contain any decision rule to resolve conflicts, resulting in the feature-list form typical of most everyday category rules (Rosch & Mervis, 1975). In contrast, because of the greater generality of informational features, as illustrated in Figure 1, exemplars can contain features that, at an informational level, overlap rival categories. When decision-making is reliant on informational features, the relatively weak feature-list rules typical of most naturalistic rules are insufficient to resolve feature conflicts, and must be supplemented by some internal rule. For example, to create our materials, we used a disjunctive rule with a numeric threshold. An item is a bleeb in our categorization scheme if it has at least two of rounded head, rounded torso, two legs or striped pattern. This decision-threshold is the internal rule that resolves feature conflict.
Hannah and Brooks (2006) split participants according to whether they produced weak, feature-list rules or strong rules, containing an explicit mechanism for resolving feature conflict, which always involved a counting mechanism. As expected, participants producing listing rules were sensitive to instantiation familiarity when making decisions, consistent with their decision-making being reliant on instantiated features. Those producing counting rules, however, showed no sensitivity to the familiarity of instantiations, consistent with their decision-making being more reliant on informational features.
Our first step is to demonstrate an effect of instantiation familiarity in even users of strong rules using behavioral measures, that is, of those reliant on informational features when making categorization decisions. This is needed to support the claim that there is a close association between instantiated and informational features, as would be expected if informational features are derived from instantiated features. This assumption is shared by both the mere sparsity and aligned abstraction hypotheses. The two hypotheses hold different implications regarding the time course of ERP activity, which is explored in Experiment 2. Although Thibaut and Gelaes (2006, Experiment 3A & 3B) provided some evidence that users of strong rules were still sensitive to feature instantiation when making categorization decisions, their training stimuli do not obviously allow for a distinction between informational and instantiated features to emerge. Furthermore. The test stimuli they used were neither characterized by a single familiar feature nor were they intact items. Their effect actually disappeared when they used intact items (Experiment 3C). It is unclear then the extent to which their data represents a sensitivity to instantiated features by people reliant on informational features.
Experiment 1: Instantiation Familiarity on Strong- and Weak-Rule Use
In our first experiment, all participants were given the same training based on the supported induction procedure used by Brooks and Hannah (2006), and using materials closely based on their stimuli. Following training, they given either a weak, feature listing rule (listing group), or a strong rule based on counting features (counting group). After demonstrating they knew the rule, participants were given novel examples of each category to classify. Given that our training stimuli, shown in Figure 2, manifest feature conflict only at the level of informational features, we expect that the use of a strong rule should orient users to informational features, while the use of a weak rule should orient users to instantiated features.
Each novel transfer item was based on the informational structure (Table 1) of a non-prototype training item, and took on three forms. For example, a non-prototype bleeb training item is shown in the centre of Figure 3, with bleeb and ramus prototypes shown above. Beneath the training item are the three versions of a test equivalent. The left test item is an entirely novel equivalent with one familiar feature seen in most bleeb training items (facilitated-familiarity item), the middle test item consists of entirely novel feature instantiations (all-novel item), while the right test item consists of novel feature instantiations, with one familiar feature seen in most training ramuses (interfering-familiarity item).
We expect that for all participants, the speed of correct identification of the test items will vary depending on whether an item contains a familiar instantiation. We expect the listing group, reliant mainly on instantiated features, will be quickest to correctly categorize facilitated-familiarity items, and slowest to categorize interfering-familiarity items. Similarly, listing participants should be most accurate on facilitated-familiarity items and least accurate on interfering familiarity items. We expect counting participants to show a smaller, but significant, effect of test item on reaction time; because they are given a perfect rule, there may not be a significant effect of test item on accuracy. We are also agnostic regarding whether the trends of the two groups will be the same or not as the behavioral measures are too crude to allow us to make inferences about the relations of the underlying feature representations. For example, the counting groups’ use of an explicit decision rule may change the pattern of overt behavioral responses from that of the listing participants regardless of the relations between informational and instantiated features.
Methods
Participants
Thirty-three McMaster undergraduate students participated in exchange for credit in first- or second-year psychology courses. Participants were randomly assigned to groups, with 16 in the listing group, and 17 to the counting group. One participant in the counting-rule condition was dropped for not using the rule. Data analyzed here therefore comes from 32 participants, 16 in each of the two rule conditions.
Stimuli and apparatus
The experiment was conducted on a 2.4 GHz Pentium 4 computer, running the Windows 2000 operating system. Stimuli were displayed on a Samsung SyncMaster 17” color monitor, with a refresh rate of 75 Hz, and a screen resolution of 1024 X 768 pixels. The screen was set 80 cm from the participant; viewing distance was fixed by a chin rest. Stimulus presentation and response collection was controlled by Presentation software.
Training items. Training items from both categories are shown in Figure 2.Stimuli consisted of line drawings of imaginary animals nearly identical2 to those used in Hannah and Brooks (2006), and consisted of two categories of animals, bleebs (left column, Figure 2) and ramuses (right column, Figure 2). Every category consisted of a prototype item having all four of the characteristic values of its category, shown in the top row of Figure 2, and four one-away items, each of which deviated from the prototype on a single dimension, for a total of ten items. Each category was defined by characteristic values on four different dimensions: head shape (round/angular), torso shape (round/angular), torso pattern (stripes/dots), and number of legs (two/four). The informational structure of the categories is summarized in Table 1.
The categories are created around a family-resemblance structure (Rosch & Mervis, 1975) so that no single feature is necessary or sufficient for classification. Instead, category membership involves a counting rule: an item is a member of category X if it has at least two of the characteristic features of category X. An item was a bleeb if it had at least two of rounded head, rounded torso, stripes or two legs; an item was a ramus if it had at least two of angular head, angular torso, dots or four legs.
One-away items deviated from their prototypes by taking on the characteristic value of the rival category, but only at an informational level. For example, the two items in the second row of Figure 2 both deviate from their respective prototypes on head shape: the bleeb has an angular head and the ramus has a rounded head. The head of the bleeb one-away, however, looks very different from the head of any ramus, and neither does the head of the ramus one-away resemble any bleeb head.
Test items. Test items were novel instantiations of the eight training one-away items. In Figure 3 a bleeb one-away from training is shown centrally, with the two prototypes above it, and three test items based on the one-away are shown beneath it, with one example from each of the three test conditions. Items in the facilitated familiarity condition (bottom left) were given a familiar feature from the actual category’s prototype; in the example shown in Figure 3, the facilitated familiarity item has the torso seen in the bleeb prototype and most bleeb training item. Items in the all-novel test condition (bottom middle) consisted of features that were perceptual novel instantiations of the one-away training item; in the example shown in Figure 3, the all-novel item contains a novel pointed head, rounded torso, stripes and two legs. Items in the interfering familiarity condition (bottom right) were given a familiar feature from the other category’s prototype; in the example shown in Figure 3, the interfering familiarity item has the pointed head seen in the ramus prototype and most ramus training items.