Knapp Recognizing Distinctive Faces 1 Running head : RECOGNIZING DISTINCTIVE FACES Recognizing Distinctive Faces : A Hybrid - Similarity Exemplar - Model Account

In recognition-memory experiments, Nosofsky and Zaki (2003) found that adding discrete distinctive features to continuous-dimension color stimuli helped participants to both identify the old items as old (the old-item distinctiveness effect), as well as identify the new items as new. The present study tests the extent to which these results generalize to the domain of face recognition. Two experiments were conducted, one using artificial faces and one using natural faces. Artificial faces were used to test memory for faces with discrete distinctive features while controlling the similarity of the faces themselves on more continuous dimensions. The natural faces experiment used the faces of 40 bald men categorized into three groups (typical, isolated, and discrete-feature distinct) based on experimental ratings of distinctiveness. In both experiments, there were strong effects of the distinctive features on recognition performance. The data were accounted for reasonably well by a hybrid-similarity version of an exemplar-recognition model (Nosofsky & Zaki, 2003), which includes a feature-matching mechanism that can provide boosts to an item’s self-similarity. Knapp Recognizing Distinctive Faces 3 According to exemplar models of old-new recognition (Gillund & Shiffrin, 1984; Hintzman, 1988; Lamberts, Brockdorff, & Heit, 2003; Medin & Schaffer, 1978, Nosofsky, 1988), people represent lists of study items in terms of individual exemplars, with each exemplar corresponding to an individual study item. In most versions of these models, test items are assumed to give rise to a global activation of the exemplar-based memory representation. The greater the degree of activation, the more familiar is the test item and so the greater is the probability that the observer judges the item to be old. A representative of this class of global-familiarity exemplar models is the generalized context model (GCM) (Nosofsky, 1986, 1991). In the GCM, exemplars are represented as points in a multidimensional psychological space, and the similarity between exemplars is a decreasing function of distance in the space. In the model, global activation or familiarity is based on the summed similarity of a test item to all of the stored study-list exemplars. The model has been applied primarily in domains involving the recognition of simple perceptual stimuli varying along a few salient dimensions, such as colors, schematic faces, and geometric forms varying in size and orientation. In these domains, fine-grained measurements can be obtained of the similarity between exemplars in the multidimensional space. Thus, the model allows one to make fine-grained predictions of the probability with which individual items are judged as old or new. Indeed, there have been numerous demonstrations of the ability of the model to account in quantitative detail for rich sets of individual-item, old-new recognition data (Nosofsky, 1991; Shin & Nosofsky, 1992; Zaki & Nosofsky, 2001; for closely related work, see Kahana & Sekuler, 2002; Lamberts et al., 2003). Knapp Recognizing Distinctive Faces 4 However, most of the successes of the GCM have involved its ability to predict how false alarm rates associated with new items vary with their similarity to old items. A potentially important limitation of the standard model is that it fails to account for patterns of results involving hit rates to old items. Specifically, because of its summed similarity rule, the standard GCM predicts that "typical" old items should have higher hit rates than "distinctive" old items (Valentine & Ferrara, 1991). Although the constructs of "typicality" and "distinctiveness" are open-ended ones, the general idea is that a distinctive item is one that is relatively isolated in the multidimensional similarity space of studied exemplars, whereas typical items lie in more densely clustered locations of the similarity space. Thus, the summed-similarity associated with typical items is greater than that associated with distinctive items. Therefore, typical old items should give rise to a greater degree of familiarity and so should have higher hit rates. A similar qualitative prediction is made by most other global-familiarity models of old-new recognition (e.g., Gillund & Shiffrin, 1984; Hintzman, 1988). Much research, however, especially from the face recognition literature (Bartlett, Hurry & Thorley, 1984; Light, Kayra-Stuart & Hollander, 1979; Valentine & Ferrara, 1991; Vokey & Read, 1992), suggests that distinctive old items have higher hit rates than do typical old items, thereby providing a direct challenge to such models. A good example of such a result is found in a face-recognition study conducted by Busey and Tunnicliff (1999). Busey and Tunnicliff had participants provide similarity ratings for a large set of naturalistic male faces. They analyzed the similarity data by using multidimensional scaling (MDS) techniques and located the faces in a six-dimensional psychological space. They then used the scaling solution in combination with the GCM Knapp Recognizing Distinctive Faces 5 to predict participants' performance in an old-new recognition task involving the same set of faces. Busey and Tunnicliff (1999) found that distinctive faces, i.e., those lying in isolated regions of the multidimensional similarity space, gave rise to higher hit rates than did typical faces, in direct contrast to the predictions from the GCM. However, as acknowledged by Busey and Tunnicliff (1999), besides lying in isolated regions of the space, many of the distinctive faces also had highly salient specific features such as beards, whereas faces lying in denser more typical regions did not. Thus, the isolation of the faces along the continuous dimensions of the space was confounded with the presence of idiosyncratic discrete features. To investigate further the basis of Busey and Tunnicliff's (1999) results, Zaki and Nosofsky (2001) and Nosofsky and Zaki (2003) explicitly manipulated individual-item distinctiveness in the domain of color. Unlike face space, the underlying dimensional structure of colors is well understood, with extensive scaling work indicating that colors vary along the three dimensions of hue, saturation, and brightness. Thus, one can explicitly manipulate typicality and distinctiveness by varying the location of items along these three dimensions. In several different old-new recognition studies, Zaki and Nosofsky (2001) and Nosofsky and Zaki (2003) tested conditions in which individual colors were located in either dense or isolated regions of the continuous-dimension color space. Importantly, across conditions, the same colors served as either typical or distinctive items, thereby removing any stimulus-specific effects from the main pattern of results. When distinctiveness was manipulated in this manner, these researchers found that hit rates for distinctive old items were not greater than for typical old items. In fact, the complete set of results was well predicted by the standard GCM. These studies Knapp Recognizing Distinctive Faces 6 suggested that mere isolation in a continuous-dimension similarity space is not sufficient to produce a robust hit-rate advantage for distinctive old items. Thus, to pursue further the findings from Busey and Tunnicliff (1999), Nosofsky and Zaki (2003) conducted additional studies in the domain of color in which distinctiveness was manipulated by including idiosyncratic discrete features on the studied objects. Specifically, they tested designs in which various discrete alphanumeric characters were added to a few members of a set of continuous-dimension color patches. Under these conditions, a strong old-item distinctiveness effect was observed: Old items with a discrete distinctive feature added to them had significantly higher hit rates than did old items without such features. This result is similar to the previously discussed results for faces from Busey and Tunnicliff (1999). In addition, as discussed more fully later in our article, adding the distinctive features to foil items made it easier for participants to realize that the foils were new. To account for the distinctiveness effects, Nosofsky and Zaki (2003) developed a modified version of the GCM called the hybrid-similarity GCM (HS-GCM) (for closely related ideas in the domains of similarity and classification, see Lee & Navarro, 2002; Navarro & Lee, 2003; Verguts, Ameel, & Storms, 2004). This extended version of GCM incorporates the ideas of Tversky’s (1977) feature-contrast model (FCM) into the MDS framework of the traditional GCM. The basic assumption in the FCM is that the similarity between two objects is based on measures of their common and distinctive features. Specifically, similarity is an increasing function of the measure of the objects’ common features, and a decreasing function of the measures of the objects’ distinctive features. Of crucial importance, whereas standard MDS approaches assume that all Knapp Recognizing Distinctive Faces 7 items have an equal degree of self-similarity, FCM allows for differing degrees of selfsimilarity, with an increase in the number of common matching features increasing the measure of self-similarity. Thus, within the FCM framework, an item with a highly salient feature will have a higher measure of self-similarity than an item without a highly salient feature. In the HS-GCM, the similarity between two objects is determined jointly by their distance in a continuous-dimension psychological space and by the extent to which they have matching or mismatching discrete features. Within this extended framework, if a distinctive item’s self-similarity is sufficiently high, its summed similarity can exceed that of more typical items. This increase in summed similarity leads to a corresponding increase in familiarity and results in an increased probability of judging the distinctive item as “old.” Indeed, the inclusion of the feature-matching mechanism allowed the HS-GCM to match the quantitative results of Nosofsky and Zaki (2003) whereas the traditional version of GCM could not. While the results of Nosofsky and Zaki (2003) provided important preliminary evidence in favor of the new model, the stimuli used in their experiments were highly artificial (color patches combined with alphanumeric markings). Objects in the natural world are rarely that simplistic. Therefore, the main goal of the present research was to explore the extent to which the results of Nosofsky and Zaki (2003) generalize to a more natural domain, namely that of face recognition. Given that it has proven difficult for GCM to explain results in this domain in the past (Busey & Tunnicliff, 1999; Valentine & Ferrara, 1991), it seems an especially appropriate test for the HS-GCM. Because faces are highly variable and complex, we conducted two experiments, one using artificial faces and one using natural faces. Artificial faces were used as an Knapp Recognizing Distinctive Faces 8 experimental intermediary between the easily manipulated domain of color and the much less easily manipulated domain of faces. Artificial faces were derived using principal component analysis (PCA) on images from the Facial Recognition Technology (FERET) database (Phillips, Moon, Rizvi & Rauss, 2000). Because the similarity of these faces on more continuous dimensions could be manipulated, foils of varying degrees of similarity to the distinctive targets could be created (see the Method section), thus allowing for direct comparison of results with those of Nosofsky and Zaki (2003). In a second experiment, we provided further tests of the modeling approach in a naturalistic-face domain. With natural faces, no predetermined continuous dimensions exist to explicitly create typical versus distinct items. Additionally, many potential discrete distinct features exist on natural faces. To create a reasonable but challenging test bed for the model, we selected a subset of the naturalistic faces used in Busey and Tunnicliff’s study. We selected the study and test items so as vary both item typicality and the number of faces with highly salient discrete distinct features. A multidimensional face space was constructed by collecting similarity and distinctiveness ratings of the faces used in the experiment. We then used the derived face space and the distinctiveness ratings in combination with the HS-GCM to predict the old-new recognition judgments.