THE EFFECT OF ENHANCING PROTOTYPE SALIENCE ON THE ACCURACY OF CATEGORIZING SEXUAL ORIENTATION FROM FACES

Transcription

1 THE EFFECT OF ENHANCING PROTOTYPE SALIENCE ON THE ACCURACY OF CATEGORIZING SEXUAL ORIENTATION FROM FACES by Konstantin Tskhay A thesis submitted in conformity with the requirements for the degree of Master of Arts Graduate Department of Psychology University of Toronto Copyright by Konstantin Tskhay 2012

2 THE EFFECT OF ENHANCING PROTOTYPE SALIENCE ON THE ACCURACY OF CATEGORIZING SEXUAL ORIENTATION FROM FACES Abstract Konstantin Tskhay Master of Arts Graduate Department of Psychology University of Toronto 2012 Recent literature in person perception suggests that sexual orientation and other perceptuallyambiguous group memberships could be discerned from faces with accuracy exceeding chance levels. No study to date has explored the question of how people acquire mental representations of sexual orientation, however. In Study 1, I show that the perception of sexual orientation is distinct from that of obvious group memberships. Then, I show that people can learn prototypical representations of sexual orientation to increase their sensitivity to artificially-generated stimuli (Study 2) and to actual faces (Study 3). Last, I demonstrate that this improvement is a product of participants attunement to the patterns of facial features in the prototypes of the two categories. Together, the results demonstrate that people can learn mental representations of sexual orientation through exposure to environmentally relevant stimuli. Furthermore, the studies provide insights about how the mind learns social categories more generally. ii

3 Acknowledgments First, I would like to thank my research supervisor, Nick Rule, for all the guidance that he has provided throughout this year. Without his oversight, this and many other projects would not be the same. Having Nick as my supervisor not only improved me as a researcher, but also as a person. He continually encourages and inspires me to reach for the stars. I also would like to express my gratitude to Adam Anderson, whose research vision advanced this project and inspired me along the way. I would like to give thanks to Wil Cunningham for agreeing to join my committee. Last, I would like to thank my colleagues, friends, and family for supporting me along the way and reminding me that one day I will be able to fit psychometric functions to these data. iii

4 Table of Contents Acknowledgments...iii Table of Contents... iv List of Figures... vi Chapter 1: Introduction Mental Representations of Group Memberships Perceptual Learning and the Ecological Approach Current Work... 5 Chapter 2: Study Method Stimuli Procedure Results and Discussion Male/Female Distinction Gay/Straight Distinction Comparison between Samples Conclusions... 9 Chapter 3: Study Method Stimuli Procedure Results and Discussion Chapter 4: Study Method Participants Stimuli Procedure iv

5 4.2 Results and Discussion Chapter 5: Study Method Procedure Results and Discussion Chapter 6: General Discussion Theoretical Contributions Conclusions References v

6 List of Figures Figure 1: Gay (left) and Straight (right) consensus-based prototype face morphs generated by an independent group of participants (N = 17) Figure 2: The overall psychometric function for sex (male/female) categorization (left) and sexual orientation (gay/straight) categorization (right) in Study Figure 3: Graphical representation of the overall improvement demonstrated in Study 2. Pretraining = Solid line. Post-training = Dashed Line Figure 4: Accuracy pre- and post-training vi

7 1 Chapter 1 Introduction Whether choosing a checkout line at a grocery store or searching for a seat at a busy coffee shop, people constantly form impressions of others. These impressions are formed rapidly, automatically, and outside of conscious awareness (Bodenhausen & Todd, 2010). Furthermore, we rely upon them to instantly and implicitly categorize people into groups (Macrae & Bodenhausen, 2000). Some of these groups are obvious because the members express distinct, salient, and natural markers of their identity; for example, people are especially good at accurately categorizing each other by sex (Macrae & Martin, 2007), race (Richeson & Trawalter, 2005; Remedios, Chasteen, Rule, & Plaks, 2011), and age (Wright & Stroud, 2002). Other times, however, the markers may facilitate inferences that are less accurate. Many of these groups, like political beliefs or sexual orientation, do not generally possess obvious markers that could aid categorization; nevertheless, research in person perception has begun to show that these groups can be distinguished from one another (Tskhay & Rule, 2012). Sexual orientation is one group distinction that has received increasing attention in person perception research in recent decades. Sexual orientation is ambiguous: we often cannot tell whether a person is gay or straight. Research shows, however, that sexual orientation can be perceived at levels greater than chance guessing (50% correct; Tskhay & Rule, 2012) when presenting experimental participants with short videos (Ambady, Hallahan, & Conner, 1999), auditory recordings (Linville, 1998), and still photographs of faces (Rule, Ambady, Adams, & Macrae, 2007). Furthermore, sexual orientation can be accurately perceived even when targets attempt to conceal their gay or lesbian identity (Sylva, Rieger, Linsenmeier, & Bailey, 2010). The perception of sexual orientation is similar to that of obvious groups: it is rapidly inferred (i.e., 50 milliseconds; Rule & Ambady, 2008a) and relies on specific cues (Gaudio, 1994; Rule, Ambady, Adams, & Macrae, 2008). Relevant cues identified to date include gait and the degree of sway in one s walk (Johnson, Gill, Reichman, & Tanissary, 2007); facial features, such as the mouth, eyes, and hairstyle (Rule et al., 2008); and pronunciation of vowel sounds (Rendall, Vasey, & McKenzie, 2008). Thus, it is possible that people have a general idea of what a gay person should look like and may even have a prototype, or mental representation, of what it is to look gay. Congruence between the prototype and actual facial features could facilitate

8 2 categorization at greater than chance levels. However, questions about whether such a prototype exists for sexual orientation and how people come to have a mental representation of sexual orientation remain unanswered. 1.1 Mental Representations of Group Memberships Abstract-features models of face categorization have been used to explain the processes involved in face perception. Abstract features are common patterns of features shared by a particular category of stimuli (Wallis, Siebeck, Swann, Balnz, & Bülthoff, 2008). Earlier models supposed that the mind created only one norm for all faces (Valentine, 1991). More recent evidence, however, suggests that the mind represents separate categories of social groups, such as distinct mental representations of men and women (Baudouin & Brochard, 2011). For example, according to the Abstract-features model, even though every woman one encounters may be uniquely distinct from all other women in her appearance, one s mind will recognize a shared pattern of features relating her to the larger population of women. This pattern of features will be synthesized to construct the mental representation of an average woman that describes the entire category. Furthermore, the mind will track the idiosyncratic deviations from the pattern and use these deviations to identify any specific woman or class of women (e.g., mother, Starbucks barista). Categorization, therefore, occurs after we implicitly and mentally compute the degree of similarity between a particular exemplar and the prototype composed of abstract features. Researchers have examined the possibility of prototypical facial representations within a face-recognition framework. Valentine (1991) proposed a model of facial representation and encoding based on distinctiveness, such that people have a tendency to construct facial prototypes by mentally averaging the features of many faces they encounter throughout life. According to this framework, prototypical faces will seem more familiar to observers. Homa, Smith, Macak, Johovich, and Osino (2001), for example, found that people miscategorized the prototypical face characterizing a set of faces (i.e., the average of the faces presented during a familiarization task) as familiar following training that did not include this face. Thus, categorization is based on visual input and comparison of that input against a prototype stored in memory (Bruce & Humphreys, 1994). Since categorical perception is a perceptual process and

9 3 could be prototype-based, it is possible that the salience of the prototype could be enhanced following some degree of training. Consistent with this framework, Levin and Beale (2000) were able to induce categorical perception of human faces following mere exposure to faces on extreme ends of a morphed unfamiliar-face continuum. They suggest that familiarization with a continuum allows for the creation of categories due to the summed-similarity rule recruited by the mind to identify distinct patterns in the environment (Valentine & Ferrara, 1991). Similarly, Quinn, Palmer, and Slater (1999) showed that people could be trained to discriminate the sex of cats based on facial features following training. It seems that similar configurations of abstract features facilitate the creation of prototypical representations, resulting in categorization (Diamond & Carey, 1986). More recent research in face perception has further demonstrated that facial prototypes can be reconstructed using reverse correlation techniques. In the reverse correlation paradigm, the participant is asked to select one of two degraded images as belonging to a certain category in each of a series of trials. On each trial, an image of a face is presented with a noise pattern added, whereas the tandem image has had the same noise pattern subtracted. Averaging the noise patterns selected by a participant to represent the target category produces a new noise pattern. This noise pattern is then applied to the original underlying face (which is identical across all trials and stimuli but appears different due to the imposed noise) to reveal the participant s mental representation of the selected group. Dotsch and colleagues have used reverse correlation to construct visualized mental representations of Moroccan and Chinese men (Dotsch, Wigboldus, Langner, and van Knippenberg, 2008), criminals and homosexuals (Dotsch, Wigboldus, & van Knippenberg, 2011), and even prototypes for people s personality traits (Dotsch & Todorov, 2011). Earlier work used reverse correlation to reconstruct specific identities. Mangini and Biederman (2004) asked participants to categorize random visual noise patterns resembling a face as either Tom Cruise or John Travolta. When averaging across the trials in which participants indicated that the noise looked more like Tom Cruise, a prototypical image of Tom Cruise emerged; and likewise for John Travolta. Hence, the images were reconstructed spontaneously and without any explicit exposure to photographs of the targets. These results suggest that individuals create prototypical representations of groups, traits, and individuals in order to simplify the categorization process. It is possible that people have mental

10 4 representations of ambiguous groups, such as sexual orientation, that is learned through continual interaction with the environment. 1.2 Perceptual Learning and the Ecological Approach Perceptual learning is typically defined as a relative change in perceptual abilities following training (Gibson, 1963). This type of learning has been recognized to be orientationspecific and localized to the brain s primary visual processing areas (V1; Florentini & Berardi, 1980). Furthermore, the process is implicit (Gilbert, 1998) and task specific (Fahle & Morgan, 1996). Seitz and Watanabe (2005) highlighted that the discriminability of features and patterns could be enhanced in adult humans with sufficient training. Moreover, other researchers found that experts (e.g., radiologists) have a fine-tuned ability to discriminate between two different patterns that may be indiscriminable to non-experts (Sowden, Davies, & Roling, 2000) and that two indiscriminable odors become perceptually different after aversive conditioning (Li, Howard, Parrish, & Gottfried, 2008). Following similar logic, Stevenage (1998) found that people could be trained to discriminate between the faces of identical twins. Perceptual learning, therefore, extends beyond the processes in the primary visual cortex to higher-order processing areas of the brain. More recently, researchers have explained perceptual learning in terms of the enhanced ability to perceive a stimulus within external noise i.e., the information from the context that is outside of the perceptual experience of an observed target (Hurlbert, 2000). This ability is not drawn from the signal to noise ratio, however. Instead, it is believed to be based in enhanced neuronal tuning to an ecologically-relevant stimulus (Gilbert, Li, & Piech, 2009) and to occur at various levels of processing. Learning, moreover, implies the occurrence of an interaction between an individual and the environment. According to the ecological approach to perception, the relevant patterns in the environment should be learned by the mind to achieve optimal and efficient functioning (Zebrowitz & Montepare, 2006). For example, recent evidence shows that the people can quickly identify relevant signals to the symptoms of disease (Schaller, Miller, Gervais, Yager & Chen, 2010) and that women become more accurate at judging sexual orientation at the peak of ovulation (Rule, Rosen, Slepian, & Ambady, 2011). This notion is consistent with the normbased perspective of prototypical face recognition, since prototypes are environmentally-relevant and thought to simplify the processes involved in recognition and categorization (Valentine,

11 5 1991). Thus, increasing the relevance of group membership could enhance the accuracy of identifying sexual orientation. Furthermore, given that the mind appears to learn abstract-patterns in the environment (Wallis et al., 2008), it is possible that increasing salience by ascribing relevance to those patterns could result in attunement to stimulus properties and subsequent enhancement in categorization accuracy. 1.3 Current Work Since the categorical perception of human faces could be prototype-based (Valentine, 1991) and the mind may learn prototypical representations through interactions with environment, the accuracy of categorizing sexual orientation from faces could potentially be enhanced following training. This possibility is particularly intriguing because sexual orientation is an ambiguous and naturally-occurring group that is perceived largely unconsciously (e.g., Rule et al., 2008). Furthermore, the accuracy of identifying members of perceptually ambiguous groups is not perfect (approximately 64% correct; Tskhay & Rule, 2012) and it is unclear whether the categorization of sexual orientation relies on separate prototypes for gay and straight identities, or if it is based solely on the mental representation of sex and the stereotype of gender inversion (Freeman, Johnson, Ambady, & Rule, 2010). The perception of sexual orientation is also consistent with the framework of face representation as occurring through prototypes (Valentine, 1991), as gay people are more accurate than straight people at distinguishing between gay and straight targets in a set of static faces (Rule et al., 2007) possibly due to increased exposure to the gay minority group or the increased ecological relevance of such categorization, both of which may promote prototype formation. Thus, increasing the salience of abstract features that cue sexual orientation should lead to increased accuracy in categorizing both computer-generated and veridical faces. Furthermore, this process should be driven by perceptual attunement to the patterns of features that are relevant to categorizing sexual orientation in the real world.

12 6 Chapter 2 Study 1 Before I could examine whether the accuracy of identifying perceptually ambiguous groups could be improved, I needed to confirm that the perception of sexual orientation is not perfect. Thus, I compared the discriminability of the defined category membership (i.e., sex) to that of the ambiguous category (i.e., sexual orientation). 2.1 Method Stimuli Thirty-three female and 33 male faces from the Karolinska database (Lundqvist, Flykt, & Öhman, 1998) posing neutral facial expressions and direct frontal eye gaze were morphed to generate female and male prototypes, respectively. Similarly, 52 photographs (n = 26 gay) of white male targets from an in-house database were used for construction of gay and straight male prototypes; these individuals were also posing a neutral expression and looking directly into the camera when photographed. The photos were pre-rated for perceived sexual orientation (i.e., expressed as the proportion of participants identifying any particular target as gay) by an independent group of judges (N = 17). The eight photographs categorized most often as gay (M = 69%, SD = 12%) and the eight photographs categorized most often as straight (M = 9%, SD = 5%) were selected for use in the study. Specifically, the faces of the men who were most often categorized as gay were morphed to create a prototype of a gay male. Similarly, the eight photographs of the men most often categorized as straight were morphed to produce a prototype of a straight male. The prototypes were based on consensual judgments and therefore represent the stereotypical faces of a gay and straight person, regardless of the targets actual sexual orientations (see Figure 1). Prior to aggregation, all of the faces were removed from their original background, converted to grayscale, cropped to the extremes of the head, and standardized in size. I morphed the male and female prototypes together using Morpheus Photo Mixer v3.16 software to create ninety-nine images resulting in a 101-image continuum, wherein each image received (a)% contribution from the female prototype and (100-a)% contribution from the male prototype. A continuum for the gay and straight faces was generated analogously.

13 Procedure Two hundred participants recruited online from Amazon s Mechanical Turk categorized the generated stimuli from the male/female continuum as male or female (n = 102) or from the gay/straight continuum as gay or straight (n = 98) in a two-choice paradigm. 2.2 Results and Discussion I estimated a psychometric logistic function for each sample: I used stimulus intensity (contribution from the male or straight prototype, respectively) as the independent variable, and the percentage of participants perceiving the face as male (straight) as the dependent variable. Furthermore, I applied the same function for each participant within the samples to estimate the threshold and slope parameters, which I used to calculate the averages and to conduct comparisons between the two samples. Thus, I divided the intensity levels into 11 bins containing nine adjacent observations, where lower numbers represented lower contribution from the male or straight prototype and higher numbers represented greater contribution from the prototype. I calculated the probability of responding male or straight within each bin for each participant. This procedure allowed for the estimation of the individual thresholds and slopes Male/Female Distinction The threshold, or point at which 50% of respondents identified the stimulus as male and the other 50% categorized the same stimulus as female, was the face with 49.85% contribution from the male prototype in the sex-categorization task. Furthermore, the slope of the function reached its maximum value [slope = 0.02] at this point, suggesting that the face located at this point along the continuum was most ambiguous and marked the most rapid change in the perception of the two categories (see Figure 2). I found similar results for the individual-level data. The threshold, or the intensity bin at which any given participant categorized the face as belonging to one of the two categories 50% of the time, averaged 6.05 (SD = 1.40); suggesting that an average participant changed his or her categorization from one category to the next when the faces from the 6 th bin were presented. This value suggests that half of the participants categorized the 54% male prototype as male. I further examined whether this estimate was significantly different from the criterion value of 5.50

14 8 (midpoint of the intensity continuum as defined by the bin number) by generating a 95% percent confidence interval around the estimates obtained from bootstrapping with 5000 resamples. The confidence interval [5.77, 6.32] did not include the criterion, suggesting that the threshold estimate was significantly and reliably different from what was theoretically expected. Similarly, the slope estimate [M = 0.51, SD =.80] was also reliable, 95% CI [.39,.69]. This suggests that participants were attuned to the features that are prototypically male and female and that the participants changed their categorization rapidly, when presented with an ambiguous face with equal contribution from the male and female prototypes Gay/Straight Distinction When looking at the aggregated data in the gay/straight categorization condition, I found a threshold of and a slope of.004 (see Figure 2). These estimates are substantially smaller than those observed in the male and female categorization. The threshold observed here indicates that the face had to contain approximately only a 15% contribution from the straight prototype to be perceived by half of the observers as straight, suggesting that the majority of the participants showed a bias towards categorizing faces as straight. The slope further suggests that categorization did not change rapidly. This result could be misleading, however, because (a) not all of the participants had variance in their data, and (b) it examines the aggregated sample, rather than individual data points. Thus, I examined the individual-level data to obtain more representative estimates. I eliminated 23 participants from the analyses because they did not display any variance in their categorization: one participant categorized all of the targets as gay, and the other 22 categorized all of the targets as straight. Elimination of these participants likely affected the estimates substantially, as it removes the most biased perceptions from the sample. I fit a logistic psychometric function similar to the procedure described for the male/female categorization. The threshold for categorization was 5.46 (SD = 2.97), which roughly corresponds to the 49% mark along the bin continuum; that is, participants responded straight 50% of the time when they saw stimuli containing at least 49% of the straight prototype. This estimate was reliable at α =.05, as the 95% confidence interval [4.80, 6.14] did include the criterion value. I observed a very similar result for the slope estimate, M =.08, SD =.11, 95% CI [0.06, 0.11]. These results suggest that the participants were attuned to distinctive features when an ambiguous face (approximately 50%

15 9 straight and 50% gay) was presented, even though their response did not change rapidly between the two categories. As such, the threshold suggests that the participants recognized the presence of both prototypes in the morphed faces and their responses were influenced by this perception. Furthermore, the slope suggests that the perception from gay to straight prototype was not immediate and rapid, but instead was rather gradual Comparison between Samples To estimate how the categorization of obvious group membership differs from the categorization of ambiguous group membership, I examined the differences in the psychometric estimates of the two samples. Both judgment types were on the same scale, which allowed for statistical comparisons of the sensitivity parameters. After comparing the estimates, I found that participants thresholds did not significantly differ between the two categorization types, t(175) = 1.59, p =.11, r =.12. This was likely because of the dichotomous nature of the categorization tasks and relative ambiguity of the face containing approximately equal contributions from the prototypes of the respective continua. Interestingly, I found an effect for the slope, with sex categorization having a greater slope at the threshold level than sexual orientation categorization did, t(175) = 5.22, p <.001, r =.37. The difference in slope suggests that perception changed significantly more rapidly when participants were asked to discriminate the defined (sex) versus ambiguous (sexual orientation) category. Thus, as expected, a minor additional contribution from either end of the defined category continuum to the 50/50 face rapidly changed the perception towards that extreme. This effect was not expected in the ambiguous group categorization, however, because the extremes are not as clearly represented Conclusions In Study 1, I demonstrated that the perception of an ambiguous group membership was perceptually different from the perception of a defined category membership. Furthermore, because categorical thinking is a natural part of social development, it is possible that people could learn a new category following implicit training. Moreover, because the change in either the perceptual threshold or the slope should suggest the presence of perceptual learning, this change following training would suggest that people learn the prototype via its features. Therefore, I conducted Study 2 to estimate how well people learn the prototypes if subjected to an implicit learning paradigm.

16 10 Chapter 3 Study 2 People can discriminate between gay and straight faces with accuracy that is significantly greater than chance guessing (Rule & Ambady, 2008a). Mental prototypes of the stigmatized (gay) minority group could facilitate such accurate categorization. The goal of this study was to examine whether such a prototype exists and whether the salience of that prototype can be enhanced following implicit training. 3.1 Method Stimuli I used the same stimuli from the gay/straight continuum as in Study 1. Additionally, a 95- decibel white noise blast 100 ms in length was generated using the Audacity software for the Mac OS X system to serve as a conditioning stimulus. This noise is loud, unexpected, and aversive, 1 and therefore should elicit a threat response similar to other fear-conditioning stimuli (LeDoux, 1996). Because the perceptual system is sensitive to detecting threatening patterns in the environment (Öhman & Mineka, 2001), the pairing of the noise and the faces should increase the salience of the prototypical features of one of the categories. This would result in perceptual attunement to these patterns in all of the faces of the continuum, resulting in categorical perception Procedure Twenty-six participants sat in front of a computer monitor and were given a pair of Bose QC15 noise-cancelling headphones to wear for presentation of the auditory, conditioning stimulus. Images were randomly presented using DirectRT experimental software. The experiment consisted of three phases and employed a within-subjects design. In the first phase, participants categorized all of the 99 morph faces from the continuum and the two original 1 At the end of the experiment, I asked the participants to indicate how comfortable they were with the noise on a 7- point Likert-like scale where greater numbers indicated more aversion. A one sample t-test indicated that the participants found the noise to be significantly more aversive than the midpoint of the scale, M = 4.19, SD = 2.22, t(25) = 2.06, p =.05.

17 11 prototype faces as either gay or straight via key-press. The stimuli were preceded by a fixation cross for 250 ms and remained onscreen until the participants made their judgments. No feedback was provided during this stage. In the second stage, all of the participants were trained to discriminate between the gay and straight prototypical features. The participants viewed the top 40 faces from the extremes of the continuum: 20 gay (those with % contribution from the gay prototype) and 20 straight (those with 0 20% contribution from the gay prototype). Each image was presented twice in random order. The gay faces were preceded by a fixation cross for 250 ms. The presentation of faces was followed by a white noise blast delivered 750 ms after stimulus onset. A backward mask consisting of visual noise appeared immediately after the presentation of the white noise and remained on screen until the participant made his or her categorization. The straight stimuli were presented similarly, except that they were not accompanied by a white noise blast. Thus, the participants were exposed to each face for 850 ms in the learning stage. Participants made a dichotomous categorization, indicating whether the face presented was either gay or straight via key-press. This procedure was employed to further increase the sensitivity of the participants to prototypical patterns of features through conditioning. In the final stage of the experiment, participants classified all of the stimuli as either gay or straight once again. No feedback was provided in this part of the experiment. Thus, the data from this stage were considered as post-test measures of participant performance. 3.2 Results and Discussion Similar to the analyses in Study 1, I estimated a logistic psychometric function for each participant s data before and after the training. I also fit the same function to the aggregate scores across the participants for demonstration purposes. In aggregate, the threshold before training was 37.94%, suggesting that the face had to be only about 38% straight for 50% of the participants to categorize it as straight. The slope at this point was.005. After the training, the threshold for categorization approached the expected value (50%), threshold = 43.24, suggesting that people became more sensitive to the gay features in the faces and took these into account when making their categorizations. The slope also

18 12 increased but was still rather small, slope =.009. See Figure 3 for a visual representation of these changes. Although these data are informative and interesting, I needed to perform participant-level analyses to ensure that the threshold and slope changes had achieved statistical significance, and to get more robust response bias estimates for these parameters. Thus, I fit the same psychometric function to each participant s data aggregated into 11 intensity bins. The intensity bins thus served as an independent variable. The thresholds, or the bin where the participants categorized the face as belonging to one of the categories 50% of the time, from the pre-training assessment (M = 5.06, SD = 1.94) were significantly smaller than those obtained after training, M = 6.31, SD = 1.99, t(25) = 2.44, p =.02, r =.44. The slopes before (M =.14, SD =.38) and after (M =.15, SD =.16) the training did not significantly differ, t(25) < 1.00, p =.94, r =.01. The difference in threshold represents how much more attuned the participants became to patterns of features in the gay prototype, whereas the slope represents how much additional contribution from either extreme is necessary to cause a perceptual shift from seeing the gay versus straight category. Thus, I had anticipated the threshold change: people became more attuned to the features of the gay prototype. A non-significant difference in the slope, however, suggests that the category membership remained ambiguous: the participants still could not rely on the changes in patterned information to facilitate a more rapid shift in perceptions. Moreover, as the group I examined here is ambiguous, I expected a negligible change in the slope, which was consistent with the data.

19 13 Chapter 4 Study 3 Study 2 demonstrated that a very short amount of time (i.e., 5-7 minutes) is necessary to produce categorical perceptions along the stimulus continuum. The utility of the results, however, could be questioned as long as they apply only to artificially-generated face morphs. In Study 3, I therefore attempted to generalize the implicit category learning to veridical, real-world faces. 4.1 Method Participants Two-hundred four undergraduates participated for partial course credit or monetary compensation Five participants did not finish the study due to computer failure. Another 42 participants were eliminated because they showed no variance either throughout the experiment or in one of the three blocks. Last, 44 participants reported decreasing the volume on the headphones, or taking the headphones off entirely, to avoid the conditioning stimulus these data were not used in the analyses because this group of participants did not engage in the training manipulation. Therefore, I analyzed the data from the remaining 113 participants Stimuli The gay and straight continuum was the same as in Studies 1 and 2. Additionally, photographs of 90 gay and 90 straight men from a previously-validated database were employed for the testing phases of the experiment (Rule, 2011; Rule & Ambady, 2008a). These images were obtained from online dating advertisements posted in various major US cities. Men in the images self-reported looking for either a male or female romantic partner or explicitly selfidentified as gay or straight. The images were available in the public domain and were anonymous. All men were Caucasian and were 18 to 30 years old. Headshot, portrait-style photos were downloaded. Only images presenting a directly-oriented face that was free of facial hair, spectacles, and jewelry were selected. The faces were converted to grayscale, removed from their original backgrounds, cropped to the limits of the head, and standardized in size. Targets

20 14 actual sexual orientations were never disclosed to participants and no photos were obtained from the local geographic area Procedure The study consisted of three parts. First, the base rate for the accuracy of categorization for actual faces was determined by asking participants to categorize 90 (45 gay, 45 straight) of the real faces as either gay or straight via key-press at a self-paced rate. Second, they were trained in prototype application using the morphed faces, as described in Study 2. Third, participants categorized the remaining 90 real faces as either gay or straight. The actual faces were therefore presented in the first and last stages, whereas stimuli from the gay-straight continuum were used for conditioning. The actual faces were randomized between and within the two (pre-training and post-training) blocks such that no participant saw any face more than once. 4.2 Results and Discussion The data were examined using analyses based on signal detection theory (Macmillan & Creelman, 2005). The images of gay men correctly categorized as gay were considered hits, and the images of straight men misperceived as gay were counted as false alarms. I calculated a nonparametric index of discriminability (A ) to estimate accuracy and the correspondent metric B to represent response bias for the judgments made before and after the training. For the pre-training judgments, I found that accuracy (M =.58, SD =.09) was significantly greater than chance guessing (.50), t(112) = 8.99, p <.001, r =.65. Furthermore, the response bias (M =.02, SD =.08) was also significantly greater than chance [t(112) = 3.90, p <.001, r =.35], such that participants generally categorized targets as straight. Repeating the same analyses for the post-training ratings, I again found that accuracy (M =.60, SD =.09) was significantly greater than chance, t(112) = 11.61, p <.001, r =.74. Response bias (M =.003, SD =.15), however, was not different from zero, t(112) < 1.00, p =.86, r =.02. The decrease in response bias is an expected product of a negative interaction with the environment, as participants should generalize their responses more after encountering threatening stimuli because avoiding the threatening information in this case, the aversive noise associated with the face should have adaptive value for an organism (Schaller, 2008) and will therefore be less

21 15 biased to categorize people as straight. It is important to note that the accuracy measure A is a response-bias-free estimate (Macmillan & Creelman, 2005) and the two values are independent of each other. Assessing the effect of training, I found that the participants became more accurate following the conditioning, t(112) = 2.00, p =.048, r =.18 (Figure 4). Similarly, the response bias became significantly lower after training, as expected, t(112) = 2.00, p =.047, r =.18. I performed an additional analysis examining how accuracy levels prior to training predicted improvement after training; thus, I correlated each participant s pre-training accuracy score with his or her improvement score (i.e., difference between the pre-training A and post-training A ). I found that pre-training scores strongly, negatively, and significantly correlated with improvement, r(111) = -.66, p <.001. This correlation suggests that people who started out displaying greater accuracy generally tended to score lower after the training, and those who showed lower accuracy before training tended to improve following training. It is expected that people with lower accuracy should improve as they become attuned to new sets of features. On the other hand, it is also sensible that people who had a more defined representation of the category (people who scored higher in the beginning) might score lower following training, as they may develop a competing representation of sexual orientation. Nevertheless, these results are particularly interesting because, on average, the participants improved their accuracy with actual faces after being exposed to the stereotypical gay and straight face prototypes. This suggests that the stereotypical features in the prototype may have some ecological validity, and are useful when applied to person categorization.

22 16 Chapter 5 Study 4 While Studies 2 and 3 showed that the perception of sexual orientation is a malleable, feature-dependent process and that the accuracy of identifying sexual orientation can be improved following just five minutes of training, it is still unclear whether this improvement was indeed due to the change in perceptual parameters. Study 4, therefore, sought to shed some light on this question by directly examining the link between the threshold and accuracy estimates before and after the training. 5.1 Method Procedure Thirty-two participants were seated individually in front of a computer. In the first part of the experiment, the participants categorized both the 101 morphed faces from the gay-straight continuum and 90 (45 gay, 45 straight) of the actual faces as gay or straight in separate counterbalanced blocks of randomly-presented trials. Immediately after, they proceeded to the training stage described in Studies 2 and 3. After the training, the participants again categorized both the morphed faces from the continuum and the remaining 90 actual faces. The actual faces were divided randomly between the two (pre-training and post-training) blocks. 5.2 Results and Discussion I first examined the participants accuracy in the blocks with actual faces both before and after the training. The participants accuracy for identifying gay and straight people after the training (M =.62, SD =.11) was significantly improved over the accuracy assessed before training, M =.57, SD =.09, t(31) = 2.56, p <.015, r =.42. Moreover, accuracy both before [t(31) = 4.09, p <.001, r =.59] and after [t(31) = 6.30, p <.001, r =.75] training was significantly greater than chance guessing. Participants were significantly more likely to categorize targets as straight before the training [M B =.05, SD =.10, t(31) = 2.95, p =.005, r =.45] but not after the training [M B =.04, SD =.25, t(31) < 1.00, p =.42, r =.14]. However, a direct comparison of response bias before and after training did not show a significant difference: t(31) < 1.00, p =.64, r =.08.

23 17 I next calculated the thresholds for each participant individually before and after the training. To get a more sensitive measure of the relationship of the thresholds, slopes, and accuracy, I correlated the threshold and slope estimates with participants accuracy before and after the training. I calculated 5000 bootstrap resamples to ensure that the estimates were reliable and to assess the significance of the relationships (i.e., 95% confidence intervals containing zero were considered to not significantly differ from chance). Before the training, the relationship between participants thresholds and their accuracy was not different from zero, r(30) =.00, 95% CI [-.46,.34], suggesting that people were not attuned to the features in the prototype that could signal sexual orientation and therefore did not utilize this information. After training, however, this relationship became negative, strong, and reached statistical significance, r(30) = -.33, 95% CI [-.59, -.04], suggesting that participants attunement to patterns of features was related to their accuracy after the training. Of main interest to the current study was the comparison of this relationship before and after the training. To examine this, I conducted a meta-analytic comparison of the correlation coefficients for the relationship of thresholds and accuracies before and after the training. Thus I compared the relationship between sensitivity to prototypes and the accuracy estimated from the judgments of actual faces. Meta-analytic comparison is an appropriate analytic technique for comparing the magnitudes of the two effect sizes (correlation coefficients). This analysis revealed that the participants accurate judgments after the training were significantly more related to the threshold estimates than in the pre-test, Z = 2.55, p =.005. It is important to note that the relationship was negative, such that people who had a greater threshold (i.e., less sensitivity to features contributed by the gay prototype) also tended to have lower accuracy following the training. Conducting the same analyses with the slope, I found the expected pattern: before the training, the relationship between the slope and accuracy was positive and significant, r =.32, 95% CI [.09,.60]. This relationship shows that people generally relied on minor changes in the information from either extreme to make accurate categorizations. Furthermore, it is possible that the participants used the most salient cues (i.e., hairstyle), rather than cues from the faces internal features, to infer group membership when judging the real faces. However, the slope was unrelated to accuracy after the training, r =.02, 95% CI [-.30,.39]. Meta-analytic comparisons showed that this relationship underwent a significant decrease before versus after training, Z = 1.99, p =.02. Again, this complements the effect for the slope

24 18 found in Study 2, where slope change was not observed: participants, realizing that the salient features like hair are no longer available, possibly ignored this information and relied on the information to which their minds became more sensitive (i.e., prototypical facial cues). Furthermore, because the quick perceptual change from one extreme to the next was not associated with threat, this information became irrelevant for categorical perception.

25 19 Chapter 6 General Discussion The categorization of sexual orientation appears to be malleable and can be learned over a short period of time without conscious effort. In my first study, I demonstrated that the perception of sexual orientation is indeed ambiguous, as compared to a defined category (i.e., sex). Next, I showed that implicit learning with threat-conditioning could attune people to detect features based on the gay and straight prototypes, which subsequently resulted in greater category discriminability. Furthermore, I demonstrated that training with prototypes (stereotypical representations of gay and straight men) affects the accuracy of categorizing men s faces as gay and straight, suggesting that stereotypes about the appearance of gay men could hold a kernel of truth (Berry, 1990). Last, in Study 4, I demonstrated that categorization accuracy became related to the threshold following training, suggesting that perceptual attunement may serve as the mechanism responsible for the increased accuracy of categorization. These results are important because they show that the perception of sexual orientation could be prototype-driven. These studies also demonstrate that prototype application could be enhanced, and consequently allow for a better understanding of why some people may be better at categorizing ambiguous group membership than others. Additionally, the study is the first to my knowledge to apply the principles of perceptual training to the categorical perception of ambiguous groups within a norm-based face perception framework (Valentine, 1991). Furthermore, these results provide additional support for Rule et al. s (2007) findings that both gay and straight people were better remembered by their ingroup members. The enhanced memory for ingroup members could be due to a more simplistic encoding and retrieval of the deviations from the prototype that result from increased familiarity with other members of the ingroup. Furthermore, some stigmatized and ambiguous identities may be ill-defined due to lack of exposure to exemplar targets. It is important to note that the category straight is also illdefined due to a large amount of variation between people within a given category, and because the generalized majority group (i.e., straight) could be defined only through opposition to the minority group (i.e., gay).

26 Theoretical Contributions The current work makes several important theoretical contributions. First, it complements the work of Rule et al. (2008) by demonstrating that the features people use to make accurate categorizations could be prototype- and stereotype-based. Second, these studies show that there may be a kernel of truth (Berry, 1990) to the stereotypes about the appearance of gay men, since participants in our sample were able to improve their categorization accuracy following training with stereotype-derived prototypes. Third, the studies support the prediction derived from the ecological approach to social perception: ascribing relevance to social stimuli results in prototype formation and increased attunement to relevant features. Last, and potentially most important, the current work provides insight to categorical perception, more generally. Specifically, by examining the perception of ambiguous groups, researchers may be able to acquire information about the processes that may be involved in the formation of obvious categories, such as sex and race. Thus, the current work suggests that the mind could have developed prototypes for obvious group members through a continual integration of reinforcement signals with physical cues communicated by the categories. Furthermore, the current set of studies demonstrates that categorical perception may be achieved through implicit learning of signals from the environment, thus suggesting that categorization processes could be dependent upon and relevant to self-preservation and survival. Future research would be encouraged to examine the prototypical category representations in neural centers associated with face identification and recognition, such as the fusiform face area (Kanwisher & Moscovitch, 2000; Moscovitch, Winocur, & Behrmann, 1997; Tong, Nakayama, Moscovitch, Weinrib, & Kanwisher, 2000) and regions involved in higherorder processing of social group membership (e.g., medial prefrontal cortex; Rule & Ambady, 2008b). Furthermore, because I show that the categorization of ambiguous group memberships can be trained, the formation and understanding of social categories, in general, could benefit from further investigations using neuroimaging. Sexual orientation is particularly useful in studying the process of categorization, as it is ill-defined, has various levels of expression, and is not perceived with perfect accuracy; yet is still a real, perceptible, and naturally-occurring social distinction. Thus, studying categorization processes through the lens of sexual orientation allows for understanding of the categorization process, more generally.

27 21 Despite the benefits that may be culled from this work, these studies were also limited in several ways. The prototypes constructed are likely not representative of the general population of gay and straight faces, and are therefore rather congruent with stereotypes that people hold about the target groups. Future studies may want to examine how the prototypes constructed from actual gay and straight faces affect the learning of categories. Nevertheless, the results show that the learning of implicit stereotypical representations influences the perception of actual targets and further increases the discriminability of category members, thus suggesting that these stereotypes could be somewhat reflective of actual physical properties. Second, perceptual learning has been shown to be largely dependent on the time interval between the training and task. Some studies show almost immediate improvement in perceptual discrimination (e.g., distinguishing between two scents; Li et al., 2008), whereas more training may be necessary to encode more complex distinctions, such as those in the face. In my studies, I found small improvements after just 5 minutes of training. Future studies could examine whether improvement is greater following more intense exposure to training (30 minutes), and whether the effect persists for longer durations (e.g., days or weeks). 6.2 Conclusions It is important to examine stereotype- and prototype-activation within identification and categorization because these processes may allow for a deeper understanding of how the mind perceives other people in the environment, more generally. The current studies were the first to examine the improvement of categorizing ambiguous human targets through the processes of perceptual learning. I found that categorization accuracy could be improved for artificiallygenerated and veridical faces. Furthermore, I demonstrated that perceptual attunement to patterns of features following environmental conditioning is responsible for improvement in categorization accuracy. These results extend theoretical arguments regarding categorical perception by demonstrating that the mind may attune to relevant patterns in the environment to create categorical social vision (Adams, Ambady, Nakayama, & Shimojo, 2010). Indeed, the mind is an incredibly complex and efficient mechanism that may adaptively create categories when doing so is relevant to protection and survival. The present work provides a step forward in better understanding this process.

28 22 References Adams, R. B., Jr., Ambady, N., Nakayama, K., & Shimojo, S. (2010). The science of social vision. New York: Oxford University Press. Ambady, N., Hallahan, M., & Conner, B. (1999). Accuracy of judgments of sexual orientation from thin slices of behavior. Journal of Personality and Social Psychology, 77, Baudouin, J.-Y. & Brochard, R. (2011). Gender-based prototype formation in face recognition. Journal of Experimental Psychology, 37, Berry, D. S. (1990). Taking people at face value: Evidence for the kernel of truth hypothesis. Social Cognition, 8, Berger, G., Hank, L., Rauzi, T., & Simkins, L. (1987). Detection of sexual orientation by heterosexuals and homosexuals. Journal of Homosexuality, 13, Bodenhausen, G. V. & Todd, A. R. (2010). Automatic aspects of judgment and decision making. In B. Gawronski & B. K. Payne (Eds.), Handbook of Implicit Social Cognition (pp ). New York, NY: The Guilford Press. Bruce, V. & Humphreys, G. W. (1994). Recognizing objects and faces. Visual Cognition, 1, Diamond, R. & Carey, S. (1986). Why faces are and are not special: An effect of expertise. Journal of Experimental Psychology, 115, Dotsch, R. & Todorov, A. (2011). Reverse correlating social face perception. Social Psychological and Personality Science. DOI: / Dotsch, R., Wigboldus, D. H. J., & van Knippenberg, A. (2011). Biased allocation of faces to social categories. Journal of Personality and Social Psychology, 100, Dotsch, R., Wigboldus, D. H. J., Langner, O., & van Knippenberg, A. (2008). Ethnic out-group faces are biased in the prejudiced mind. Psychological Science, 19, Fahle. M. & Morgan, M. (1996). No transfer of perceptual learning between similar stimuli in the same retinal position. Current Biology, 6, Florentini, A. & Berardi, N. (1980). Perceptual learning specific for orientation and spatial frequency. Nature, 287, Freeman, J. B., Johnson, K. L., Ambady, N., & Rule, N. O. (2010). Sexual orientation perception involves gendered facial cues. Personality and Social Psychology Bulletin, 36, Gaudio, R. P. (1994). Sounding gay: Pitch properties in the speech of gay and straight men. American Speech, 69,

29 23 Gibson, E. J. (1963). Perceptual learning. Annual Review of Psychology, 14, Gilbert, C. D. (1998). Adult cortical dynamics. Psychological Reviews, 78, Gilbert, C. D., Li, W., & Piech, V. (2009). Perceptual learning and adult cortical plasticity. Journal of Physiology, 587, Homa, D., Smith, C., Macak, C., Johovich, J. & Osorio, D. (2001). Recognition of face prototype: The importance of categorical structure and degree of learning. Journal of Memory and Language, 44, Hurlbert, A. (2000). Visual perception: Learning to see through noise. Current Biology, 10, R231 R233. Johnson, K. L., Gill, S., Reichman, V., & Tassinary, L. G. (2007). Swagger, sway, and sexuality: Judging sexual orientation from body motion and morphology. Journal of Personality and Social Psychology, 93, Kanwisher, N. & Moscovitch, M. (2000). The cognitive neuroscience of face processing: An introduction. Cognitive Neuropsychology, 17, LeDoux, J. (1996). The Emotional Brain: The Mysterious Underpinnings of Emotional Life. New York: Simon and Schuster. Levin, D. T. & Beale, J. M. (2000). Categorical perception occurs in newly learned faces, otherrace faces, and inverted faces. Perception & Psychophysics, 62, Li, W., Howard, J. D., Parrish, T. B., & Gottfried, J. A. (2008). Aversive learning enhances perceptual and cortical discrimination of indiscriminable odor cues. Science, 28, Linville, S. E. (1998). Acoustic correlates of perceived versus actual sexual orientation in men s speech. Folia Phoniatrica et Logopaedica, 50, Lundqvist, D., Flykt, A., & Öhman, A. (1998). The Karolinska Directed Emotional Faces KDEF. CD ROM from Department of Clinical Neuroscience, Psychology Section, Karolinska Institutet, ISBN Macmillan, N. A. & Creelman, C. D. (2005). Detection theory: A user s guide. Mahwah, NJ: Lawrence Erlbaum Associates, Inc. Macrae, C. N., & Bodenhausen, G. V. (2000). Social cognition: Thinking categorically about others. Annual Review of Psychology, 51, Macrae, C. N., & Martin, D. (2007). A boy primed Sue: Feature-based processing and person construal. European Journal of Social Psychology, 37, Mangini, M. C. & Biederman, I. (2004). Making the ineffable explicit: Estimating the information employed for face classifications. Cognitive Science, 28,

30 24 Moscovitch, M., Winocur, G., & Behrmann, M. (1997). What is special about face recognition? Nineteen experiments on a person with visual object agnosia and dyslexia but normal face recognition. Journal of Cognitive Neuroscience, 9, Öhman, A., & Mineka, S. (2001). Fears, phobias, and preparedness: Toward an evolved module of fear and fear learning. Psychological Review, 108, Quinn, P. C., Palmer, V., & Slater, A. M. (1999). Identification of gender in domestic-cat faces with and without training: Perceptual learning of a natural categorization task. Perception, 28, Remedios, J. D., Chasteen, A. L., Rule, N. O., & Plaks, J. E. (2011). Impressions at the intersection of ambiguous and obvious social categories: Does gay + black = likable? Journal of Experimental Social Psychology, 47, Rendall, D., Vasey, P. L., & McKenzie, J. (2008). The Queen s English: An alternative biosocial hypothesis for the distinctive features of gay speech. Archives of Sexual Behavior, 37, Richeson, J. A., & Trawalter, S. (2005). On the categorization of admired and disliked exemplars of admired and disliked racial groups. Journal of Personality and Social Psychology, 89, Rule, N. O. (2011). The influence of target and perceiver race in the categorization of male sexual orientation. Perception, 40, Rule, N. O., & Ambady, N. (2008a). Brief exposures: Male sexual orientation is accurately perceived at 50 ms. Journal of Experimental Social Psychology, 44, Rule, N. O., & Ambady, N. (2008b). First Impressions: Peeking at the neural underpinnings. In N. Ambady & J. J. Skowronski (Eds.), First Impressions (pp ). New York, NY: The Guilford Press. Rule, N. O., Ambady, N, Adams, R. B. Jr., & Macrae, C. N. (2008) Accuracy and awareness in the perception and categorization of male sexual orientation. Journal of Personality and Social Psychology, 95, Rule, N. O., Ambady, N., Adams, R. B., Jr., & Macrae, C. N. (2007). Us and them: Memory advantages in perceptually ambiguous groups. Psychonomic Bulletin & Review, 14, Rule, N. O., Rosen, K. S., Slepian, M. L., & Ambady, N. (2011). Mating interest improves women s accuracy in judging male sexual orientation. Psychological Science, 22, Schaller, M. (2008) Evolutionary bases of first impressions. In N. Ambady & J. J. Skowronski (Eds.), First Impressions (pp ). New York, NY: The Guilford Press. Schaller, M., Miller, G. E., Gervais, W. M., Yager, S., & Chen, E. (2010). Mere visual

31 25 perception of other people s disease symptoms facilitates a more aggressive immune response. Psychological Science, 21, Seitz, A. & Watanabe, T. (2005). A unified model for perceptual learning. Trends in Cognitive Science, 9, Sowden, P. T., Davies, I. R., & Roling, P. (2000). Perceptual learning of the detection of features in X-ray images: A functional role for improvements in adults visual sensitivity? Journal of Experimental Psychology: Human Perception and Performance, 26, Stevenage, S. V. (1998). Which twin are you? A demonstration of induced categorical perception of identical twin faces. British Journal of Psychology, 89, Sylva, D., Rieger, G., Linsenmeier, J. A. W., & Bailey, J. M. (2010). Concealment of sexual orientation. Archives of Sexual Behavior, 39, Tong, F., Nakayama, K., Moscovitch, M., Weinrib, O., & Kanwisher, N. (2000). Response properties of the human fusiform face area. Cognitive Neuropsychology, 17, Tskhay, K. O. & Rule, N. O. (2012). Accuracy in categorizing perceptually ambiguous groups: A review and meta-analysis. Accepted for publication at Personality and Social Psychology Review. Valentine, T. (1991). A unified account of the effects of distinctiveness, inversion, and race in face recognition. Quarterly Journal of Experimental Psychology, 43A, Valentine, T. & Ferrara, A. (1991). Typicality in categorization, recognition and identification: Evidence from face recognition. British Journal of Psychology, 82, Wallis, G., Siebeck, U. E., Swann, K., Blanz, V. Bülthoff, H. H. (2008). The prototype effect revisited: Evidence for an abstract feature model of face recognition. Journal of Vision, 8, Wright, D. B., & Stroud, J. N. (2002). Age differences in lineup identification accuracy: People are better with their own age. Law & Human Behavior, 26, Zebrowitz, L. A. & Montepare, J. (2006). The Ecological Approach to Person Perception: Evolutionary Roots and Contemporary Offshoots. In M. Schaller, J.A. Simpson, & D.T. Kenrick (Eds.), Evolution and Social Psychology (pp ). New York: Psychology Press.

32 Figure 1. Gay (left) and Straight (right) consensus-based prototype face morphs generated by an independent group of participants (N = 17). 26