ادراک جمعیت در پروزوپاگنوزیا

Abstract Prosopagnosics, individuals who are impaired at recognizing single faces, often report increased difficulty when confronted with crowds. However, the discrimination of crowds has never been fully tested in the prosopagnosic population. Here we investigate whether developmental prosopagnosics can extract ensemble characteristics from groups of faces. DP and control participants viewed sets of faces varying in either identity or emotion, and were asked to estimate the average identity or emotion of each set. Face sets were displayed in two orientations (upright and inverted) to control for low-level visual features during ensemble encoding. Control participants made more accurate estimates of the mean identity and emotion when faces were upright than inverted. In all conditions, DPs performed equivalently to controls. This finding demonstrates that integration across different faces in a crowd is possible in the prosopagnosic population and appears to be intact despite their face recognition deficits. Results also demonstrate that ensemble representations are derived differently for upright and inverted faces, and the effects are not due to low-level visual information.

Introduction Every day we interact with crowds of people. Whether it is on a city bus, in a classroom, or in a business meeting, we routinely view and extract important information from groups of faces, and do so rather rapidly. Indeed, recent studies have shown that people are adept at recognizing crowd characteristics, such as average gender, identity or emotion, even when crowds are viewed so briefly that information about any specific individual is not extracted (De Fockert and Wolfenstein, 2009, Haberman and Whitney, 2007 and Haberman and Whitney, 2009). For example, as a passenger on a bus, we form a general impression of important characteristics of a crowd standing on the street corner, even if we are only able to view the crowd for a split-second as we ride by. Given the frequency with which we interact with crowds, a deficit in perceiving crowd characteristics would likely pose a hindrance in a host of social situations. Anecdotal evidence suggests that individuals with prosopagnosia, a deficit in discriminating individual faces, feel overwhelmed in crowded situations, perhaps in part due to their inability recognize familiar faces in a crowd. For example, one prosopagnosic describes his experience walking into a reception hall, “There are a lot of people there, perhaps as many as a hundred or so people. These are all people I am supposed to know, each with a supposedly unique face. My goal is to find just one specific individual. I can scan the room for hours in frustration… ( Asprin, 2011).” Another prosopagnosic expresses frustration saying, “Faces in public are just all faces to me, I don’t see them individually. This is especially [true] in crowded public areas. When I look into a crowd, most look very much alike to me (BP, 2011).” Can prosopagnosics’ discomfort with crowds be explained entirely by their deficits in perceiving single faces? Or could it reflect a more general impairment in integrating and extracting face-related information from a crowd? On the other hand, might prosopagnosics actually be better at ensemble coding because they do not perceive crowd members as distinct individuals? The perceptual characteristics of developmental prosopagnosics2 (DPs), individuals who have never fully developed the ability to recognize faces, have been increasingly studied during the last decade. However, almost all of the previous research used single faces to investigate processing in DPs. Although the study of individual face processing in DP added essential information aiding the understanding of problems related to individual face recognition, we know virtually nothing about how DPs extract information from groups of faces and whether it is normal or not. When processing crowds, typical viewers initially discount individual faces in a group and instead formulate unitized percepts that accurately describe crowd characteristics (De Fockert and Wolfenstein, 2009, Haberman and Whitney, 2007 and Haberman and Whitney, 2009). The ability to generate a gestalt percept of the crowd, independent of information derived from individual faces, can be viewed as a mechanism that compensates for the limited capacity of the visual system to process multiple items simultaneously. Redundant information across items in a scene is compressed into an average representation of the entire set, referred to as the “ensemble code” (Alvarez, 2011, Ariely, 2001 and Chong and Treisman, 2003). This average representation provides a more precise description in comparison to individual evaluations of each member of the set because noise from one individual evaluation cancels out uncorrelated noise from another individual evaluation (Alvarez, 2011). As such, it has been shown that typical viewers can accurately extract both the mean emotional expression and mean identity of the crowd, although performance is at chance when they are asked to discriminate, identify, or localize individual members of a previously seen set (De Fockert and Wolfenstein, 2009, Haberman and Whitney, 2007 and Haberman and Whitney, 2009). Previous research suggests that DPs have trouble integrating individual face features into a gestalt (Behrmann et al., 2005, de Gelder and Rouw, 2000 and Lobmaier et al., 2010), and may be generally impaired at identifying the global shape of a stimulus, showing such deficits for objects as well as faces (Avidan et al., 2011, Behrmann and Avidan, 2005, Behrmann et al., 2005, Bentin et al., 2007 and Palermo et al., 2011). For alternative findings see: Le Grand et al. (2006), Duchaine, Yovel, & Nakayama (2007), Schmalzl, Palermo, Green, Brunsdon, and Coltheart (2008) and Lee, Duchaine, Wilson, and Nakayama (2010). Ensemble coding, like other holistic processing tasks, requires the integration of features across space (Alvarez, 2011) or time (Haberman, Harp, & Whitney, 2009). If DPs have difficulty with this type of integration in general, we may expect that they will have trouble forming a unitary percept of any attribute of a crowd, not just average identity. Alternatively, it is possible that the deficits DPs experience during individual face recognition tasks will be minimized via the process of ensemble coding. As mentioned previously, ensemble coding involves canceling out “noisy” individual evaluations, thereby achieving a more precise representation of the group as a whole. Although individual face evaluations are suboptimal in DP, the averaging process inherently reduces such imprecision. This leaves open the intriguing possibility that DPs, who are impaired at individual face identification, may be able to extract the mean identity of the crowd just as well as controls. If DPs do not experience interference by individual faces in the crowd, they could potentially be better than normal perceivers at extracting ensemble information. The aim of this study was to explore whether DPs can successfully perceive ensemble characteristics of face sets, or “crowds.” In order to distinguish between deficits specific to the perception of face identity and impairment in ensemble coding in general, we measured the ability to estimate not only the average identity of upright faces, but also the average emotional expression, an attribute for which DPs typically exhibit little impairment when performing judgments on individual faces (Bentin et al., 1999, Dobel et al., 2007, Duchaine et al., 2003, Humphreys et al., 2007 and Jones and Tranel, 2001; for an different view see Palermo et al., 2011). Accordingly, we limited our group of participants to those who reported no or very little impairment in emotional processing of faces. Furthermore, we included conditions in which the face sets were inverted to control for low-level visual effects during ensemble coding.

. Results Fig. 6 and Fig. 7 show the rectified standard deviation of the error distributions for both controls and DPs during the heterogeneous and homogeneous conditions. The pattern clearly indicates that DP's performance falls well within the distribution of control performance during the heterogeneous condition and suggests that DPs can successfully perform ensemble coding on crowds of faces. Small sample t-tests further confirm that prosopagnosics’ performance is similar to control performance. There were no significant differences between prosopagnosic and control performance, except in one case, where the prosopagnosic performed better than her matched control sample (DP1 Emotion Inverted SD = 17.62). Matched control mean SD = 19.90 see Table 4 below. The rectified standard deviation of the error distribution for the individual ... Fig. 6. The rectified standard deviation of the error distribution for the individual participants during the heterogeneous condition. Again, each DP is shown as a single triangle in a given color with their controls shown in the same color as the DP to which they were matched. In contrast to their performance on standardized face tests, DP performance in the experimental tasks is scattered across the range of control performance, indicating that DPs successfully ensemble code crowds of faces. *In this condition, this participant was exposed to the display for 3000 ms. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of the article.) Figure options The rectified standard deviation of the error distribution for the individual ... Fig. 7. The rectified standard deviation of the error distribution for the individual participants during the homogeneous condition. *In this condition, this participant was exposed to the display for 3000 ms. Figure options Table 4. The scores from small sample t-tests for each prosopagnosic compared to their matched control group within each condition of the ensemble coding task. There are no significant differences between controls and prosopagnosics except in one case, where the prosopagnosic performed better than her matched control group. DP1 vs. Controls DP2 vs. Controls DP3 vs. Controls DP4 vs. Controls Emotion Upright 1.46(4), p = .22 −1.26(4), p = .28 −.20(4), p = .85 .88(4), p = .43 Emotion Inverted −2.90(4), p = .04 * .01(4), p = .99 −.79(4), p = .47 −2.00(4), p = .11 Identity Upright .28(4), p = .79 −.98(4), p = .38 1.61(4), p = .18 −.44(4), p = .68 Identity Inverted .31(4), p = .77 −.27(4), p = .79 .31(4), p = .77 −.01(4), p = .99 * The prosopagnosic performed better than their matched control group. Table options We further tested whether individual face recognition abilities were correlated with ensemble coding performance by conducting a non-parametric, bi-variate correlational analysis between the performance on the Berkeley Famous Face Test and performance on the ensemble coding tests. We examined controls and prosopagnosics as one group because this combined data set best exhibits variance of performance on the Berkeley Famous Face Test. These scores were not correlated with performance on the ensemble coding tasks during any condition (Emotion-Upright: r(24) = −.056, p > .79; Emotion-Inverted: r(24) = .173, p > .42; Identity-Upright: r(24) = −.150, p > 48; Identity-Inverted: r(24) = .085, p > 69). Thus, the wide range of performance in individual face discrimination tasks was not correlated with ensemble coding scores, further dissociating individual face recognition ability from performance on the ensemble coding task. This pattern is not unique to the Berkeley Famous Face Test. All standard face tests were correlated with each other using parametric measures and no test was correlated with the ensemble coding task (see Appendix Table A1). If the ensemble coding task were unreliable, this would contribute to a lack of correlation. To minimize this potential, we conducted a Chronbach's alpha test, which confirmed that the ensemble coding task was reliable, α = .775. 5 One potential account for these results is that both controls and DPs were overwhelmed by the difficulty of the task. To compensate for task difficulty, both groups may potentially choose a face at random from the display rather than engaging in ensemble coding to extract a summary of the crowd. If this scenario were true, the participants would not be processing the group of faces as a whole, but only extracting the features from one face. To ensure that both groups were extracting mean representations and thereby engaging in some manner of integration across the faces in the display, we designed a model that simulated performance based on picking only one face in the display. For the simulation, a face was selected at random from the heterogeneous display. Next, noise was added based on each individual's standard deviation in the homogeneous condition. Finally, the program chose a face within the noise distribution and the corresponding value was subtracted from the mean of the display. By repeating this Monte Carlo procedure, we obtained a simulated error distribution and the associated standard deviation. We compared the simulated standard deviation against participants’ true performance for both control and DP groups independently. Paired t-tests revealed significant differences between simulated and non-simulated results. The simulated standard deviations for the controls were significantly higher, compared to the participants’ true performance (Emotion-Upright: t(19) = 15.82, p < .0001; Emotion-Inverted: t(19) = 11.51, p < .0001; Identity-Upright: t(19) = 10.41, p < .0001; Identity-Inverted: t(19) = 8.11, p < .0001). Similarly, the DPs simulated standard deviations were also significantly higher compared to each DP's true performance (Emotion-Upright: t(3) = 4.60, p < .02; Emotion-Inverted: t(3) = 19.22, p < .0002; Identity-Upright t(3) = 8.97, p < .01; Identity-Inverted: t(3) = 5.21, p < .01). The smaller variance in the real data compared to the simulation indicates that performance genuinely reflected ensemble coding and was not an artifact of task difficulty. Fig. 8 shows simulated vs. true performance for controls, while Fig. 9 shows simulated vs. true performance for DPs. Controls true performance (grey) compared to the simulated performance (black) ... Fig. 8. Controls true performance (grey) compared to the simulated performance (black) for individual participants. Controls performed significantly better than the simulation predicts if judgments were based on picking one face at random from the display. This comparison confirms that controls participants engaged in some sort of ensemble coding. Figure options Prosopagnosics true performance (grey) compared to the simulated performance ... Fig. 9. Prosopagnosics true performance (grey) compared to the simulated performance (black) for individual participants. Controls performed significantly better than the simulation predicts if judgments were based on picking one face at random from the display. This comparison confirms that controls participants engaged in some sort of ensemble coding. Figure options To ensure that our methods were sensitive enough to detect differences across conditions, we conducted a 3 way ANOVA on the control data. The factors were Dimension (emotion, identity), Orientation (upright, inverted), and Study Display Condition (heterogeneous, homogenous). The ANOVA revealed that the SD was significantly larger with inverted compared to upright faces [F(1,19) = 15.08 p < .001, ηp2 = .443]. The better performance with upright than inverted faces is consistent with reliance on holistic processing in the upright case to achieve ensemble coding and replicates earlier results ( Haberman & Whitney, 2009). The Orientation × Dimension interaction was not significant. As expected, the ANOVA also revealed better performance for the homogeneous compared to the heterogeneous condition [F(1,19) = 57.57, p < .001, ηp2 = .752]. This is not surprising because the homogeneous task should be easier (requires only matching) than the experimental task (requires ensemble coding). All other main effects and interactions did not reach significant levels.