Postdiscrimination gradients with familiar and unfamiliar faces.
Gradient shift is suspected to have a number of counterparts in the natural environment. That is, in these cases it is possible to identify stimuli that function in a similar manner as S+ and S-, and the organism tends to respond to relatively extreme stimuli (i.e., stimuli very different from S-). Examples include the evolution of sexual dimorphism (ten Cate et al. 2006; Weary et al. 1993) and aposematic coloration (Gamberale and Tullberg 1996; Gamberale-Stille and Tullberg 1999; Yachi and Higashi 1998), the preference for supernormal stimuli and caricatures (Ghirlanda and Enquist 1999, 2003; Ramachandran and Hirstein 1999; Zimmer 2003), changes in speech patterns (Martindale 2006), and extreme responses in certain forms of mental illness (Derenne 2010; Dunsmoor et al. 2009, 2011).
Research that incorporates naturalistic stimuli into the laboratory can also shed light on the underlying causes of gradient shifts by revealing whether and how it is expressed under a wide range of conditions. For example, whereas researchers have traditionally studied stimulus generalization and gradient shift using unidimensional stimuli (Honig and Urcuioli 1981; Purtle 1973; Riley 1968), a number of researchers have more recently studied gradient shift using naturalistic images of human faces (e.g., Derenne and Breitstein 2006; Dunsmoor et al. 2009; McLaren and Mackintosh 2002; Spetch et al. 2004). The pattern of generalization emerging from studies with facial stimuli notably resembles that obtained with the more traditional procedures. This finding is consistent with the possibility that gradient shift-like phenomena occur widely in the natural environment and further suggests that a single process (or set of processes) is responsible for all the varied instances of gradient shifts and shift-like behavior. Also of theoretical importance is that gradient shift with complex and non-unidimensional stimuli is not easily explained by traditional models, and the finding of gradient shift in such cases has helped foster alternative models of stimulus generalization and gradient shift (e.g., Ghirlanda 2002; Harris 2006; McLaren and Mackintosh 2000, 2002; Saksida 1999; Yachi and Higashi 1999).
The exact form of the generalization gradient is dependent, in part, on how well participants can distinguish among the various stimuli constituting the stimulus set. For example, when the stimuli are more highly discriminable, the generalization gradient tends to have a narrower base and steeper slopes (e.g., Blough 1972; Guttman and Kalish 1956). Gradient shift more specifically is affected by the discriminability of S+ and S-, with pronounced shifts typically occurring when S+ and S- are relatively similar but not when they are relatively dissimilar (Baron 1973; Derenne 2006; Hanson 1959).
A seemingly-related finding is that gradient shift is dependent on the amount of training subjects have with S+and S-. For example, Wisniewski et al. (2010) reported that undergraduates trained to discriminate between complex sounds showed peak shift when they received 140 or 180 training trials, but not when they received either less training (60 or 100 trials) or more training (220 or 260) trials. Based on this finding, the authors suggested that "peak shift is not always an end state of learning but, rather, can be a transitional state occurring at intermediate levels of training" (p. 814). The literature on the effects of brief discrimination training on gradient shift is not large, but other studies have shown that subjects receiving extensive discrimination training show little or no shift (e.g., Galizio and Baron 1976; Terrace 1964).
The present experiment addressed the possibility that "familiarity" with the stimuli in general can alter gradient shifts in a manner consistent with enhanced training. To examine this possibility, a comparison was made in the amount of shift produced by a stimulus set that was relatively familiar to participants with one that was not by virtue of participants' pre-experimental histories. The "familiar" stimuli were images based on the participants' own face and the unfamiliar stimuli were images based on the face of a stranger. In both cases, the stimuli varied in terms of bilateral facial symmetry.
Participants were 20 undergraduates (6 male, 14 female) who were enrolled in lower-level psychology courses and compensated with extra course credit.
Apparatus and Materials
Pictures of the faces were taken with a Sony Cyber-Shot DSC-w1 5.1 megapixel digital camera. Participants completed discrimination training and the generalization test on a Gateway Profile 6 personal computer equipped with a 43.18-cm (17-in.) color monitor, keyboard, and mouse.
Participants also completed the Body Image Coping Strategies Inventory (or BICSI; Cash et al. 2005). This was used to occupy participants while a stimulus set was created from the picture of the participant's face. The BICSI is a 29-item inventory of three responses to threats or challenges to body image: appearance fixing (10 items), positive rational acceptance (11 items), and avoidance (8 items). Responses are on a 4-point Likert scale (definitely like me, mostly like me, mostly not like me, or definitely not like me). The items were presented in the same randomized sequence to all participants.
The general procedure was as follows: the participant first read a description of the study and signed an informed consent form approved by the Institutional Review Board. The participant next had his/her pictures taken by the researcher. Then, the participant completed the BICSI. While the participant completed the BICSI, the researcher used the picture of the participant to create the set of images based on the participant's own face. Finally, participants sat at the computer and underwent discrimination training and generalization testing.
The procedure for generating the face stimuli was similar to that described by Derenne (2010). Each time a set of stimuli was created, a photograph of a face was taken with a Sony Cyber-Shot DSC-w1 5.1 megapixel digital camera. The photograph was then imported into MS Paint XP and turned into a 256-color bitmap image. This procedure removed some of the detail from the face, and facilitated the creation of a "seamless" symmetrical image. Highlights on the forehead, cheeks, and chin were also removed from the image to help ensure that facial symmetry, and not variations in light and shadow between the two halves of the face, functioned as the discriminative stimulus.
To create the symmetrical image, one side of the face (including the eye, cheek, and parts of the nose, forehead, and chin) was copied, "reversed," and "pasted" onto the other side of the face. Lastly, five images intermediate to the naturally asymmetrical and symmetrical versions of the face were created using the morphing software WinMorph 3.0. For graphical display and data analysis purposes, the stimuli were numbered from Image 1 (naturally asymmetrical) to Image 7 (symmetrical). WinMorph uses a linear algorithm, which keeps constant the degree of change across successive stimuli in a set. For every stimulus set, S+ was an image midway between the naturally asymmetric and the symmetric versions of the face. Figure 1 shows an example of images created with this procedure.
One of six sets of "stranger" faces, prepared in advance, was used with each participant. The strangers were one male and five female college-student volunteers. The faces of the volunteers were judged to lack either marked bilateral symmetry or asymmetry (the same was true of the typical participant). Random assignment was used to match stranger faces with participants.
Each participant completed discrimination training and generalization testing twice: once when the images were based on his/her own face, and once when the images were based on the face of a stranger. For 10 of the 20 participants, training and testing with images based on the participant's own face preceded training and testing with images based on the stranger's face; for the other half of participants the reverse sequence was used. To determine whether gradient shifts could be produced toward either symmetrical or asymmetrical faces, the stimulus selected for S- was also varied. For each participant, the naturally asymmetrical face served as S- with one set of images, and the symmetrical face served as S- with the other set of images. Within the set of participants assigned to one of the training and testing sequences described above, which S- was used with which set of faces was counterbalanced. That is, for half of the participants, S- was the naturally asymmetrical version of the familiar faces and S- was the symmetrical version of the stranger faces; for the other half of participants, the reverse arrangement was used. Random assignment was used to match participants with the above conditions, with the caveat that each possible combination of conditions occurred equally often.
At the beginning of discrimination training the following instructions were displayed on the computer screen.
You are now about to start the experiment. At first, you will see a version of the picture that I took. This picture will be shown for 10 seconds. Study it carefully and try to remember what it looks like. Your task will be to choose this picture whenever it is shown. Next you will see two pictures presented side-by-side. One of them will be the same as the first picture, the other will be different. Choose the one that is the same as the first picture. The two pictures will be very similar to each other, and you may not be able to tell which is the same. That's OK. You will be given this choice a number of times, and after each choice you will be told whether you were correct or incorrect. If you pay close attention to the two pictures, you should be able to figure out which one is correct. After a while, you will be shown one picture at a time and you will have to indicate whether it is the same or different than the first picture. You will not be told whether your choices are right or wrong during this phase.
[FIGURE 1 OMITTED]
The participant began the experiment by using the mouse to click a button labeled "Start" located immediately below the instructions. At that time, the instructions were replaced by S+, and the message "The original image" displayed above it. After 10 s, the screen changed to show the first of 15 discrimination training trials. On each training trial, S+ and S- were shown side-by-side, along with the query "Which one is the original picture?" The participant indicated their choice by clicking on a gray button located immediately below the corresponding image (one button was labeled "Left" and the other "Right"). For the first stimulus set (whether self or stranger), the relative placement of S+ across trials was left (L), right (R), R, L, L, R, L, R, R, L, R, L, L, L, R. For the second stimulus set, the relative placement of S+ was right (R), left (L), R, L, R, R, R, L, L, L, R, R, L, R, L. Each time a choice was made, the participant was briefly shown the message "Correct" or "Incorrect," depending, respectively, on whether the participant had selected S+ or S-. The message "Please Wait" was then displayed for 10 s before the next trial began.
The generalization test immediately followed the final training trial. The generalization test was organized into five cycles of images in which each of the seven images in a stimulus set was shown once (for a total of 35 trials). The sequence of the images within a cycle was randomized. On each test trial, the message "Is this the same as the original?" appeared at the top of the screen, and two buttons marked "Yes" and "No" appeared immediately below the image. Each time a choice was made, the message "Please Wait" was displayed for 10 s before the next trial began. During the test, feedback about response accuracy was withheld.
The analysis was based on participants who made no more than four errors during discrimination training, including none on the last six trials, in both the familiar and unfamiliar face conditions. (Data collection continued until 20 participants met these criteria; 5 participants failing to meet the training criteria were replaced). The participants whose results are reported below averaged 0.95 errors during training. The analysis ignored counterbalancing in the sequence of conditions (i.e., whether participants were first trained and tested with images based on their own face or that of a stranger) and instead was designed to assess the effects of familiarity and the location of S-.
Figure 2 shows the mean number of "Yes" responses emitted in the presence of each stimulus during the generalization test as a function of stimulus familiarity and the relative position of S-. The upper panel shows the postdiscrimination gradients when S- was a symmetrical face; the lower panel shows the postdiscrimination gradients when S- was naturally asymmetrical. Visible trends in the data include gradient shifts away from S-, and differences in the gradients due to stimulus familiarity.
In line with past research (e.g., Thomas et al. 1991; Wisniewski et al. 2010), the degree of gradient shift was measured using the mean the generalization gradient. The mean was calculated as follows: each "Yes" response during the generalization test was assigned a numerical value equivalent to the image that was present at the time (e.g., 1 for each response to the naturally asymmetrical image, 4 for each response to S+, and 7 for each response to the symmetrical image); the sum of these values was divided by the total number of "Yes" responses.
Each gradient was subjected to a one-sample t-test to determine whether gradient shift occurred (i.e., the means of the gradients were compared against a criterion value of 4, the numerical value assigned to S+). This analysis indicated that gradient shift was obtained when participants responded to images based on their own face (for the naturally asymmetric S-, M=4.88, [t.sub.9]=-3.47, p<0.01; for the symmetric S-, M= 3.24, [t.sub.9]=3.22, p=0.01), but not when they responded to images of a stranger's face (for the naturally asymmetric S-, M=4.29, [t.sub.9]=-1.89, p=0.09; for the symmetric S-, M=3.26, [t.sub.9]=.66, p=0.52). A repeated measures ANOVA was conducted to determine whether the degree of shift from S- significantly differed as a function of stimulus familiarity. To control for variation in the direction of shift (i.e., the gradient shifted away from either Stimulus 1 or Stimulus 7, depending on which served as S-), we compared the absolute difference between the mean of the gradient and S+. This analysis indicated that differences in the degree of shift between the two conditions did not reach the level of statistical significance (for familiar faces, M=.82; for stranger faces, M=.52), F(1, 19)=.684, p=.418.
[FIGURE 2 OMITTED]
Stimulus generalization traditionally is assessed through visual examination of the gradients. To aid in interpretation of the data, we conducted two quantitative comparisons based on key features of the gradients. First, a comparison was made of the proportion of total "Yes" responses made to S+ (greater generalization is indicated by a lower proportion). In this regard, participants were found to make a lower proportion of "Yes" responses to S+when evaluating stranger faces (M=.20) than their own face (M=.31), F (1, 19)=4.97, p=.038. Second, a comparison was made in the slopes of the generalization gradient (greater generalization is indicated by flatter slopes). Slopes were calculated using the formula (Y2-Y1)/(X2-X1); to create a single measure of slope for each participant, the slope of responding to relatively asymmetrical stimuli (Images 1-3) was averaged with the slope of responding to relatively symmetrical stimuli (Images 5-7). In this manner, the slopes of the gradient were found to be significantly flatter when participants responded to images based on the face of a stranger (M=.44) than when they responded to images based on their own face (M=.96), F (1, 19)=6.450, p=.020.
Visual inspection of the individual generalization gradients on which Fig. 2 is based revealed that 19 of the 20 participants demonstrated gradient shift with at least one set of images. When participants examined images based on their own face, gradient shift was a typical occurrence (13 of 20 participants). For the other seven participants, responding was limited to Images 3-6 (i.e., S+ and the stimuli immediately adjacent to it). Performances were considerably more variable when participants examined images based on the face of a stranger. Gradient shifts again repeatedly occurred, but for five participants the shift occurred toward S- rather than away from it. In only two cases was responding tightly clustered around S+ (i.e., Images 3-6).
Discrimination training has been observed to produce gradient shifts with many different stimuli and species (e.g., Cheng et al. 2006; Ghirlanda and Enquist 2003; Purtle 1973; Schneider and Lickliter 2010; ten Cate et al. 2006). However, discrimination training does not inevitably produce gradient shifts. It appears necessary for S+ and S- to be relatively similar to each other for gradient shift to occur (e.g., Derenne 2006). Gradient shift also may depend on the amount of discrimination training subjects receive (e.g., Wisniewski et al. 2010), with gradient shifts occurring only after some minimum number of trials have occurred. Research also has shown that modified discrimination training can turn a task that normally gives rise to gradient shifts into one in which gradient shifts are mitigated or eliminated. This includes the use of extended discrimination training (Wisniewski et al. 2010), errorless discrimination training (Terrace 1964), and verbal labels assigned to the stimuli to make them more salient (Ban and Minke 1984; Galizio and Baron 1976).
The present research addressed the possibility that preexperimental experience with the stimuli is another key factor that determines whether or not gradient shifts occur. For example, pre-experimental "familiarity" with the stimuli might facilitate discrimination training, with the result that gradient shifts appear after less training with familiar stimuli than with unfamiliar stimuli. Conversely, past familiarity with the stimuli might mitigate discrimination training in much the same way that gradient shifts are reduced by extended or enhanced discrimination training.
In the present experiment, gradient shifts were obtained with the "familiar" faces but not with the stranger faces. In other words, the results are consistent with the possibility that familiarity can facilitate discrimination training and promote gradient shifts. Important to this conclusion is the shape of the generalization gradients obtained in the two conditions. The gradients with familiar faces had a distinct peak and steep slopes and were generally of the form reported in many studies of human and nonhuman stimulus generalization and gradient shift. The gradients with the stranger faces lacked a distinctive peak and had relatively flat slopes. The exact cause of such flat gradients is uncertain, but two possibilities include that participants in the stranger condition had greater trouble learning to discriminate S+ and S-, or that they had greater trouble remembering the characteristics of S+ and S- during the generalization test. The former possibility seems less likely as participants were replaced if they could not meet the training criteria with both sets of faces.
Although the research appears to show that preexperimental familiarity with the stimuli affects gradient shifts, a more robust test is needed to determine whether familiarity effects can act in both of the suggested manners. This could be accomplished by comparing familiar and unfamiliar stimuli across multiple levels of training, in which case one would expect that after relatively brief training, gradient shifts would occur with the familiar stimuli, but not with the unfamiliar stimuli (replicating the present results). However, with extensive training the reverse might be true: gradient shifts would not occur with the familiar stimuli, but they would with the unfamiliar stimuli. An assumption is that the relatively flat gradients and the lack of gradient shift in the stranger condition are both a result of limited training with unfamiliar stimuli. That is, it is presumed that not only would more extensive training create gradient shifts with unfamiliar stimuli, but the generalization gradients would come to resemble those obtained with less extensive training and more familiar stimuli.
Although the results and conclusions of this research are considered preliminary, the findings nevertheless suggest that further research in this area is worth pursuing. The method described here shows that it is possible to compare gradient shift with familiar and unfamiliar images while controlling for other factors known to affect gradient shift (such as the selection of S+ and S-). Furthermore, the familiarity effect was relatively robust, despite the fact that it was necessary to include in the method features that may have undermined stimulus familiarity effects. Participants' faces had to be altered to create the stimulus dimension, and the task that subjects were asked to complete likely had unfamiliar features.
Published online: 4 November 2014
Ban, P. K., & Minke, K. A. (1984). The use of generalization gradients for the study of mediational processes. The Journal of General Psychology, 110, 115-128.
Baron, A. (1973). Postdiscrimination gradients of human subjects on a tone continuum. Journal of Experimental Psychology, 101, 337-342.
Blough, P. M. (1972). Wavelength generalization and discrimination in the pigeon. Perception & Psychophysics, 12, 342-348.
Cash, T. F., Santos, M. T., & Williams, E. F. (2005). Coping with body-image threats and challenges: Validation of the Body Image Coping Strategies Inventory. Journal of Psychosomatic Research, 58, 191-199.
Cheng, K. (2002). Generalisation: Mechanistic and functional explanations. Animal Cognition, 5, 33-40.
Cheng, K., Spetch, M. L., Kelly, D. M., & Bingman, V. P. (2006). Small-scale spatial cognition in pigeons. Behavioural Processes, 72, 115-127.
Derenne, A. (2006). Effects of S+, S- separation on gradient shifts in humans. The Journal of General Psychology, 133, 163-173.
Derenne, A. (2010). Shifts in postdiscrimination gradients within a stimulus dimension based on bilateral facial symmetry. Journal of the Experimental Analysis of Behavior, 93, 485-494.
Derenne, A., & Breitstein, R. M. (2006). Gradient shifts with naturally occurring human face stimuli. The Psychological Record, 56, 365-370.
Dunsmoor, J. E., Mitroff, S. R., & LaBar, K. S. (2009). Generalization of conditioned fear along a dimension of increasing fear intensity. Learning & Memory, 16, 460-469.
Dunsmoor, J. E., Prince, S. E., Murty, V. P., Kragel, P. A., & LaBar, K. S. (2011). Neurobehavioral mechanisms of human fear generalization. Neuroimage, 55, 1878-1888.
Galizio, M., & Baron, A. (1976). Label training and auditory generalization. Learning and Motivation, 7, 591-602.
Gamberale, G., & Tullberg, B. S. (1996). Evidence for peak-shift in predator generalization among aposematic prey. Proceedings of the Royal Society of London B: Biological Sciences, 263, 1329-1334.
Gamberale-Stille, G., & Tullberg, B. S. (1999). Experienced chicks show biased avoidance of stronger signals: An experiment with natural colour variation in live aposematic prey. Evolutionary Ecology, 13, 579-589.
Ghirlanda, S. (2002). Intensity generalization: physiology and modeling of a neglected topic. Journal of Theoretical Biology, 214, 389-404.
Ghirlanda, S., & Enquist, M. (1999). The geometry of stimulus control. Animal Behaviour, 58, 695-706.
Ghirlanda, S., & Enquist, M. (2003). A century of generalization. Animal Behaviour, 66, 15-36.
Guttman, N., & Kalish, H. I. (1956). Discriminability and stimulus generalization. Journal of Experimental Psychology, 51, 79-88.
Hanson, H. M. (1959). Effects of discrimination training on stimulus generalization. Journal of Experimental Psychology, 58, 321-334.
Harris, J. A. (2006). Elemental representations of stimuli in associative learning. Psychological Review, 113, 584-605.
Honig, W. K., & Urcuioli, P. J. (1981). The legacy of Guttman and Kalish (1956): 25 years of research on stimulus generalization. Journal of the Experimental Analysis of Behavior, 36, 405-445.
Martindale, C. (2006). Peak shift, behavioral contrast, and phonological evolution with explanations of the Great English Vowel Shift and Grimm's Law. Journal of Cultural and Evolutionary Psychology, 4, 5-26.
McLaren, I. P. L., & Mackintosh, N. J. (2000). An elemental model of associative learning: I. Latent inhibition and perceptual learning. Animal Learning and Behavior, 28, 211-246.
McLaren, I. P. L., & Mackintosh, N. J. (2002). Associative learning and elemental representation II: Generalization and discrimination. Animal Learning and Behavior, 30, 177-200.
Purtle, R. B. (1973). Peak shift: A review. Psychological Bulletin, 112, 408-421.
Ramachandran, V. S., & Hirstein, W. (1999). The science of art: A neurological theory of aesthetic experience. Journal of Consciousness Studies, 6, 15-51.
Riley, D. A. (1968). Discrimination learning. Boston: Allyn & Bacon.
Saksida, L. M. (1999). Effects of similarity and experience on discrimination learning: A nonassociative connectionist model of perceptual learning. Journal of Experimental Psychology: Animal Behavior Processes, 25, 308-323.
Schneider, S. M., & Lickliter, R. (2010). Operant generalization in quail neonates after intradimensional training: Distinguishing positive and negative reinforcement. Behavioural Processes, 83, 1-7.
Spetch, M., Cheng, K., & Clifford, C. (2004). Peak shift but not range effects in recognition of faces. Learning and Motivation, 35, 221-241.
ten Cate, C., Verzijden, M. N., & Etman, E. (2006). Sexual imprinting can induce sexual preferences for exaggerated parental traits. Current Biology, 16, 1128-1132.
Terrace, H. S. (1964). Wavelength generalization after discrimination learning with and without errors. Science, 144, 78-80.
Thomas, D. R., Mood, K., Morrison, S., & Wiertelak, E. (1991). Peak shift revisited: A test of alternative interpretations. Journal of Experimental Psychology: Animal Behavior Processes, 17, ISOMO.
Weary, D. M., Guilford, T. C., & Weisman, R. G. (1993). A product of discriminative learning may lead to female preferences for elaborate males. Evolution, 47, 333-336.
Wisniewski, M. G., Church, B. A., & Mercado, E. (2010). Temporal dynamics of generalization and representational distortion. Psychonomic Bulletin & Review, 17, 809-814.
Yachi, S., & Higashi, M. (1998). The evolution of warning signals. Nature, 394, 882-884.
Yachi, S., & Higashi, M. (1999). Modeling associative learning with generalization for a case of warning signals. Ecological Research, 14, 243-248.
Zimmer, R. (2003). Abstraction in art. Philosophical Transactions of the Royal Society of London B, Biological Sciences, 358, 1285-1291.
Author Note Thanks are extended to Billea Ahlgrim, Lauren Henry, Stephanie Joppa, Jessica Kahnke, Cory Klein, Elizabeth Klosterman, and Amanda Quinn for their assistance with this research.
A. Derenne ([mail]) * E. A. Loshek * B. Bohrer
Department of Psychology, University of North Dakota, PO Box 8380, Grand Forks, ND 58202-8380, USA e-mail: firstname.lastname@example.org
|Printer friendly Cite/link Email Feedback|
|Title Annotation:||ORIGINAL ARTICLE|
|Author:||Derenne, Adam; Loshek, Eevett A.; Bohrer, Brittany|
|Publication:||The Psychological Record|
|Date:||Mar 1, 2015|
|Previous Article:||The behavior of controllers.|
|Next Article:||Assessing stimulus control in a discrimination task with compound stimuli: evaluating testing procedures and tracking eye fixations.|