Edited by: Marcus Cheetham, University of Zürich, Switzerland
Reviewed by: Rosemarie Velik, Carinthian Tech Research, Austria; Anthony Paul Atkinson, Durham University, UK
*Correspondence: Tyler J. Burleigh, Department of Psychology, University of Guelph, 50 Stone Road East, Guelph, ON N1G 2W1, Canada e-mail:
This article was submitted to Cognitive Science, a section of the journal Frontiers in Psychology.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
The uncanny valley (UCV) hypothesis describes a non-linear relationship between perceived human-likeness and affective response. The “uncanny valley” refers to an intermediate level of human-likeness that is associated with strong negative affect. Recent studies have suggested that the uncanny valley might result from the categorical perception of human-like stimuli during identification. When presented with stimuli sharing human-like traits, participants attempt to segment the continuum in “human” and “non-human” categories. Due to the ambiguity of stimuli located at a category boundary, categorization difficulty gives rise to a strong, negative affective response. Importantly, researchers who have studied the UCV in terms of categorical perception have focused on categorization responses rather than affective ratings. In the present study, we examined whether the negative affect associated with the UCV might be explained in terms of an individual's degree of exposure to stimuli. In two experiments, we tested a frequency-based model against a categorical perception model using a category-learning paradigm. We manipulated the frequency of exemplars that were presented to participants from two categories during a training phase. We then examined categorization and affective responses functions, as well as the relationship between categorization and affective responses. Supporting previous findings, categorization responses suggested that participants acquired novel category structures that reflected a category boundary. These category structures appeared to influence affective ratings of eeriness. Crucially, participants' ratings of eeriness were additionally affected by exemplar frequency. Taken together, these findings suggest that the UCV is determined by both categorical properties as well as the frequency of individual exemplars retained in memory.
Categorization is a critical determinant of human survival. In the absence of categories, humans would be required to learn whether each stimulus that we encountered was desirable or noxious as well as whether the conspecifics that we encountered were kin or competitors. The variability in cross-cultural folktaxonomies (Medin and Atran,
Group membership and identity form an especially relevant class of categories for humans (for a review, see Fiske and Taylor,
In the context of human factors, Mori's (
While the essential phenomenon of the uncanny valley has a number of cultural antecedents (Schoenherr and Burleigh,
It is important to distinguish between the classic and generalized forms of the UVH. The classic account of the UVH provided by Mori (
Evidence that has been interpreted as supporting the classic UVH has been obtained from studies using a variety of stimuli selected from a number of ontological categories. A majority of these studies have observed affective functions that are consistent with the UVH when using stimuli representing computer-generated morph sequences of human and non-human entities, including non-human animals, robots, and anthropomorphic dolls (MacDorman and Ishiguro,
Few studies have examined the possibility that perceptual continua representing non-human entities could produce UCV phenomena. To the best of our knowledge, only two studies have examined this possibility. In Yamada et al. (
Although the UVH provides a description of the non-linear response function, it does not explain why this function occurs, nor does it specify the mechanisms that are responsible. A common explanation is that the negative affect associated with uncanny stimuli might be a consequence of biological adaptations for threat avoidance behaviors (e.g., MacDorman et al.,
From this account, it might be reasonable to assume that the UCV phenomenon is specific to observers viewing images of conspecifics—an assumption that would be consistent with the classic UVH. Given that the spread of communicable diseases depends on the genetic similarity between the observer and the stimulus entity, it is possible that a species could have evolved mechanisms that allow them to respond to pathogen cues in conspecifics, but not heterospecifics. Communicable diseases, however, are not the only source of contamination that members of a species have had to contend with in their environments. As Rozin and Fallon (
Another theory that accounts for threat avoidance behavior is based on the premise that appearances provide information that allows individuals to predict behavior, and thus to anticipate potential threats in their environment. Some uncanny valley stimuli can be seen to present mismatched features (Seyama and Nagayama,
If feature mismatch is the result of the inclusion of features from neighboring categories, then a crucial feature of any general account of the UVH is the specification of category learning systems that acquire the category structure, as well as the representations that are retained within them (for a recent review, see Goldstone et al.,
Empirical evidence is consistent with accounts of the uncanny valley based on categorical perception. For instance, Cheetham et al. (
Any explanation of the UCV phenomenon based on categorical perception must consider categorization processes and representational assumptions (e.g., prototype-related models were recently considered by Moore,
However, while a category boundary model might provide an adequate explanation of the uncanny “valley,” which is a simple U-shaped quadratic function, it cannot account for the entire UCV response function, which is a more complex N-shaped cubic function (e.g., Mori,
Thus, a key difference between category boundary and exemplar-based models is how individuals become sensitized to perceptual space. Category boundary models suggest that individuals can only typically become sensitized to a single region of perceptual space, namely where the category boundary is located; whereas exemplar-based models suggest that individuals can become sensitized to multiple regions of perceptual space, due to the distributions of individual members (Nosofsky,
Crucially, affective processing of stimuli might not require the instantiation of categorical processes. The mere-exposure effect (Zajonc,
We suggest that at least two broad relationships are possible between cognitive and affective processing of stimuli, which we conceptualize as categorical perception (Figure
The categorical perception model (Figure
Alternatively, the frequency-based exposure model (Figure
The present study was designed to test the predictions of the multistage models of the uncanny valley presented in Figures
Importantly, the stage models do not make predictions concerning the specific nature of categorical processing, only the relationship between categorical and affective responses. Therefore, a nested prediction concerns whether categorization will reflect category boundary or exemplar-based representations. The first possibility is that individuals will only have access to a category boundary representation that partitions the response continuum. Therefore, categorization accuracy and affective responses should increase, and response times should decrease, as a function of a stimulus' distance away from the category boundary. If participants are insensitive to individual characteristics, then categorization uncertainty should also be evidenced by a linear increase in response latencies as a function of proximity to the category boundary. Alternatively, if exemplar-based representations are acquired for two contrasting categories and used for categorical processing, then the location of the central tendencies for each category should determine the location of the maxima and minima of the response functions for affective and categorization responses. Response latencies should evidence a similar trend. Specifically, if uncertainty in category membership is a function of exemplar frequency, then we would expect exemplars presented with comparatively high frequency during training to be associated with fast responses whereas exemplars presented with comparatively low frequency to be associated with slow responses.
In order to test the predictions of these models, our experimental design uses a category-learning paradigm in which we manipulate exemplar frequency along the perceptual continuum. Experiment 1 consists of two training conditions. In the first condition, stimuli within response categories are presented with equal frequency, with each category having an equivalent distribution (EFED). In the second condition, both category distributions are equivalent, but the exemplars were presented with unequal frequency (UFED) such that stimuli near the extrema of Categories A and B training sets are presented with the greatest frequency, and stimuli adjacent to the category boundary were presented with the lowest frequency. An important aspect of our design is that individuals are not exposed to the continuum extrema during the training phase. Thus, while the category boundary of the EFED and the UFED conditions should be identical, differences in exemplar frequency should decrease affective responses outside the training range if frequency-based information is a determinant of categorical and/or affective responses.
The results of Experiment 1 should provide a straightforward tests of our predictions. Left unaddressed, however, is what we consider to be a tacit property of UCV as discussed by Mori (
EFED | – | – | 4 | 4 | 4 | 4 | 4 | – | 4 | 4 | 4 | 4 | 4 | – | – | 40 |
UFED | – | – | 8 | 6 | 4 | 2 | – | – | – | 2 | 4 | 6 | 8 | – | – | 40 |
UFUD | – | – | 8 | 6 | 4 | 2 | – | – | 4 | 4 | 4 | 4 | 4 | – | – | 40 |
Crucially, we were also interested in determining whether the affective response patterns could reasonably support a UCV interpretation. We distinguish between “strong” and “weak” interpretation as follows. The UCV function is a non-linear response function that is defined by a slope, indicating a category preference attributable to familiarity (e.g., for humans over robots), and a valley region that is located near the category boundary but skewed toward the preferred category. Thus, support for a strong interpretation of the UCV would be obtained if a response function possessed all of these features; support for a weak interpretation of the UCV would be obtained if a response function possessed some of these features, such as a valley region without a slope. We anticipate the possibility that the EFED and UFED conditions might provide support for a weak interpretation, but not for the strong interpretation, due to their symmetry. In contrast, the UFUD condition might provide support for a strong interpretation of the UCV due to the asymmetry of the response function.
Although the stimuli that we use all represent non-human entities, we believe the findings of these studies are pertinent to human-like stimuli. By using non-human stimuli we hope to minimize the influence of stimulus familiarity or preference for human stimuli. This novelty facilitates the task of training participants to learn different category structures in an experimental setting with practical limitations (e.g., time). This manipulation also allows us to illustrate that response patterns associated with the UCV are generally patterns that can be attributed to stimulus familiarity rather than human-likeness,
Experiment 1 was designed as an initial test of our predictions derived from the hypothesized multistage models, and to provide evidence in support of the UCV phenomenon. We manipulated the frequency of stimulus presentation to differentially sensitize participants to regions of the stimulus response continuum. An equal frequency condition (EFED) was provided to half of the participants, wherein all stimuli within a category were presented with equal frequency, thereby creating a uniform distribution. An unequal frequency condition (UFED) was provided to the remaining half of the participants, wherein stimuli located within the middle of each category distribution were presented with higher frequency, thereby approximating a normal distribution. In each case, distributions of stimuli from Category A and Category B were symmetrical. Thus, by the end of training we hypothesized that participants should learn the distribution of the training stimuli equally well. Following training, participants responded to stimuli selected from the entire continuum. In the UFED condition, we additionally predicted that participants should show changes in affective responses due to less familiarity with the extreme values that in fact share fewer features with the contrasting category.
A total of 60 participants were recruited online for this study (31 female,
Three morph sequences were generated, comprising the permutations of three distinct non-human creatures: a beast, a reptile, and an alien. These creatures were selected given our assumption that participants would have less familiarity with these categories thereby allowing us to more readily manipulate their frequency of exposure in the experimental context. Creatures were created using Daz Studio 4.6 Pro (daz3d.com) by modifying the morphology and texture of the
At the beginning of training, participants were instructed that they would be presented with “models of unfamiliar living creatures” and that their task was to “learn what categories they belonged to.” They were told that each creature was either a “Cax” or a “Miv” and that they were to press the “C” or “M” key depending on which type of creature they thought they saw. Participants were instructed to balance the demands of speed and accuracy. Key assignment was counter-balanced across participants.
Participants completed 1 training session per day over the course of 3 days. Each training session was composed of 10 blocks of 40 trials each, for a total of 400 trials per training session, and each session required approximately 20 min to complete. For each trial, a fixation point was presented for 500 ms, followed by a randomly selected stimulus from the training distribution for 750 ms (these timings were selected to be consistent with Cheetham et al.,
In order to test our predictions, we analyzed training and test responses separately in terms of categorization accuracy, response time, eeriness ratings, and the shape of categorization and affective response functions. A series of repeated-measures analyses of variance (ANOVA) were conducted. Greenhouse-Geisser adjusted values are reported with unadjusted degrees of freedom. All reported pairwise comparisons were conducted using a Bonferroni adjustment. We also report partial-eta squared as a measure of effect size. Following this, we use curve fitting analyses in order to facilitate our interpretation of the affective response functions.
A repeated-measures ANOVA was conducted on categorization accuracy, using stimulus location relative to the category boundary (4) and response category (2) as within-subjects variables, and stimulus training distribution (2) as a between-subjects variable. Here, response categories (i.e., “Cax” and “Miv”) were randomly assigned to stimuli located on the left- and right-halves of the stimulus continuum. Given the counterbalancing of stimulus sets, we collapsed across morph models prior to analysis. Stimuli directly adjacent to the category boundary in the UFED training condition were also removed prior to analysis. This adjustment was made due to the fact that these stimuli were not present in the EFED training condition and might introduce bias in the analysis. Similarly, the stimulus located at the category boundary was not presented during training and was therefore absent from the analysis of test responses.
Our analysis of training response accuracy revealed a significant main effect for stimulus distance from the category boundary,
Using the same design as the analysis of accuracy, a repeated-measures ANOVA was conducted on categorization response time. To eliminate outlying observations, we first computed an unadjusted mean response time for each participant and identified trials wherein their responses were 3 standard deviations above the mean. This accounted for 2.1% of trials. Consistent with our analysis of categorization accuracy, categorization response time decreased as a function of distance from the category boundary,
As Figure
A repeated-measures ANOVA was conducted on accuracy of responses obtained during the test phase. This analysis was comparable to that of the training phase with the exception that due to the uniform distribution used during the test phase for both training groups, stimuli adjacent to the category boundary as well as novel extrapolation items outside the range of the training items were also included in the analysis. Again, the stimulus located at the category boundary was eliminated from the analysis due to its ambiguity (it was entered into another analysis, see the Categorization response times and Affective ratings of eeriness sections below). Thus, stimulus location relative to the category boundary (7) and response category (2) were entered as within-subjects variables, and stimulus training distribution (2) was entered as a between-subjects variable.
Replicating the findings of categorization accuracy obtained in the training phase, we observed a significant main effect of stimulus distance from the category boundary,
A repeated-measures ANOVA was conducted on categorization response time in the test phase. In order to compare to affective response functions (see below), the stimulus at the category boundary was also included. Therefore, unlike previous analyses, the entire stimulus continuum was tested. Thus, stimulus location (15) was entered as a within-subjects variable and stimulus training distribution (2) as a between-subjects variable. An analysis of response time outliers was again conducted on individual participants' responses. After obtaining an unadjusted mean, no responses were observed to be larger than 3 standard deviations above the mean. This result is not surprising given the reduced number of replications in the test phase.
As with the response time analysis in the training phase, a main effect was observed for stimulus location,
In the categorical perception literature, category boundaries were originally assessed by examining category response frequencies across a stimulus continuum (cf. Pisoni and Tash,
Figure
In order to examine the location and property of the global minima in eeriness ratings that correspond to the uncanny valley, we conducted a repeated-measures ANOVA that included stimulus location (15) as a within-subjects variable and stimulus training distribution (2) as a between-subjects variable. Relative to the previous analyses of categorization accuracy and response time, this approach allows for a straightforward comparison between the shapes of the function that fit affective responses provided below. Our analysis revealed a marginally significant effect of stimulus location,
Although an interaction was not observed between stimulus location and stimulus training distribution, Figure
The variation in patterns across our dependent measures prompted an examination of the relationship between these measures. Test accuracy, test response time, and affective ratings were included in a correlational analysis. We examined the EFED and UFED conditions separately.
As Table
Accuracy | −0.485 (0.079) | - | |
Eeriness | 0.076 (0.796) | −0.059 (0.842) | |
Accuracy | −0.517 (0.059) | - | |
Eeriness | 0.136 (0.643) | 0.076 (0.797) |
Mori's original proposal assumed that the uncanny valley is characterized by a non-linear response function. In the present experiment, we sought to directly test this assumption by fitting curves to the obtained response functions (see also Burleigh et al.,
A number of non-linear functions were selected on theoretical grounds that were not included in Mori's original characterization of the model. In particular, our manipulation of frequency effects in the context of a categorization experiment was motivated by the belief that when participants are sensitized to specific regions of the response continuum, the location of affective minima and maxima can be manipulated. As we noted, a category boundary representation would be evidenced by a U-shaped quadratic function, whereas an exemplar-based category representation would be evidenced by an M-shaped quintic function.
We used a curve fitting analytic approach to test these possibilities, by fitting polynomials of degree 0 through 5 (i.e., constant, linear, quadratic, cubic, quartic, and quintic) to the means. Curve fitting was performed using Origin Lab (originlab.com) software. We used the Akaike Information Criterion (AIC; see Burnham and Anderson,
EFED | Constant1 | 3.521 | −19.43 | 52.81 | <0.001 | 0.054 |
Condition | Linear2 | 0.431 | −48.26 | 23.98 | <0.001 | – |
Quadratic3 | 0.152 | −60.70 | 11.54 | 0.002 | – | |
Cubic4 | 0.057 | −71.61 | 0.63 | 0.396 | – | |
Quartic5 | 0.040 | −72.24 | 0.00 | 0.542 | – | |
Quintic6 | 0.036 | −67.84 | 4.40 | 0.060 | – | |
UFED | Constant1 | 0.386 | −52.59 | 4.76 | 0.064 | 0.069 |
Condition | Linear2 | 0.328 | −52.36 | 4.99 | 0.057 | – |
Quadratic3 | 0.246 | −53.45 | 3.90 | 0.099 | – | |
Cubic4 | 0.235 | −50.36 | 6.99 | 0.021 | – | |
Quartic5 | 0.135 | −54.01 | 3.34 | 0.131 | – | |
Quintic6 | 0.073 | −57.35 | 0.00 | 0.693 | – | |
UFUD | Constant1 | 1.145 | −36.28 | 23.26 | <0.001 | 0.069 |
Condition | Linear2 | 0.534 | −45.03 | 14.51 | <0.001 | – |
Quadratic3 | 0.188 | −57.52 | 2.02 | 0.252 | – | |
Cubic4 | 0.181 | −54.22 | 5.32 | 0.049 | – | |
Quartic5 | 0.093 | −59.54 | 0.00 | 0.693 | – | |
Quintic6 | 0.091 | −53.99 | 5.54 | 0.043 | – |
In the EFED condition, the constant, linear, and quadratic models fell outside the confidence set. Thus, accounts based on random responding or based on the association of equivalent negative affect for each stimuli are not supported (constant). Participants also did not appear to be solely biased by one end of the response continuum (linear). Perhaps most importantly for our purposes, a failure to obtain a fit for the quadratic function suggests that a category boundary model does not provide a good fit to the data in the absence of other assumptions. Instead, the model within the confidence set that was most likely to represent the data was the quartic model,
In the UFED condition, we again observed that the constant and linear models fell outside the confidence set. In the same manner as the EFED condition, this suggests that a constant response bias or uniform negative affect were not evidenced in participants' responses (constant) and that a single category was not used as the sole basis for comparison (linear). Instead, the model within the confidence set that is most likely to represent the data was the quintic model.
The results of Experiment 1 add further evidence to the literature for the existence of the UCV phenomenon. Our results, however, qualify categorical perception accounts. Our measures of categorization accuracy, response frequencies, and response latencies all produced categorical response functions indicating that participants successfully learned the category structures. These analyses also suggest that the equivalent categorization performance was obtained on either side of the category boundary, which indicates that categories (i.e., Cax and Miv) were learned equally well. Similarly, whether participants were trained with exemplars with equal frequencies or unequal frequencies did not appear to alter the location of the category boundary. In both equal frequency (EFED) and unequal conditions (UFED), the category boundary was located at stimulus 8. Greater accuracy was obtained for items adjacent to the category boundary in the equal frequency condition relative to those in the unequal frequency condition. While such findings indicate categorical perception, it is not necessarily the case that categorical processing is the primary determinant of affective responses.
In order for the UCV to be understood in a manner similar to Mori's initial conceptualization, an affective relationship must be established with the location of exemplars along a continuum. Our curve-fitting analysis of eeriness ratings indicated that there were differences between the frequency training conditions. In the equal frequency condition, the response pattern was best fit by a cubic function. This pattern was evidenced by a slope, indicating that one response category was preferred to the other, and also a non-linear component at one end of the stimulus continuum. This pattern did not conform to our
The differences in response functions for categorization accuracy and eeriness ratings also suggest an important relationship that has been neither specified nor explicitly examined in the literature examining the UVH. Namely, we found that categorization accuracy and affect did not significantly co-vary. Such a finding has important implications for studies of the UVH that claim that it can be accounted for by categorical perception. We suggest that the methods of Experiment 1 played a key role in dissociating affective and categorical responses. By requiring affective responses immediately after stimulus presentation and prior to categorical responses, the probability that categorical information was available was reduced. Two differences in findings provide clear demonstrations of this dissociation. First, the small amount of error variance in the categorization responses observed for the items on the end of the distribution in the test session can be sharply contrasted against the larger error variance for the affective responses. Second, whereas categorization accuracy was uniformly high and response times were uniformly fast for items located near the ends of the distributions, affective ratings instead showed asymmetric effects with the response functions.
Another interesting finding was the absence of a relationship between response time and eeriness. Long response latencies are typically taken as evidence of response uncertainty. If eeriness is a consequence of uncertainty in the category membership of an exemplar, then eeriness and response times should exhibit a positive correlation. Instead, the absence of a significant correlation suggests that the processes that determine the uncertainty in category membership and the processes underlying eeriness might be supported by different affective and cognitive processes. Thus, whereas the uncanny valley appears to be a product of experience with exemplars, these two processes appear to be separable. It is necessarily the case that at some level of processing these processes must be influenced by the same stimulus information. Yet stages of processing appear to be evidenced such that affective ratings were influenced more by novel exemplars that represented extrapolations for the range of training stimuli, whereas categorization responses appear to be primary influenced by categorical representations of the stimulus continuum stored in long-term memory.
Experiment 1 therefore provides preliminary evidence in support of the frequency-based exposure model of the UVH, and against the categorical perception model. However, categorization responses were consistent with a category boundary representation. Therefore, we suggest that frequency-based memory representations and category boundary representations are both stored in long-term memory, but that the representation of this information produces different patterns of performance in affective and categorical responses. A remaining possibility is that the UCV phenomenon requires both unequal frequencies within a category, and unequal distributions for both reference categories. Experiment 2 examines this possibility.
Experiment 2 was conducted to clarify the relationship between variables observed in Experiment 1 while also further investigating the effect of distributional properties on categorical and affective responses. One way to interpret Mori's (
A total of 30 participants were recruited online for this study (12 female,
The same stimuli were used to ensure a direct comparison with Experiment 1.
All procedures were identical to Experiment 1 with the exception of training frequency. In Experiment 2, we presented participants with one category from the equal frequency condition in Experiment 1 and another category from the unequal frequency condition. We refer to this distribution as the unequal frequency, unequal distribution (UFED) condition below.
As in Experiment 1, we analyzed categorization accuracy, categorization response times, categorization response frequencies, and affective responses in the training and test phases. We also conducted curve fitting analyses to facilitate our interpretation of the affective response pattern. Again we report Greenhouse-Geisser unadjusted values and unadjusted degrees of freedom.
A repeated measures ANOVA was performed on categorization accuracy, using stimulus location relative to the category boundary (4) and response category (2) as within-subjects variables. Unequal and equal frequency categories were assigned to stimuli located on the left- and right-halves of the stimulus continuum, respectively. Similarly, we also collapsed across morph models and excluded the category boundary stimulus, as well as stimuli directly adjacent to the category boundary, because training for these stimuli occurred for only one of the two response categories. As in Experiment 1, we obtained a significant main effect for stimulus distance from the category boundary,
Figure
A similar repeated measures ANOVA was conducted for categorization response time. Prior to this analysis, we removed outlying trials with response times greater than 3 standard deviations from a participant's mean response time for each stimulus. This resulted in a removal of 2% of all trials. As with the analysis of accuracy, main effects were observed for exemplar distance from the category boundary,
Figure
A repeated measures ANOVA was also conducted on categorization response accuracy for stimuli presented during the test phase. As in Experiment 1, this analysis included stimuli adjacent to the category boundary, because participants received an equal frequency of these stimuli in each response category during test. Again, the category boundary stimulus was not included, as there was no objective criteria that could be used to determine accuracy. Thus, stimulus location relative to the category boundary (7) and response category (2) were entered as within-subjects variables.
Replicating the findings of the training phase, we observed significant main effects for stimulus distance from the category boundary,
A repeated measures ANOVA was conducted with categorization response time as the dependent variable. Unlike the analysis of training stimuli, no outliers met the removal criterion. Therefore, all responses were entered into the response time analysis. A main effect for stimulus distance from the category boundary reached significance,
As in Experiment 1, we determined the frequency of “Cax” responses as a proportion of total category responses for each level along the stimulus continuum. These response frequencies were plotted across the stimulus continuum, and fitted a sigmoid function to the data.
Figure
In order to test the possibility that the category boundary in the UFUD condition was biased due to the unequal frequency category, we conducted a follow-up analysis. We fit a sigmoid function to the response frequencies for each participant in all three of the training conditions. Then, we used two-tailed
A repeated measures ANOVA was conducted on affective ratings, with stimulus location (15) entered as a within-subjects variable. The main effect of stimulus location did not reach significance,
Test accuracy, test response time, and affective ratings were included in a correlational analysis. Replicating the results of Experiment 1, we again observed a significant correlation between response time and categorization accuracy,
Next, we were interested in assessing the affective response trend. A visual inspection of Figure
As Table
The results of Experiment 2 further qualify those obtained in Experiment 1. Although categorical performance patterns were similar to those of Experiment 1, we observed an asymmetry in categorization accuracy. Specifically, accuracy was lower for boundary-adjacent stimuli in the unequal frequency response category that was defined by fewer exemplars near the boundary, relative to the response category with more exemplars near the boundary. This suggests that the frequency training had the intended effect on categorical performance. A similar asymmetry was also observed in affective responses. However, given that measures of response times and accuracy did not correlate with affective responses, it is clear that the asymmetry in categorical perception was not the only, or even the primary, determinant of the affective asymmetry. In line with the proposed frequency-based exposure model of the UVH, we suggest that these patterns have a common cause, but that they are due to separable processes. Namely, they are both rooted in the memory traces that are encoded into long-term memory, as a result of participants' exposure to stimuli. Importantly, whereas categorical perception is the result of categorical processes which draw upon these long-term memory stores in order to determine category membership, affective responses appear to be driven by separate sub-categorical processes used to assess familiarity.
Another finding of particular importance to assessing the claims of the UVH is that the affective response pattern was in closer correspondence to the function described by Mori (
In the present study, we obtained patterns across multiple dependent measures that are consistent with the UVH. Our use of affective and categorization responses further allows us to draw specific conclusions concerning the relationship between these processes, thereby establishing the phenomena of the uncanny valley as well as determining its specific properties. Our analyses of cognitive responses strongly suggests that participants perceived the stimuli categorically (Cheetham et al.,
Our observation that the uncanny valley is not solely the result of categorical perception stands in sharp contrast to recent accounts of the phenomenon (e.g., Burleigh et al.,
In the present study, we found strong evidence that supports the role of exemplar frequenies in determining affective responses. Unlike categorization responses wherein the category boundary was the primary determinant of performance, extrapolation items outside the initial training range were associated with greater negative affect relative to items within the training range. Such a finding is of considerable interest given that it goes against a number of well-established findings in the psychological literature. Specifically, end effects are observed when stimuli are presented along a stimulus continuum, and extreme items are identified more quickly and accurately then intermediate items (for a recent exemplar-based model, see Kent and Lamberts,
In contrast to categorical perception accounts of the UVH, the affective responses of our participants clearly demonstrate sensitivities to the distributional properties of categories that resulted from the manipulation of exemplar frequencies (see also, Förster et al.,
Interesting parallels have been drawn between the UVH and the speech perception literature (Moore,
Our study is not the first to consider the role of categorical perception in the UCV phenomenon (Cheetham et al.,
In our study, we trained participants on stimuli belonging to non-human ontological categories to which they had little or no prior exposure. We designed our training regimen to approximate the differing levels of experience that participants would have with natural categories (e.g., human or non-human animal groups). In addition to finding evidence for categorical perception, including fitting the data with logistic functions, our analysis of affective responses demonstrated frequency effects that are not consistent with categorical perception. In a similar way to exemplar-based models that have been provided in speech perception literature to account for prototype effects (Lacerda,
The present study, as well as the current literature on the UVH, leaves open a crucial question asked by researchers studying cognition and affect: what is the causal relationship between affect and cognition? Presently, we proposed two models of the uncanny valley phenomenon. Whereas the categorical perception model assumes that categorical and affective responses are integrated at some level (Stage 3), the frequency-based exposure model assumes that they are separable (Stage 2 and Stage 3 produce different responses). Both models share in the assumption that responses are a function of stimulus comparisons with representations in long-term memory. This elementary distinction leaves open still further possibilities. Elsewhere in the affective processing literature, a number of hypotheses have been put forward which merit investigation (for a review, see Cacioppo and Gardner,
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
We would like to thank Karl F. MacDorman and Mark Brown for their constructive feedback on earlier drafts of the manuscript.
The studies reported in this article were approved by the Research Ethics Board at the University of Guelph (REB #14JA020).
1 A secondary analysis of arcsine transformed data revealed the same pattern as the analysis of untransformed data.
2 A secondary analysis of arcsine transformed data revealed the same pattern as the analysis of untransformed data.
3A secondary analysis of arcsine transformed data revealed the same pattern as the analysis of untransformed data.
4A secondary analysis of this data was conducted using arcsine transformed proportion correct values. This analysis revealed a similar pattern as the analysis of untransformed data. A change from significant to marginally significant results were observed for the main effect of response category [