Weaker face recognition in adults with autism arises from perceptually based alterations

Face recognition has been shown to be impaired in autism spectrum disorders (ASD). However, it is still debated whether these face processing deficits arise from perceptually based alterations. We tested individuals with ASD and matched typically developing (TD) individuals using a delayed estimation task in which a single target face was shown either upright or inverted. Participants selected a face that best resembled the target face out of a cyclic space of morphed faces. To enable the disentanglement of visual from mnemonic processing, reports were required either following a 1 and 6 second retention interval, or simultaneously while the target face was still visible. Individuals with ASD made significantly more errors than TD individuals in both the simultaneous and delayed intervals, indicating that face recognition deficits in autism are also perceptual rather than strictly memory based. Moreover, individuals with ASD exhibited weaker inversion effects than the TD individuals, on all retention intervals. This finding, that was mostly evident in precision errors, suggests that contrary to the more precise representations of upright faces in TD individuals, individuals with ASD exhibit similar levels of precision for inverted and upright faces, for both simultaneous and delayed conditions. These results suggest that weakened memory for faces reported in ASD may be secondary to an underlying perceptual deficit in face processing.


Abstract
Face recognition has been shown to be impaired in autism spectrum disorders (ASD). However, it is still debated whether these face processing deficits arise from perceptually based alterations. We tested individuals with ASD and matched typically developing (TD) individuals using a delayed estimation task in which a single target face was shown either upright or inverted. Participants selected a face that best resembled the target face out of a cyclic space of morphed faces. To enable the disentanglement of visual from mnemonic processing, reports were required either following a 1 and 6 second retention interval, or simultaneously while the target face was still visible. Individuals with ASD made significantly more errors than TD individuals in both the simultaneous and delayed intervals, indicating that face recognition deficits in autism are also perceptual rather than strictly memory based. Moreover, individuals with ASD exhibited weaker inversion effects than the TD individuals, on all retention intervals. This finding, that was mostly evident in precision errors, suggests that contrary to the more precise representations of upright faces in TD individuals, individuals with ASD exhibit similar levels of precision for inverted and upright faces, for both simultaneous and delayed conditions. These results suggest that weakened memory for faces reported in ASD may be secondary to an underlying perceptual deficit in face processing.

Lay Summary
Individuals with autism spectrum disorder (ASD) are known to have difficulties recognizing faces on a day-to-day basis. This research delves into the potential reasons underlying this difficulty. Results suggest that the facial processing impairment within ASD may not be only a memory difficulty but also a perceptual difficulty. These findings add to the understanding of the core abilities within ASD and may have direct implications for the treatment of facial processing difficulties within this population.

INTRODUCTION
Individuals with autism spectrum disorder (ASD) have persistent deficits in social communication and social interaction across multiple contexts (DSM À5). It is diagnosed on the basis of behavioral symptoms in two domains: social communication and restricted, repetitive patterns of behaviors, interests, or activities (American Psychiatric Association, 2013). Atypical perception has also been reported within ASD demonstrating a range of modulated perceptual functioning manifested in different sensory modalities (e.g., Dakin & Frith, 2005;Hadad & Yashar, 2022). Perhaps the most robust modulated perception that appears directly related to the core social impairments experienced by individuals with ASD are processes associated with face recognition. Consistent with the central symptoms of social dysfunction, face processing and recognition is thought to be a domain of weakness for many individuals with ASD (e.g., Hartston et al., n.d.;Dawson et al., 2005;Tanaka & Sung, 2016). Griffin et al. (2021) found comparable and large deficits in ASD for face identity and recognition and reported that scores of an average individual with ASD were nearly 1 SD below the average typically developing (TD) individual on face processing tasks. Howard et al. (2000) found that people with highfunctioning autism showed neuropsychological profiles characteristic of the effects of amygdala damage, in particular selective impairment in the recognition memory for faces. As face processing is strongly associated with social interaction skills, the prime difficulty within ASD (i.e., Giannou et al., 2020;Lander et al., 2018), impaired face recognition has been reported to affect these individuals on a day-to-day basis (Stanti c et al., 2021).
However, the underlying cause of this weakness remains unresolved. The literature in this area of research has been inconsistent (See Tang et al., 2015;Weigelt et al., 2012), with studies debating whether deficits in face processing are perceptually based or rather arise only from memory impairment. Weigelt et al. (2012) found quantitative, but not qualitative, differences in face identification in individuals with ASD. They suggest that face processing markers, such as the inversion effect (the advantage in processing upright over inverted faces) appear to be present similarly within TD individuals and individuals with ASD. Moreover, they have claimed that quantitatively (i.e., how well faces are remembered or discriminated), individuals with ASD perform worse than TD individuals in face memory tasks in which a delay intervenes between sample and test, and less so in tasks with minimal mnemonic demands. By contrast, a more recent review by Tang et al. (2015) depicts both qualitative and quantitative differences in face identification between individuals with and without ASD.
Indeed, several recent findings point to a perceptual deficit. Hadad and collogues have demonstrated weaker face discrimination in Autism even for discrimination of simultaneously presented faces . Furthermore, face discrimination in ASD is shown to be qualitatively different, rather than being simply less efficient than in TD individuals. Vettori et al. (2019) and  all demonstrated weaker or no inversion effect in ASD and claim that individuals with ASD present atypical perceptual abilities. The face inversion effect, that is typically shown to be larger for faces compared with nonfacial objects (Civile et al., 2016;Yin, 1969), has been taken to support the configural information hypothesis, which states that faces are processed in a holistic representation. Inverting faces disrupts configural processing insisting on the use of local facial information. Thus, the reduced face inversion effect observed in ASD, reflects a part-based processing of faces and thus demonstrate alterations occurring primarily at the encoding stage of face processing, rather than at the storage of face representations in memory (Freire et al., 2000). It also implies a reduced sensitivity to the predominant, upright faces (e.g., . Consistent with this interpretation, people with ASD also show reduced other-race effect in face discrimination and similar perceptual representations for frequently and less frequently encountered faces (own-vs. other-race faces) (Hartston et al., n.d.;. The current study was designed to further extend this new line of studies by examining perceptual processing directly comparing perceptual and memory tasks with varying mnemonic requirements. Stanti c et al. (2021) tested face perception and face memory in the same individuals and reported (a) poorer face recognition abilities in individuals with ASD and (b) perceptual difficulties along with the memory difficulties by these individuals within the face processing tasks. However, tasks differed between perception and memory and thus it was difficult to determine the contribution of each to the weaker face processing in ASD. To examine both perceptual and mnemonic deficits in ASD simultaneously, it is necessary to directly pit against each other conditions that share their visual processing requirements but differ only in their mnemonic requirements. Specifically, it is necessary to assess performance in a task that consists of a condition in which there is no mnemonic encoding as well as conditions with variable retention intervals. In the case of a perceptual deficit, we would expect to find a consistent impairment in all conditions as all rely on visual processing. However, impairment in mnemonic processes, such as retrieval and maintenance, should be reflected in an impairment following a retention interval and no difficulty within the simultaneous condition (Pertzov et al., 2020).
To balance the perception and mnemonic conditions, we used trials with a single face but also maintained a dynamic range of performance by using a delayed estimation task (e.g., Ma et al., 2014;Pertzov et al., 2020) in which participants select a matching item out of a variety of options that are very similar to each other. In the delayed estimation paradigm, participants are required to reproduce a previously observed stimulus from a gradually changing cyclic scale. This allows an insight into the types of errors made during face perception. Subjects choose between a circle of faces of which half the faces are morphed versions of the target face (precision errors) and the other half morphed versions of a random face (random errors). These tasks have a large dynamic range of performance (not likely to lead to ceiling and floor effects) and were found to be more sensitive to capture impairments in clinical populations, compared with more traditional memory tasks (Zokaei et al., 2015). They also enable the documentation of the distribution of errors, thus providing data on the type of errors made by the participants (Ma et al., 2014). For example, a complete failure to access a memory representation should be manifested as a uniform distribution of errors across the scale (random errors), whereas a degradation in the fidelity of a memory should lead to a broader distribution of errors around the correct value (precision errors). The results obtained on these delayed estimation tasks suggest that extending the retention interval influences both types of errors: it increases the number of errors distributed randomly on the reporting scale, as well as broadening the distribution of errors around the correct target (Pertzov et al., 2017). As no previous study has used such a paradigm in ASD, little is known about the precision with which faces are stored by the individual with ASD and which type of errors would accompany a deficit in these individuals. Here we inspect these two types of errors, precision and random, in ASD and in TD individuals.
Krill et al., 2018 used a face delayed estimation task with several retention intervals in TD individuals and found that longer retention intervals increase the precision errors but more so the proportion of random errors. These findings imply that forgetting of faces reflects decreased accessibility of the memory representations over time within the TD population. To examine the specificity of the effect, Krill et al. (2018) also included inverted faces and measured the face inversion effect (Yin, 1969). In contrast to the effect of retention interval (i.e., forgetting), face inversion led to larger errors that were mainly associated with decreased precision of recall. This effect was similar in all retention interval conditions even when memory was not required in the task. Krill et al. (2018) concluded that upright faces are remembered more precisely compared with inverted faces due to perceptual, rather than mnemonic mechanisms. Importantly, in a follow up study the paradigm was shown to be suitable for testing neuropsychological populations and was successfully applied to individuals with congenital impairments in face processing (congenital prosopagnosia) (Pertzov et al., 2020).
While there is clear evidence of difficulties within facial processing in ASD, the question regarding the nature of the deficits remains unclear. The current study is the first to test individuals with ASD in a single task containing both perceptual and mnemonic conditions and therefore the first to directly compare perceptual versus memory causes of facial processing difficulties. Through the delayed estimation paradigm and the documentation of the distribution of errors, we were able to analyze the type of errors performedprecision or randomand were able to make inferences regarding their underlying causes.
An overall reduced inversion effect is expected for ASD, indicating that face processing is qualitatively altered in these individuals . Strong support would specifically come from reduced inversion effect on precision errors (similar distribution of errors around the correct value), indicating that upright and inverted faces are processed with similar levels of precision in ASD. Moreover, poorer performance and reduced inversion effect observed in both retention intervals and the simultaneous condition would suggest qualitatively perceptual alterations in face processing in ASD and that difficulties within face processing in ASD is based on atypical perceptual processing.

METHOD Participants
Twenty-six adults with high functioning ASD (five females; mean age = 25.6, range = 18-36) and 26 TD adults (17 females; mean age = 26.2, range = 18-38) participated in the current study. We set the sample size in advance to match the sample size that was used in previous studies measuring perceptual processes in Autism employing within-subjects designs. Past experience has indicated that this sample size is sufficient to show significant differences between groups (e.g., Hadad et al., 2017;Hadad & Ziv, 2015). In addition, we calculated the sample size that was required for statistical power of 0.8, based on effect sizes in Pertzov et al. (2020) which indicated that 24 participants were required for each group. Participants were native Hebrew speakers and reported normal or corrected to normal vision. The groups were matched based on their intelligenceas measured by the Test of Nonverbal Intelligence (TONI4 test, Brown et al., 2010). The TONI allows testing of intelligence without the confounding effects of linguistic skills. All ASD participants underwent the ADOS-2 (Lord et al., 2000) assessment and only participants with a confirmed diagnosis of ASD were included in the study. Participants were recruited via the local community and the "Beit Ekstein Center" for adults with communication difficulties. Participants with ASD were compensated for their time by 50 NIS per hour and TD participants by course credits or 35 NIS per hour. This research was prospectively reviewed and approved by the Institutional Review Board (IRB) of the University of Haifa, IRB #046/20.

Experimental stimuli, materials, and procedures
One hundred and ninety realistic, color pictures of faces (78 female and 112 male) were taken from the Productive Aging Lab Face Database (Minear & Park, 2004), The IMM Face DB (Nordstrøm et al., 2004), and the Glasgow Unfamiliar Face Database (Mike Burton et al., 2010) databases. The stimuli were mixed gender throughout the experiment while each trial used the same gender (circular display of faces). All the stimuli were of Caucasian faces and all participants were Caucasian with no known extensive exposure to other-race faces. All faces used had a neutral expression and were cropped in a fixed round form, without hair (using Adobe Photoshop CS6). To prevent verbal tagging and hamper the usage of potential long-term memory strategies, all faces were displayed only once on a given block and all faces in a trial had similar age, gender, skin tone and facial shape (e.g., cheekbones, jaw line). The stimuli in each trial consisted of three original faces (from the pool) and five morphed faces (Abrosoft FantaMorph deluxe V5) between each pair (83%A/17%B, 67%A/33%B, 50% A/50%B, 33%A/67%B, 17%A/83%B).
Stimuli were presented on a 17.3-inch Dell FDH (1920_1080) Truelife LED-Backlit Touch Display on a Dell Inspiron laptop. Participants were positioned at a comfortable distance from the screen, and the size of stimuli in degrees of visual angle was calculated assuming 60 cm distance. The computer was situated in a quiet room with dimmed light. The experiment was programmed in MATLAB (MathWorks, Inc.) and Psychophysics Toolbox (Brainard, 1997;Pelli, 1997).
The experimental design is illustrated in Figure 1. Each trial began with the presentation of a central fixation cross (white, 3 pixels, 0.08 of visual angle) for 1000 millisecond. This was followed by a display of a single upright or inverted face at the center of the screen (200 Â 200 pixels, 4.5 Â 4.5 ) for 1500 ms. In the memory conditions, a black screen was displayed for 1 or 6 seconds during the retention interval. Participants were instructed to remember the target face and, at the end of the retention interval, to report its identity by touching one of the faces out of a cyclic array of 18 faces situated on an imaginary circle around the fixation within a distance of 470 pixels (10.6 ). The selected face was then displayed at the center (mnemonic conditions) or to the left of the target face (nonmnemonic condition). Participants had up to 25 seconds to modify their selection and once they reached their final decision, they clicked the space bar to register their selection and proceed to the next trial. Figure 2 describes the way in which the circular array of report circle was designed (the colored circles, frames and text were not displayed in the experiment): the previously displayed face (target) was shown in the circular array with two other faces (original faces shown in a red frame). Five equally spaced morphed faces were generated between each pair of the original faces. Thus, each target face was accompanied by five morphed faces between it and each one of the two other original faces. Five other morphed faces between the two nontarget original faces were also displayed in the circular array, these morphs were not generated from the target face. The entire circular array was randomly rotated in every trial to avoid strategic learning. The nonmnemonic block (simultaneous condition) consisted of 30 trials in which a single face was presented simultaneously with the circular array of 18 faces. Participants were instructed to select the identical face on the report circle (see Figure 1). Each participant performed three memory blocks and one F I G U R E 1 (a) Experimental design of memory trials. A single target face (upright or inverted) was presented, followed by 1 or 6 second of a blank screen retention interval. Next, a circular array of 18 optional faces was displayed. The participants were required to select the face that was identical to the target face that they had just seen. (b) Experimental design of the simultaneous condition in which there was no retention interval between the display and the report. The participants were required to select the face that matches the target face. (Reproduced with permission from Elsevier, license number 5180810574482). perception (nonmnemonic) block. More memory blocks were tested than perception blocks due to the testing of the two memory conditions (1 second interval and 6 second interval), and the larger variability within the memory blocks (see Krill et al., 2018). Each memory block was composed of 30 trials consisting of all conditions, half with a 1 second delay and half with a 6 second delay. Each trial was randomly assigned to the upright or the inverted condition. Trials were displayed in random order and none of the faces were repeated within a block.
To ensure participants' engagement in the task, feedback was presented every 10 trials, depicting the average error rate on the last 10 trials (the error rate calculation is described in the data analysis section). A score of 100 was given if the participant's average magnitude of errors was less than 1 across the last 10 trials, and the score linearly decreased with an increased error rate to a score of 60. The feedback provided during the experiment was minimal, presented in small written font in English at the top left-hand corner of the screen and no mention was made of this feedback by the examiner to the participants.

Analysis
We followed Ma's recommendations (Ma, 2018) and used a nonparametric measure of error-the mean precision error and the proportion of random errors. This approach avoids the use of mixture modeling for assessing the uniform component and the precision errors (Ma, 2018). Hence, we explored the different types of errors directly using the following approach (as was performed in Krill et al., 2018, Pertzov et al., 2020: First we analyzed the mean raw errors by simply averaging errors of all sizes and of all trials within a condition. Next, we extracted two summary statistics from each distribution of errors: (1) Proportion of random errors: when a participant selected a face from the circle that did not have any resemblance to the target face [morph did not include any fraction of the target face (errors 6, 7, 8, 9 in absolute value)]. In such cases, we assumed that the participant did not remember the target face, and therefore just guessed. To obtain a proportion value, the number of such errors was divided by the overall number of trials, (2) Mean precision error: calculated based on trials in which a participant reported a face that had some resemblance to the target face (i.e., was a morph of the target face). In such cases, we assumed that participants had some recollection of the target face. To quantify the degree of precision recall we averaged the magnitude of the absolute errors (range of 0-5). Note that when participants did not have any recollection of the target face, they were likely to guess a random face and therefore sometimes report a face with some resemblance to the F I G U R E 2 The circular report array comprised of 18 faces. The three original (un-morphed) faces are marked by red rectangular frames (the colored circles, frames and text were not displayed in the experiment). Face "0" was the target face that the participant was required to report, hence the correct answer. Faces 6 and À6 are the other two original faces. All the faces between the three original faces are linear morphs between the two closest original faces. Selection of a face that was generated from a morph with the target face and therefore had some resemblance to it (À5 to 5) was treated as a precision error. Selection of a face that was not generated from a morph with the target face and therefore had no resemblance to it (errors above 5 and below À5) was treated as a random error. (Reproduced with permission from Elsevier, license number 5180810574482). target face (see uniform distribution in Figure 3). Also note that the two summary statistics, (1) proportion of random errors and (2) mean precision error, are somewhat independent of each other. The proportion of random errors is sensitive only to the proportion of trials defined as random while mean precision error is sensitive to the magnitude of errors, which is related to the shape of the error distribution rather than to the proportion of trials in it. Thus, it is reasonable that an experimental manipulation would modulate the proportion of random errors but not the mean precision error, and vice versa. Figure 3 is an illustration provided to emphasize that a small proportion of the random errors were within the threshold and could therefore be classified as either random or precision errors. Thus, the average number of random errors was extracted for each subject and this number of errors was classified as random even when they were within the "precision" threshold. For statistical analysis, within this mixed-design study, we applied a repeated measures ANOVA with face orientation (upright or inverted) and retention interval (simultaneous, 1 or 6 s) as within-subject factors, and group (ASD or TD) as between-subject factor. The dependent variables were mean raw error, proportion of random and the mean precision error. We conducted a separate repeated measure ANOVAs for each dependent variable.

Mean raw error
First, before extracting the distribution of errors to precision and random errors, we analyzed the mean size of all the errors using a repeated measures ANOVA with retention-interval (simultaneous, 1 and 6 seconds) and orientation (upright vs. inverted) as within participant factors, and group (TD/ASD) as the between-subject factor. As can be seen in Figure 4, performance was poorer in the ASD group; F(1,50) = 29.64, p < 0.001, η 2 p = 0.37, and for the inverted faces; F(1,50) = 7.46, p < 0.01, η 2 p = 0.13. A significant effect of the interval condition was observed; F(2,100) = 233.9, p < 0.001, η 2 p = 0.82, that interacted with group, F(2,100) = 3.19, p < 0.05, η 2 p = 0.06. Further analysis testing group difference in each interval revealed differences between the groups in all retention intervals, with the ASD group performing poorer than the TD group, in the simultaneous condition, F(1,50) = 13.6, p < 0.001, η 2 p = 0.21, the 1 s condition, F(1,50) = 31.6, p < 0.001, η 2 p = 0.39 and in the 6 s condition, F(1,50) = 17.4, p < 0.001, η 2 p = 0.26. Next, we examined random errors, which are uniformly distributed errors presumably originating from a complete failure to access information regarding the probed face (and therefore the reports are spread randomly over the reporting scale). Precision errors are errors that are distributed around the correct face in memory, reflecting some rudimentary recollection of the face.
F I G U R E 3 An example of an error distribution in one condition (upright face, long retention interval) for the TD group (left) and ASDs (right).
In accordance with the color code used in Figure 2, precision errors are illustrated in green and random errors in pink. The number of random errors per bin (average number of errors of 6, 7, 8, 9 in absolute value) was subtracted from the precision errors and added to the proportion of random errors (see Krill et al., 2018).
Further analysis carried out by group revealed that only the TD individuals showed a significant inversion effect, F(1,25) = 6.99, p < 0.05, η 2 p = 0.22, while no face orientation effect was observed for the ASD group, F (1,25) = 0.005, p = 0.94, η 2 p = 0.00 in the precision errors. This once again reiterates the inversion effect seen in TD individuals which is not seen here by the ASD group.

Response times
Participants were instructed to report as accurately as they can, regardless of time (but respond within a time frame of 25 s). As can be seen in Figure 7, there was no significant difference between groups in any of the intervals; simultaneous F(1,48) = 1.17, p = 0.28, 1 second interval F(1,50) = 2.41, p = 0.13, 6 second interval F (1,45) = 2.61, p = 0.11. Note that in the simultaneous condition the target face was displayed during the entire F I G U R E 4 Mean raw errors for upright and inverted faces across retention intervals in the TD group (blue) and the ASD group (red). Mean raw errors are denoted in morph level units (see Figure 2). Error bars indicate standard error of the mean across participants. * and ** denote p values of 0.05 and 0.001 of direct contrasts between groups, respectively.
F I G U R E 5 Proportion of random errors for upright and inverted faces across retention intervals. Conventions as in Figure 4.
reporting phase, allowing the participants to correct their reports until the selected face seemed to match to the target face. This correction procedure takes timeleading to slower reports. The ASD group responded quicker, possibly due to their more impulsive nature, than the TD group. Despite the more impulsive nature of the responses of the ASD group, the results remained clear. In conclusion, the group deficits in performance do not appear to be due to speed-accuracy tradeoff.

DISCUSSION
Individuals with ASD and TD individuals performed a novel, single face, delayed estimation task with variable retention intervals to gain insight into whether their face processing difficulties arise due to a perceptual difficulty or solely due to a memory difficulty, as previously suggested (Weigelt et al., 2014). Overall, individuals with ASD exhibited higher rates of raw errors, random errors, F I G U R E 6 Precision errors for upright and inverted faces across retention intervals. Conventions as in Figure 4. Bottom panel: Precision errors according to TD and ASD groups.
F I G U R E 7 Response time for upright and inverted faces across retention intervals. Conventions as in Figure 4. and precision errors than TD individuals. Group differences between the ASD group and the TD group were found in all retention interval conditions, including in the simultaneous condition in which there were no mnemonic requirements. These findings provide direct evidence for an atypical perceptual processing of faces in ASD and suggest that alterations in processing faces in Autism are likely to occur at the level of perceptual encoding. This is mainly supported by the weaker performance in face processing that is shown in all interval conditions, including the simultaneous presentation of the faces, and by the reduction of performance in ASD being of similar magnitude when faces are presented simultaneously as when they must be remembered for up to 6 s. The memory load affected the two groups similarly with higher error rates in the longer interval seen for both groups. Altogether, the results suggest that weaker face processing in autism arises from altered perceptual processing.
Moreover, the inversion effect, that often characterizes face perception in the TD population, is not observed in the ASD group. The reduced inversion effect was observed for both delayed intervals and the simultaneous condition. This suggests that the atypical perceptual processing that is seen within ASD does not solely reflect a quantitatively weaker system; individuals with ASD are not simply less good in face processing; rather, they seem to process faces in a qualitatively different manner. Contrary to the more precise reports of upright versus inverted faces in the TD group, upright and inverted faces are processed with similar levels of precision in the ASD group, implying the configural manner of processing upright faces is weaker in this group. This interpretation of the results is consistent with Schultz et al. (2000) who concluded that individuals with ASD demonstrate a pattern of brain activity during face discrimination that is consistent with feature-based strategies.
Face processing in neurotypical individuals is shown to rely on configural information in a face, that is, processing not just the individuals features but also the relations among them (e.g., Avidan et al., 2011;Maurer et al., 2002). This ability typically develops over years of exposure and experience in differentiating among frequently encountered faces (e.g., upright faces). Since the classic demonstration of the inversion effect (Yin, 1969), it has been used as a marker for configural face processing and reduction or absence of this marker has been interpreted as stronger reliance on part-based processing, associated with weaker face discrimination and recognition (e.g., Joseph & Tanaka, 2003). To the extent to which reduced inversion effect reflects weaker configural processing that is more feature-based, the results in the ASD demonstrate an overall weak central coherence (e.g., Happé & Frith, 2006), and in faces specifically, reflect their difficulty to extract the identity of the face arising from integrating features and their interrelations. It could also reflect a general failure of the perceptual system within individuals with ASD to reach typical levels of face processing specialization. Weaker face recognition is specifically found in ASD for the more frequently encountered faces (i.e., upright, own-race faces;  suggesting that perception is less refined by exposure (Karaminis et al., 2016). Perceptual specialization within face recognition has been shown to continue through to the age of 30 within TD individuals (Germine et al., 2011); however, individuals with ASD show a different pattern of development (O'Hearn et al., 2019). Reduced sensitivity, within faces and nonfaces, is shown in ASD specifically to native discrimination (own-race faces, native phonemes, and cardinal orientations), while nonnative discriminations may remain intact (Hadad & Yashar, 2022). As specialization to these stimuli normally emerges during infancy (e.g., Quinn et al., 2013), it is likely that the reduced specialization to predominant faces in adults with Autism begins during the first months of age. Indeed, the duration of time spent fixating on faces varies consistently across individuals (Guy et al., 2019), and young children with ASD were shown to focus less on faces compared with TD individuals (Avni et al., 2020). Thus, a developmental perspective could add significant insight into the knowledge regarding specialization of face processing within TD and ASD individuals.
One limitation of this current study is the gender imbalance between the groups. This limitation came about with the strong male prevalence within ASD (4:1); fewer females with ASD within the near-by population and fewer who were willing to participate in the study. Note, however, that Griffin et al. (2021), following their extensive review on studies of face processing in ASD, found no systematic relation between the proportion of male participants in the study sample and the magnitude of face identity recognition differences between the ASD and TD groups. Furthermore, as a further precaution we have carried out statistical analyses within groups, for both random and precision errors, which have provided insignificant results between the genders (see Supplementary Material).
The present findings show that contrary to previous claims (Weigelt et al., 2012), weakened memory for faces in ASD may be secondary to the underlying perceptual deficit in face processing. Along similar lines, an apparent spatial working memory deficit has been shown to be the result of an inability to perform the complex information processing typically associated with spatial working memory tasks, rather than attributed to an impaired working memory (e.g., Minshew & Goldstein, 2001;Williams et al., 2005). As in the present study, this consideration could lead to the conclusion that memory is not solely the cognitive domain responsible for the poor performances obtained by individuals with autism; atypical perceptual processing appears to be responsible along with the memory difficulties. Thus, better understanding and treatment of weakened face processing skills, which may often appear as a deficiency in memory, must be targeted primarily at the processing stage of face perception.

ACKNOWLEDGMENT
This research was funded by the Israel Science Foundation (ISF), grant #882/19 to Bat-Sheva Hadad.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available from the corresponding author upon reasonable request.