Color Afterimages in Autistic Adults

Sixteen adults (8 male) diagnosed with an autism spectrum condition took part, recruited through two local autism charities. All adults had received an independent clinical diagnosis of autism (n = 7) or Asperger’s syndrome (n = 9) according to ICD-10 (World Health Organization 1992) or DSM-5 (American Psychiatric Association 2013) criteria. For inclusion in the data analysis participants also had to exceed threshold for autism on the self-report Social Responsiveness Scale-2nd edition (SRS-2) (T-score ≥ 60) (Constantino and Gruber 2012) or the Adult Autism Quotient (AQ; Baron-Cohen et al. 2001) (score ≥ 30, a threshold shown to correspond to a high degree of specificity (i.e. low false-positive rate) and sensitivity (i.e. high hit-rate) (Woodbury-Smith et al. 2005)). Two participants (both male) failed to meet threshold for autism on either questionnaire and were therefore excluded from analysis. Participants were also screened for IQ as an index of their cognitive functioning. One male participant was excluded both due to a low-average full-scale IQ score (of 72) on the Wechsler Abbreviated Scale of Intelligence-Second Edition (WASI-II; Wechsler and Psychological Corporation 2011) and for failure to follow the task instructions, another male participant gave responses consistent with a color vision deficiency on the Ishihara test (Ishihara 1973) and was excluded, as was another female participant who did not follow the task instructions. A final sample of 11 adults (4 male) formed the autism group. The gender ratio in this sample is unusual, given the male preponderance in diagnosed cases of autism.

Sixteen typical adults were recruited from community contacts. Data from two participants were removed as they scored over 60 on the SRS-2. Three further participants were excluded for matching purposes, yielding a final sample of 11 typical adults (4 male). The final groups of typical and autistic participants were well-matched at the group level in terms of gender proportion, mean age and mean IQ, but, as expected, differed significantly on mean SRS-2 and mean AQ score (see Table 1).

All 22 participants reported normal or corrected-to-normal visual acuity and were assessed as having normal color vision using Ishihara plates (Ishihara 1973) and the Lanthony tritan test (Lanthony 1998). Participants were paid standard rates of £7.50 per hour. All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. The research protocol was approved by the University of Sussex Sciences and Technology Cross-School Research Ethics Committee (approval reference ER/JJM29/3) and all participants gave written informed consent prior to participation in this study.

Four photographs were used as the basis for all stimuli presented in the experiment. Each photograph depicted one of the following four scenes: books on shelves (“books”); Culzean Castle in Scotland (“castle”); Bells beach, Australia (“beach”); and a row of painted, wooden beach huts in Whitby, England (“huts”). The books (color non-diagnostic) and castle (color diagnostic) images were identical to those used by Lupyan while the huts and beach photographs were selected as good examples of images which are color-diagnostic (beach) or color non-diagnostic (huts)¹ (see Oliva and Schyns 2000) (Fig. 1).

The adapting stimuli were generated following the Adobe Photoshop actions described in John Sadowski’s online demonstration of the Spanish Castle illusion (Sadowski 2006). The procedure involves the flattening of the luminance profile of the image, inverting the colors in HSL space and enhancing the color intensity of the resulting image. These adapting images are not equated in color saturation, so the image orientation manipulation is used to disentangle the effect of the image content from low-level differences in the adapting stimuli. Since the rotation through 180 degrees does not change the low-level properties of the image, if aftereffects are solely determined by the intensity of the colors in the adapting image, one would expect orientation to have no effect on the magnitude of the aftereffect. In contrast, if the afterimage magnitude is affected by top-down knowledge this may be disrupted by rotation (Lupyan 2015). Therefore, the orientation manipulation provides a control for differences in the low-level properties of the adapting images, and allows for the detection of purported top-down effects on afterimage strength. Inverted and greyscale versions of each original image were also generated using the hue rotation and desaturate functions of Photoshop in HSL-mode. For each of the four basic images, a smooth continuum consisting of 51 images between the original, “positive” image, through the greyscale image and to the inverted “negative” image were created through linear transformation of the RGB values for each pixel.² This continuum was the basis for the adjustment phase of each trial. It is important to note that this stimulus space is arbitrary and does not necessarily represent homogeneous or equal steps in terms of perceptual difference. Comparisons across images (e.g., comparing books to castle) in terms of these raw units should therefore be avoided, particularly as there is some variation in the perceptual size of steps across images. Figure 2 illustrates this variation—the relationship between the adjustment scale and mean saturation of the image is linear, but there is a small amount of variation in the perceptual difference that each step represents across images (see the differences in slope). The mean (SD) step size in CIE chroma units for each image was as follows: castle—2.40 (0.36); books—1.45 (0.19); beach—2.25 (0.27); huts—1.81 (0.45). To address the issue of non-comparable units on the adjustment scale, after the experiment, the continuum images were subsequently re-coded according to their difference from greyscale using CIE chroma, a perceptual color space—this is further explained in the results section. Further details of the colormetric properties of the stimulus images are given in Table 2. Throughout the experiment, a background grey with chromaticity equivalent to D65 (CIE (1931) xyY = 0.313, 0.329, 20) was used.

These images (in all colored and achromatic forms) were always presented centrally on the monitor with a fixation cross in the center. All images subtended 25° of visual angle horizontally and 17° (24° for the books image) of visual angle vertically, and were presented on a grey background. The background subtended 39° of visual angle horizontally and 29° of visual angle vertically.

The adaptation tasks were completed on a 22-inch Mitsubishi DiamondPlus 2070SB Diamondtron CRT monitor, with a resolution of 1600 × 1200 pixels, 24-bit color resolution, and a refresh rate of 100 Hz. Participants made their responses using the appropriate keyboard press. A ColorCal colorimeter (Cambridge Research Systems) was used to measure the monitor and calibrate the primary values (RGB) for the background color regularly during the data collection period. The tasks took place in a blacked-out room, with the monitor as the only source of light. A cardboard viewing tunnel lined with black felt was used to eliminate effect of peripheral objects and colors and a chin rest was used to constrain viewing distance at 57 cm, ensuring consistency of the perceived size of the images and overlap of the after-images with adjustment image.

The experiment involved two versions of an adaptation task, both identical in procedure, length and task, but with different images in each. Version A included the books and castle images and version B the huts and beach images. Each version had two images—one color diagnostic (castle/beach) and one color non-diagnostic (books/huts). In each version one image would be presented upright (hereafter referred to as “up”), and the other upside-down (hereafter “down”). This diagnosticity-orientation pairing was reversed in the second version (e.g. version A—castle up, books down; version B—huts down, beach up). Thus, each participant would adapt to color-diagnostic scenes in both orientations and color non-diagnostic scenes in both orientations. The order in which the participant completed the versions (A/B) was also counterbalanced across participants.

For the purpose of analysis the trials were classified according to the class of the image (i.e. books and huts are “non-diagnostic” and castle and beach are “diagnostic”) and orientation (upright or upside down), such that, across both versions of the adaptation nulling experiment, each participant provided settings for four conditions—diagnostic upright, diagnostic upside-down, non-diagnostic upright and non-diagnostic upside down. The specific image used in each condition varied between participants (e.g. for some the diagnostic upright image was the castle but for others it was the beach) but was counterbalanced across groups.

Participants were seen individually at the university either for a single session lasting approximately 2 h or for two shorter sessions. Within the session, the participant completed one experimental task followed by the two questionnaires tapping autism symptomatology (the AQ and SRS-2) and the WASI-II, followed by the second experimental task.

Prior to beginning the task proper, participants were briefed with instruction sheets explaining the trial procedure and the participant’s task. These instructions encouraged the participants to try to ‘make their adjustments as quickly and accurately’ as they could and emphasized the importance of maintaining a steady gaze, fixed on the central cross.

To begin, participants completed eight ‘norming’ trials, in which there was no adaptation phase, only adjustment to greyscale (trials started at a random point on the positive–negative color continuum). Four norming trials for each stimulus image were completed, with the image only displayed in the orientation it would be presented in for the adaptation trials. Norming was carried out for each image that would be subsequently presented in the adaptation trials. This was ostensibly for practice, allowing participants to become accustomed to the adjustment procedure and range of stimuli, but also allowed recording of the pre-adaptation settings for each image and observer (following Lupyan 2015).

Next, participants were presented with version A or B of the experimental task. Each experimental trial required participants to stare at a fixation cross in the center of an adapting (color) image for 20 s, which was immediately replaced by a greyscale version of the same image in the same orientation. A nulling procedure was used to measure the strength of the afterimage in terms of the shift in perception of this greyscale image towards the positive (natural) colors of the photograph. The participant was instructed to adjust the image until it appeared subjectively greyscale. This was done using the up and down arrow keys on the keyboard, which were assigned at random on each trial to either adjust the image towards the positive-color image or the negative-color image along the adjustment continuum. When the participant felt the image was greyscale they pressed the space bar. Participants were given 30 s between each trial to allow for the dissipation of the color aftereffects.

Each version consisted of eight adaptation trials (four for each of two images, of which one image would be upright and one upside down) presented in a pseudo-random order, with the constraint that trials presenting the same image could only appear a maximum of two times in succession. After having completed one version of the task (version A or version B), participants filled in the AQ and SRS-2 self-report questionnaires, then completed the second version. The order of versions was counter-balanced across participants.

The primary aim of the present study was to investigate whether color adaptation is reduced in autistic adults. Evidence for attenuated adaptation to faces (Ewing et al. 2013a, b; Fiorentini et al. 2012; Pellicano et al. 2013; Rhodes et al. 2014) and number (Turi et al. 2015) in autistic children may reflect a generalized reduction in adaptation in autism, as the hypo-priors account of autism (Pellicano 2013; Pellicano and Burr 2012) would predict. Contrary to predictions, there were no significant group differences observed in the strength of color afterimages following adaptation to whole-scene images in our sample of autistic adults, relative to a group of typical adults.

There are a number of implications of this finding, relating to the understanding of perceptual atypicalities in autism, and the interpretation of this study’s results. First, the results of the Bayesian analysis shows that the present study provides positive support for the null hypothesis (Dienes 2011; Raftery 1999) in relation to the effect of autism on basic adaptation magnitude. Similarly, the effect size associated with the group factor is very small, indicating a high degree of overlap in the distributions of adaptation strength across the autistic and typical groups. Therefore, we are confident that these findings are reliable.

Second, the basic phenomenon of color afterimages is attributable to adaptation very early on in the visual stream, at the level of photoreceptors and ganglion cells in the retina (Zaidi et al. 2012), whereas the effects of luminance contours which help to make the Spanish Castle effect so vivid reflect the contribution of areas of visual cortex modifying the afterimage signal received from the retina to create the perception of filling-in (Anstis et al. 2012; van Lier et al. 2009; Vergeer et al. 2015). In contrast, previous demonstrations of attenuated adaptation in autism have been for faces and number, which are processed at a higher level, in regions of the temporal (i.e., fusiform face area—Kanwisher et al. 1997) and parietal cortex (see Piazza and Izard 2009, for a review), respectively. In the current study, no general reduction in afterimage strength was observed in the autistic relative to the typical group. This may suggest that only higher-level adaptation is attenuated in autism, where priors and top-down influence may have more impact on perception, whilst low-level visual adaptation mechanisms in autistic adults are intact.

Third, the present study has tested autistic adults, while the majority of previous work demonstrating attenuated adaptation has been in children. For example, whilst children on the autism spectrum may have atypical adaptation aftereffects for facial identity and expression (Ewing et al. 2013b; Rhodes et al. 2014), afterimages in autistic adults are no different in strength to those of typical adults (Cook et al. 2014; Walsh et al. 2015). These latter results are, however, at odds with the finding that the parents of autistic children also exhibit reduced face aftereffects (Fiorentini et al. 2012), suggesting that attenuated adaptation of certain stimuli may also be reduced in genetically-related family members, and is manifest even in typical adults. It is important to note that while Fiorentini et al. found that the effect generalizes to adults who are relatives of autistic children, the effect is small. The evidence for reduced adaptation to numerosity (Turi et al. 2015) is also restricted to autistic children, with no similar study published on autistic adults.

There are relatively few studies which have examined adaptation in autistic adults. It remains possible that color adaptation is atypical in autistic children, even if not attenuated in adulthood. Children have a more limited experience of the world than adults and may not yet have developed compensatory strategies for any perceptual atypicality they experience. Indeed, some studies of typically developing children have found differences between children and adults in afterimages of facial identity and attractiveness, under some conditions (e.g., Hills et al. 2010; Jeffery et al. 2010, 2011). These studies suggest that children have more flexible norm-based coding of faces, due to their more limited experience of the world (see Jeffery and Rhodes 2011 for a review). Other studies have found afterimage magnitude is comparable between adults and children for face aftereffects (Anzures et al. 2009; Nishimura et al. 2008, 2011; Pimperton et al. 2009). The effect of age on color afterimages has not been widely researched. There is a suggestion that afterimages persist for longer in older people in the typical population (Kline and Nestor 1977) but whether the intensity of color afterimages changes with age is unknown. Further studies investigating adaptation in both the typical and autistic populations at different stages in development for a range of perceptual domains are needed to fully examine the question of whether age affects the intensity of adaptation afterimages.

Since we found no difference in adaptation, we fail to support the suggestion of suppressed retinal activity in the retina of autistic individuals, at least in adults (Ritvo et al. 1988). Likewise, there is no evidence that the purported differences in chromatic discrimination in autistic individuals (Cranwell et al. 2015; Franklin et al. 2008; Fujita et al. 2011; Heaton et al. 2008) has any effects on the responses made on this task. While it is the case that poorer discrimination may lead to more variability in achromatic settings, it is unlikely that this would represent a systematic shift of responses. Furthermore, differences in chromatic discrimination do not necessarily imply differences in adaptation. Finally, in adults, differences in chromatic discrimination may be associated with non-verbal ability [at least when measured using the Farnsworth-Munsell 100-Hue Test (FM100) (Cranwell et al. 2015)], whereas our groups are matched for Perceptual Reasoning Index (PRI), the non-verbal sub-scale of the WASI-II (see Table 1). The present study, it should be noted, cannot contribute to the question of chromatic discrimination in autism since the measure used is likely to be relatively insensitive and there are competing task demands related to nulling the adaptation afterimage.

The data from the present study show typical color adaptation in a group of autistic adults, contrasting to evidence for attenuated adaptation in other domains associated with autism in children. This may be due to the lower-level sensory nature of color adaptation compared to those domains in which adaptation has been shown to be attenuated in autism, or due to developmental changes in adaptation between childhood and adulthood. Further research should seek to clarify the role of top-down knowledge on color adaptation in both children and adults with and without autism. Another potentially fruitful approach may be to measure multiple forms of adaptation (e.g. color, motion, faces etc.), with the same task and procedure for each stimulus domain, within individuals in order to compare adaptation across domains and levels of representation.

Our secondary hypothesis related to a possible effect of scene content on color adaptation. Lupyan (2015) proposed that there are top-down effects of knowledge of a scene on the strength of adaptation and that natural scenes therefore have stronger afterimages than man-made scenes, which are not color diagnostic. While we observed that natural scenes elicited stronger afterimages than man-made scenes, and this effect was similarly strong across groups, we nevertheless failed to replicate Lupyan’s finding of an interaction of this effect with the orientation of the image. This interaction between orientation and image content is crucial to claims of top-down knowledge affecting afterimage magnitude, since effects of image content alone may be attributed to low-level properties of the different adapting images (greater luminance or chromatic intensity). Afterimages in our study were of equivalent strength regardless of whether the adapting image was upright or upside down. This lack of an interaction suggests that the difference in adaptation for type of scene may well be due to low-level differences in the images (e.g., chroma) rather than top-down effects of color diagnostic scene content. The lack of a verified top-down effect on color adaptation in our study means that, unfortunately, we are unable to determine whether effects of knowledge on adaptation are reduced in autistic adults, since we do not find such effects in the group of typical adults.

The failure to replicate Lupyan’s (2015) top-down effect may be due to a number of factors. The method and the stimuli were very similar to Lupyan’s and the samples were of a similar age (early adulthood). Nevertheless there were some methodological differences between studies, which may help explain the discrepancy between the findings. First, the after-effect was measured in different units across studies. Specifically, we quantified the after-effect using the perceptually meaningful measure of chroma rather than units in computational HSL space. However, when we analyzed our results in terms of HSL steps, we still found no effect of orientation and no orientation-diagnosticity interaction.

Third, the average time taken for participants to make the nulling adjustment is another potential factor. In the present study, we found no relationship between mean response times and the mean measured afterimage magnitudes (i.e. mean response times to all conditions are similar, see results). Lupyan (2015) reported an average adjustment time of just under 10 s, while our participants appear to have made their adjustments more rapidly, the average adjustment taking around 6 s. With time, the strength of the afterimage fades (e.g., Kline and Nestor 1977), it is therefore likely that Lupyan’s (2015) participants were making their achromatic settings for relatively weaker afterimages, compared to the participants in the present study. It is possible that these effects are somewhat sensitive to the time-course of adaptation—perhaps being more likely to be evident when the observer is adjusting to null a less intense afterimage (later) than when the afterimage is stronger (earlier). A weaker afterimage may cause the observer to be less certain of the bottom-up sensory input, and thus more likely to rely on top-down information in making their judgment.

Further research into these factors (within-participant orientation manipulations and response time) may help to clarify whether these aspects of the task design are crucial to identifying top-down effects. However, if it is to generate further interesting and testable research questions about top-down effects on visual perception, the effect should be generalizable beyond the specific parameters of the original study, to other stimuli, paradigms and designs. Our study adds a somewhat cautious view of the generalizability of Lupyan’s effect.

Finally, it is important to note that the Bayesian analysis confirms that the data provide only ‘weak’ (Dienes 2011; Raftery 1999) support for the null hypothesis relating to main effects of orientation and image type and the interaction between these factors. Considering Fig. 3, the aggregate data do appear to tend towards the expected direction of the effect—afterimage intensity is reduced by turning the image upside down to a greater extent for the diagnostic scenes than for the non-diagnostic. However, this trend does not appear at the individual level, with only a few participants’ individual data revealing this expected pattern. It appears from our study that the effect of top-down knowledge of afterimage strength, if present, is relatively subtle (see also Hansen et al. 2006; Olkkonen et al. 2008). The present task may not be the most sensitive test of these effects, and an alternative task which evokes stronger and more reliable effects should be sought to better test predictions relating to top-down influences on perception.

Other studies have shown the effect of prior knowledge or expectation on visual perception. An auditory label which is congruent with an image otherwise suppressed from visual awareness by inter-ocular rivalry can improve detection speed and sensitivity, and is specific to the object (e.g. hearing “pumpkin” speeds detection of pumpkins but not chairs) (Lupyan and Ward 2013). Similarly, when asked to adjust a familiar and prototypically-colored object to grayscale, observers appear to overcompensate for the influence of memory colors. For example, observers’ grayscale settings for a banana tend to be slightly biased in the blue direction, apparently to counteract the influence of the yellow memory color (Hansen et al. 2006; Olkkonen et al. 2008), although some of these top-down effects are notably small. As such, the brain appears to use prior experience of familiar objects with distinct colors to predictively code their appearance (Bannert and Bartels 2013). If such effects reflect top-down modulation of the cortical representation of those objects through the application of priors to current percepts, then we should find individual differences in the strength of these effects depending on autism symptomatology, age and experience with the objects in question. In order to fully explore this possibility, however, one must first find a perceptual task for which a reliable effect of top-down knowledge can be demonstrated in a typical group.