The task used in this study was designed to assess whether individuals are able to extract associative information at higher-levels of perception, through semantic or feature-based regularities. The results do not directly support the claims that autistic individuals may experience difficulties in either acquiring or generalising statistic regularities at the category-level. There have been suggestions that autistic individuals might develop overly specific priors which are not generalisable to other contexts [45] (Van de Cruys et al., 2014). Specifically, [53] Van de Cruys et al. (2017) claim that high precision of sensory information would lead to weaker high-level predictions at the expense of overfitted low-level ones. In relation to these heightened low-level predictions, they explain how this can lead to difficulties in processing more complex associations:
"While this happens at the expense of detecting more abstract regularities, note that the basic capacity of forming predictions remains unaffected. Rather, encoding of noise hampers discovery of regularities when these are embedded in more complex, noisy inputs." ([53] Van de Cruys et al., 2017)
We assessed this hypothesis at two separate levels, first by assessing whether information presented at a categorical level could be extracted by autistic individuals and then by testing whether information that was learnt in a specific context could be generalised to other contexts. The results reported here found that autistic participants, when compared to non-autistic controls, showed a modest but significant effect of reduced recall following statistical learning. This result was only present when data were considered across all 3 conditions of the task and this should be considered when interpreting the findings of this study.
As no interaction effect was found between the group and condition variables, there was no evidence to suggest that the autism group showed any specific difficulties in processing information at the category level or generalising information at this level. This suggests that the results do not directly support the hypotheses that autistic individuals experience difficulties in acquiring statistic regularities at higher-levels or difficulties in generalising prior expectations across different contexts.
As a significant overall difference in recall was found between our groups, when testing across all 3 task conditions, we can consider how this finding fits in with the wider literature. While a lack of group differences between autistic and neurotypical controls is common across the implicit learning literature, some studies have reported reduced performance in autistic individuals. Such results could be the result of other deficits in autism and not the implicit learning mechanisms per se. For example, motor abilities can be impaired in autism [54] (Fournier et al., 2010), which might explain reports of impaired implicit learning in motor-sequence tasks [55] (Gidley Larson and Mostofsky, 2008). [56] Foti et al. (2015) carried out a meta-analysis of implicit learning studies in autism and found no differences when combining the results of 11 different studies. This was followed up by an additional meta-analysis [57] (Obeid et al., 2016) which set out to build on and overcome the methodological issues of the study by [56] Foti et al. (2015). The concerns that [57] Obeid et al. (2016) raised regarding the results by 56] Foti et al. (2015) were due to the fact that the differences in statistical learning that they assessed were based on the extent to which reaction times reduced across blocks in which participants were presented with predictive sequences, rather than the more standard approach of comparing differences between blocks where predictive sequences occur and blocks where trials are random [58] (Nissen and Bullemer, 1987). However, [57] Obeid et al. (2016) still did not find any differences between the autistic and non-autistic participants in their study after correcting these methods.
A study that looked at visuospatial statistical learning in autism reported superior task performance in autistic adults but not in autistic children [59] (Roser et al., 2015), but it should be noted that the sample sizes used were very modest. [60] Jones et al. (2018) looked at visual statistical learning in a large sample of autistic children vs typical controls. When they compared performance across participants in the two groups, no differences were found. However, when they used a discriminant function to test the similarity of the autistic participants’ responses to the typically developing children’s responses, they found two distinct subgroups within the autism group. They also found that the autistic subgroup that performed similarly to the typically developing children tended to have reduced levels of autistic symptoms compared to the other autism subgroup.
The present study includes a large sample of participants and found evidence to suggest that autistic individuals do show reduced learning during our visual statistical learning task. If the results are taken as an example of reduced implicit learning in autistic individuals, it is important to note the specific nature of the task presented in the present study. The type of stimuli used in the tasks presented here should be highlighted, as the majority of previous studies looking at visual statistical learning in autism have used simple, abstract shapes. [38] Brady and Oliva (2008) found a series of results that suggest statistical learning can occur in typical individuals for scene-based stimuli, however this is the first study that we are aware of to use such stimuli in a group of autistic participants. The recognition and categorisation of scenes is thought to involve the processing of feature-based statistics which help us to infer the type of scene we are viewing [61] (Stansbury et al., 2013). The statistical regularities that occur across scene categories are thought to be processed at multiple levels of the visual systems hierarchical structure, acquiring information at both local and global levels [62] (Jun and Chong, 2016).
There is a body of evidence that shows that typically developed individuals are able to process and categorise the overall semantic content of scenes quickly but show limitations when detailed feature representation is required [63] (Fabre-Thorpe, 2011). Based on reports of a tendency of autistic individuals to focus on low-level details to a greater extent [16, 42, 64, 43] (Happé and Frith, 2006; Koldewyn et al., 2013; Mottron et al., 2000; Plaisted et al., 1999), it is possible that these differences would affect how semantic information is processed during rapid serial presentation tasks such as the one presented by [38] Brady and Oliva (2008). While differences in scene processing have been reported in autistic individuals, the stimuli used in these studies tends to include social features such as faces and therefore it is unclear whether such differences are specific to social scenes or extend to non-social scenes as well [65] (Wang et al., 2015). A better understanding of how non-social scenes are processed by autistic individuals, and whether any atypicalities do exist, would enable us to better interpret the results of the present study to determine whether the results are indicative of a more general deficit in visual statistical learning or whether the result is specific to the stimuli used.
It is important to highlight that high level perceptual differences in how sensory information is processed in autistic individuals should be viewed as differences in the style of perceptual processing rather than deficits. Instead of regarding autistic individuals as ‘better’ or ‘worse’ at aspects of perception, it can be helpful to think of them as having different processing styles that are beneficial in certain situations but detrimental in others. Indeed, suggestions of increased discrimination of similar stimuli [66] (Plaisted et al., 1998b) that are potentially associated with difficulties in category formation and generalisation [24, 25, 27] (Alderson-Day and McGonigle-Chalmers, 2011; Hartley and Allen, 2014, 2015; Plaisted, 2015) are also linked to the commonly reported occurrences of heightened perceptual sensitivity that lead to superior performance in autistic individuals compared to non-autistic individuals [9, 67, 68] (Shah and Frith, 1993; Plaisted et al., 1998a; O’Riordan, 2004). An important question to answer, should future studies find further evidence to suggest that autistic individuals show a reduction in the extent to which they extract statistical regularities from the environment, is whether the reduced reliance on expected properties of visual stimuli is directly linked to an increase in perceptual sensitivity.
Limitations
While the overall sample size of the full cohort was fairly large for a behavioural study, the sample sizes reduced when we stratified the full sample into smaller subsamples based on task condition and diagnostic status. Recruitment was conducted such that the full sample of participants comprised 6 different subsamples of participants, each of which contained only autistic or control participants who completed one of each of the 3 different task conditions. Significant effects of learning were found across the different subsamples of participants, except for the autistic participants who completed the generalisation condition. While this could be due to a lack of statistical power, as the sample sizes for these individual subgroups were modest, it does suggest that there was reduced recall performance for the participants in this subgroup relative to the other subgroups. This would be in line with the hypothesis that difficulties in generalising prior expectations may exist in autistic individuals.
While this is by no means a comprehensive assessment of the ability to generalise prior expectations, it does provide one example of an instance in which autistic individuals were found to process statistical information at a high level. However, it is important to note that this lack of a significant difference could be due to a lack of power. The specific hypotheses were assessed by considering only participants who completed the individual conditions within the task. This reduced the sample sizes considerably and may have been a factor in the fact that a main effect of group was found but there were no differences in any of the specific task conditions. Further investigation in this area needs to be carried out before drawing any firm conclusions.
It is also important to note that the interpretation of these results rests on the assumption that the approach used in the recall phase accurately captures the level of learning that occurred in the training phase. A 2-alternative forced-choice (2AFC) paradigm was used to assess participants memory for triplets in the recall phase [69] (Green and Swets, 1966). There are some criticisms of the 2AFC approach that suggest order effects can influence participants’ responses, but these issues are more problematic when testing for perceptual sensitivities rather than recall [70] (García-Pérez and Alcalá-Quintana, 2011). An alternative method for assessing recall is the yes/no paradigm [71] (Ahumada Jr and Lovell, 1971), but this approach could potentially be problematic when comparing autistic and non-autistic individuals due to differences in the interpretation of the task which could result in biases [72] (Hauck et al., 1998). Broadly speaking, there are a number of issues that can occur with explicit test phases at the end of implicit learning tasks, such as large amounts of noise when averaging across participants due to chance guessing and less power to detect true effects due to insufficient numbers of trials [73] (Siegelman et al., 2017). Nonetheless, when considered alongside the results from the serial reaction time task, the present finding seems to suggest that autistic individuals extract predictive information from their environment to a lesser extent than non-autistic individuals.
It is also important not to assume that the learning effects found in the category and generalisation conditions were based on semantic information, as it is possible that the higher levels of correlation between low-level features within same-category images compared to different-category images led to the observed recall effects [61] (Stansbury et al., 2013). While, [38] Brady and Oliva (2008) carried out experiments specifically to show that semantic information was processed during the task, this does not necessary extend to the sample in the present study. Indeed, it is a possibility that the two groups showed similar recall effects while actually processing different levels of information within the task. This possibility could be clarified by the inclusion of a word-based recall phase, as was included in the original study by [38] Brady and Oliva (2008). However, the present study was unable to include this due to the additional demand on data collection that the inclusion of further conditions would have resulted in. An alternative approach would be to assess whether within-category correlations of low-level features accounted for variation in the observed memorability of the different categories and, if so, whether this effect differed between the two groups [74-76] (Khosla et al., 2015, 2012; Squalli-Houssaini et al., 2018). This could be explored in future studies to build a more complete picture of the level at which autistic and non-autistic individuals process category-level information.