Quote: Isenstein EL, Park WJ, Tadin D (2021) Atypical and inflexible visual coding in autism spectrum disorder. PLoS Biol 19 (6): e3001293. https://doi.org/10.1371/journal.pbio.3001293
Released: June 8, 2021
Copyright ©: © 2021 Isenstein et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which allows unrestricted use, distribution, and reproduction in any medium, provided that the original author and source are acknowledged.
Financing: The author (s) did not receive any special funding for this work.
Competing interests: The authors have stated that there are no competing interests.
The fleeting understanding of autism spectrum disorder (ASD) primarily links the disorder to social and communication difficulties, as well as repetitive behaviors. However, differences in sensory perception are often reported in people with ASD, especially in the visual area  . In recent years there have been increasing efforts to elucidate the causes of these differences. The goal is not only to understand perceptual processing in ASD, but also to consider how these differences in sensory processing could have cascading effects on cognitive and motor functions  . These efforts face a major obstacle: perceptual processing is complex and involves several interconnected phases that use more than half of the human cortex. This creates a difficult credit allocation problem in which it is difficult to link an observed atypicality in perceptual processing to a particular component of the perceptual path. The work of Noel and colleagues in this issue ofPLOS biology an important step in meeting this challenge.
By and large, our perception of the world is the result of an “encoding-decoding” process. Coding describes how low-level sensory representations of the world are created by mapping sensory inputs to noisy and resource constrained neural representations. However, this is not the end of the exercise. Perception processing essentially depends on how these sensory representations are decoded based on various “rules”, including the combination of sensory measurements with prior knowledge (Fig. 1A). The result is an effective and flexible perceptual function that uses our extensive perceptual knowledge (Fig. 1B). In ASD research, much of the recent work has focused on changes in the decoding phase to explain sensory differences and how they translate into behavior in this population  . In contrast, Noel and colleagues used a combination of behavioral methods and computer modeling to isolate the early “coding” phase of perceptual processing.
Fig. 1. Perceptual coding and decoding.
(A) Medical imaging can be used as an analogy to perceptual coding and decoding. Coding is the mapping of sensory inputs onto neural representations, which, in our analogy, can be thought of as a CT machine that takes an image of a lung mass. A doctor’s interpretation of these images illustrates the perceptual decoding. The doctor’s assessment is influenced by previous knowledge (e.g. background, age and symptoms of the patient). Our perception is also strongly influenced by our many years of experience with our sensory environment. (B) In this analogy, the patient returns for a second CT. If the CT machine has highly precise coding, even a small increase in the measured size of the mass can indicate a significant change and the doctor will order a biopsy. If the CT machine has a low-accuracy coding, a small increase in the measured size of the mass may be within the margin of error and the doctor may not order a biopsy. Similarly, a highly precise perception-related coding leads to a rapid detection of changes in sensory statistics and to a suitable adaptation of the perception processing. However, such changes in sensory statistics can go undetected if the perceptual coding is poorly accurate. (C) Noel and colleagues found that individuals with ASD have poor coding precision of visual orientation (left), shown here as a stronger coding of 4 orientations (vertical, horizontal and 2 oblique; a noisy coding is seen as less pronounced differences between dark orientation signals and the surrounding pixels). After prolonged exposure to an artificial statistical distribution of orientations, neurotypic individuals showed a significant improvement in coding, which corresponds to the new statistic (top right), in contrast to the ASD group, which showed little change (bottom right). ASD, Autism Spectrum Disorder; CT, computed tomography.
To study the visual coding, Noel and colleagues used a simple orientation guessing task. Here the participants looked briefly at an oriented stimulus and then reported on the perceived orientation. This task shows a characteristic pattern of repulsive tendencies away from the cardinal orientations  . For example, a stimulus inclined by 15 ° from the vertical is perceived as inclined by 25 °. This distortion pattern has been associated with the “efficient coding” of common orientations in the natural environment, where the horizontal and vertical are overrepresented [4,5] . In the first block of the experiment, both neurotypical and ASA groups showed this characteristic bias pattern. However, people with ASD had a 60% higher variance in their orientation estimates. This result agrees with previous work that showed a stronger sense of orientation in ASD. showed  . The key advance of the present study is the use of bias and variance data to compute the “fisherman information” – how much information the participants’ responses contain about actual stimulus orientation. Here the authors relied on a well-known property of an estimator in statistics – the “Cramer-Rao limit” – which gives us the best performance we can ever expect from an unbiased estimator. Simply put, poor stimulus coding results in highly variable stimulus estimates. When the estimator is skewed, as is the case with perception of orientations, this limit establishes a relationship between Fisher information and the two things an experimenter can measure: bias and variance. The most important finding is that this relationship is “independent” of the differences in the decoding scheme and allows the experimenter to isolate the coding capacity.
As shown by the increased variance in response, subjects with ASD had lower baseline coding capacity than neurotypic controls (Figure 1C, left). Fisher information for both groups peaked at the cardinal points. The neurotypical group, however, had an overall coding capacity of about 30% higher. This finding is consistent with louder, more variable sensory processing in ASA that was reported in both neuroimaging  and behavioral work  (but see  ).
The second important finding provides insights into how people with ASD adapt to changes in stimulus statistics. Under the efficient coding hypothesis, visual neurons are optimized to maximize the information they carry about the natural environment  . Therefore, given changes in the visual input statistics, the hypothesis predicts that the visual system will reallocate coding resources to accommodate the new statistics. The present study provided the participants with increasing experience with an artificial environment in which, in contrast to natural statistics, the distribution of orientations was uniform. As expected, in the neurotypic group, the participants’ coding capacity increased over the course of the experiment and was reallocated to better match the even distribution of orientations used in the experiment. However, people with ASD showed no change in coding strategy (Fig. 1C, right). As a result, the group difference in coding capacity increased from 30% to 50% at the end of the experiment.
What are the implications of this pair of findings? Reduced perceptual coding could lead to less reliable and less predictable sensory representations of the environment. This can affect how people with ASD interact with the sensory world [1,3,6] . The finding of inflexible adaptation to a new sensory environment also suggests that people with ASD may need additional time to adjust to new circumstances, a concept that has been reported for the behavioral inflexibility and adherence to routine that is common in this group is central. These results are consistent with other work that has identified challenges in updating expectations based on new circumstances at ASD [9,10] . We emphasize that these interpretations of the possible cascading effects of atypical sensory coding are speculative. Furthermore, the empirical results of Noel and colleagues, with large differences between the groups, while quite convincing, are not without qualifications. The main open question is to what extent these results generalize to the entire perceptual function – this topic requires further investigation. The data on cognitive function in ASD are quite heterogeneous and include examples of both decreased and improved cognition  . For example, one theory postulates a highly precise but inflexible coding in ASD  this apparently contradicts the current results. Technical questions must also be taken into account. In the work of Noel and colleagues, the stimulus disappeared before participants gave their answers. Therefore, processes that may not be related to sensory coding, such as perceptual memory, may play a role in the reported results.
An additional obstacle in finding the root causes of behavioral differences is to distinguish the atypicals, which are fundamental to ASD, from the compensatory ones. Noel and colleagues provide a good example here. The finding of an inflexible fit to the stimulus statistic may seem basic, but the authors show that this behavior correlates with the initial differences in coding capacity. One explanation for this result is that someone with poor sensory coding would have greater difficulty recognizing a change in the sensory environment and consequently delay the use of suitable adaptation mechanisms (Fig. 1B). We suspect that there are analogous explanations for many of the symptoms reported with ASD, where the observed differences do not reflect abnormal brain function but rather the brain’s adaptive response to something else that is atypical. A better understanding of this distinction will likely be critical in assessing the impact of atypical and inflexible visual coding on clinical behavior in ASD, as well as in identifying more specific targets for interventions.