ArticlePDF Available

Abstract

There have been numerous attempts to explain the enigma of autism, but existing neurocognitive theories often provide merely a refined description of 1 cluster of symptoms. Here we argue that deficits in executive functioning, theory of mind, and central coherence can all be understood as the consequence of a core deficit in the flexibility with which people with autism spectrum disorder can process violations to their expectations. More formally we argue that the human mind processes information by making and testing predictions and that the errors resulting from violations to these predictions are given a uniform, inflexibly high weight in autism spectrum disorder. The complex, fluctuating nature of regularities in the world and the stochastic and noisy biological system through which people experience it require that, in the real world, people not only learn from their errors but also need to (meta-)learn to sometimes ignore errors. Especially when situations (e.g., social) or stimuli (e.g., faces) become too complex or dynamic, people need to tolerate a certain degree of error in order to develop a more abstract level of representation. Starting from an inability to flexibly process prediction errors, a number of seemingly core deficits become logically secondary symptoms. Moreover, an insistence on sameness or the acting out of stereotyped and repetitive behaviors can be understood as attempts to provide a reassuring sense of predictive success in a world otherwise filled with error. (PsycINFO Database Record (c) 2014 APA, all rights reserved).
Precise Minds in Uncertain Worlds: Predictive Coding in Autism
Sander Van de Cruys, Kris Evers, Ruth Van der Hallen, Lien Van Eylen,
Bart Boets, Lee de-Wit, and Johan Wagemans
KU Leuven
There have been numerous attempts to explain the enigma of autism, but existing neurocognitive theories
often provide merely a refined description of 1 cluster of symptoms. Here we argue that deficits in
executive functioning, theory of mind, and central coherence can all be understood as the consequence
of a core deficit in the flexibility with which people with autism spectrum disorder can process violations
to their expectations. More formally we argue that the human mind processes information by making and
testing predictions and that the errors resulting from violations to these predictions are given a uniform,
inflexibly high weight in autism spectrum disorder. The complex, fluctuating nature of regularities in the
world and the stochastic and noisy biological system through which people experience it require that, in
the real world, people not only learn from their errors but also need to (meta-)learn to sometimes ignore
errors. Especially when situations (e.g., social) or stimuli (e.g., faces) become too complex or dynamic,
people need to tolerate a certain degree of error in order to develop a more abstract level of represen-
tation. Starting from an inability to flexibly process prediction errors, a number of seemingly core deficits
become logically secondary symptoms. Moreover, an insistence on sameness or the acting out of
stereotyped and repetitive behaviors can be understood as attempts to provide a reassuring sense of
predictive success in a world otherwise filled with error.
Keywords: autism spectrum disorder, predictive coding, uncertainty, adaptive control, learning
[Funes] was disturbed by the fact that a dog at three-fourteen (seen in
profile) should have the same name as the dog at three-fifteen (seen
from the front). His own face in the mirror, his own hands, surprised
him on every occasion. . . . He was the solitary and lucid spectator of
a multiform world which was instantaneously and almost intolerably
exact....Hewasnotvery capable of thought. To think is to forget
a difference, to generalize, to abstract. In the overly replete world of
Funes there were nothing but details, almost contiguous details.
—Jorge Luis Borges, 1942
Autism spectrum disorder (ASD) refers to a group of neuro-
developmental conditions with an early onset and characterized
by sociocommunicative impairments and stereotyped, restricted
behavior patterns and interests (American Psychiatric Associa-
tion, 2013). Although ASD has a strong polygenetic component
with heritability around 70% (Geschwind, 2011), no biological
marker is available yet, and, thus, diagnosis mainly relies on
behavioral assessment. The prevalence of ASD is estimated to
be 1%, with males being more affected than females (Baird et
al., 2006;Pinborough-Zimmerman et al., 2012). ASD is asso-
ciated with increased comorbidity for other disorders (e.g.,
attention-deficit/hyperactivity disorder, anxiety disorders, tic
disorders, learning disabilities and epilepsy; J. L. Matson &
Nebel-Schwalm, 2007). In addition, a significant proportion of
the ASD population is intellectually disabled (Elsabbagh et al.,
2012).
The neurocognitive frameworks put forward to account for
behavioral symptoms in ASD can be broken into two groups,
depending on which symptoms are considered to be central and
preceding the others. Social first theories put problems with social
cognition or motivation front and center. The most prominent
contender is the theory of mind framework (Baron-Cohen et al.,
2000). It focuses on the social problems and argues that the core
deficit lies in the understanding of the behavior of others in terms
of one’s own underlying mental states. Nonsocial theories, on the
other hand, consider general cognitive or perceptual problems to
be the primary causal factor. Among them, the weak central
coherence theory (WCC; Frith & Happé, 1994;Happé & Booth,
2008;Happé & Frith, 2006) and the enhanced perceptual func-
tioning theory (EPF; Mottron & Burack, 2001;Mottron, Dawson,
Soulières, Hubert, & Burack, 2006) focus on the perceptual pecu-
Sander Van de Cruys, Laboratory of Experimental Psychology, KU
Leuven; Kris Evers and Ruth Van der Hallen, Laboratory of Experimental
Psychology, Leuven Autism Research, and Department of Child Psychia-
try, University Psychiatric Center, KU Leuven; Lien Van Eylen, Leuven
Autism Research and Parenting and Special Education Research Unit, KU
Leuven; Bart Boets, Leuven Autism Research and Department of Child
Psychiatry, University Psychiatric Center, KU Leuven; Lee de-Wit, Lab-
oratory of Experimental Psychology, KU Leuven; Johan Wagemans, Lab-
oratory of Experimental Psychology and Leuven Autism Research, KU
Leuven.
This research was supported by a Methusalem grant by the Flemish
Government (METH/08/02) to Johan Wagemans. We are incredibly grate-
ful to John Brock, Floris de Lange, Vebjörn Ekroll, Marie Gomot, Jakob
Hohwy, Ilona Kovacs, Laurent Mottron, Colin Palmer, Jean Steyaert,
Jeroen van Boxtel, Peter van der Helm, and Raymond van Ee for support
and invaluable feedback on a draft of this paper and to our colleagues from
Leuven Autism Research for many interdisciplinary discussions.
Correspondence concerning this article should be addressed to Sander
Van de Cruys, Laboratory of Experimental Psychology, KU Leuven,
Tiensestraat 102, 3000 Leuven, Belgium. E-mail: sander.vandecruys@
ppw.kuleuven.be
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
Psychological Review © 2014 American Psychological Association
2014, Vol. 121, No. 4, 649– 675 0033-295X/14/$12.00 http://dx.doi.org/10.1037/a0037665
649
liarities in ASD and argue for a locally (as opposed to globally)
oriented processing style in individuals with ASD. Accounts that,
prompted by the symptom cluster of repetitive and inflexible
behavior patterns, situate the core deficit in an executive dysfunc-
tion (e.g., E. L. Hill, 2004) also belong in this group of nonsocial
theories.
These theoretical frameworks are not mutually exclusive but
focus on different behavioral symptoms. Each theory has been
highly influential in shaping the fieldand has shifted the research
and clinical focus from an exclusively descriptive behavioral ap-
proach toward an enhanced desire to understand the atypical
neurocognitive mechanisms in ASD. Nevertheless, serious limita-
tions with these frameworks have become evident over the years.
First, although local processing styles, theory-of-mind difficulties,
and executive problems are common in ASD, they are neither
specific to the disorder nor apparent in all cases. Second, although
they are called neurocognitive, these mechanisms do not readily
connect to underlying neural mechanisms, except in terms of broad
networks of neural activation associated with each domain of
function. Part of the problem is a lack of specificity in the proposed
cognitive mechanism. Finally, although each of these frameworks
attempts to incorporate more than the symptom cluster or behavior
on which it is based, this often seems contrived, precisely because
each theory is too closely intertwined with the cluster of symptoms
in question.
We argue that the way in which individuals with ASD process
and respond to errors (or violations to their predictions) provides
an excellent candidate for a primary dysfunction that, when viewed
in the context of a complex developmental trajectory, provides a
mechanistic explanation for the different symptoms of ASD. This
imbalance in the brain’s handling of prediction errors could result
from different genetic and neurophysiological pathways, thus
highlighting that different pathogenetic factors could in fact con-
tribute to a common information processing imbalance (Ge-
schwind, 2011).
We structured the current article as follows. In the first
section (The Anticipating Brain) we briefly introduce the pre-
dictive coding framework as it originated from perception re-
search but evolved into a unifying theory of brain functioning.
In the second part (Predictive Coding in ASD) we propose a
specific etiological mechanism for ASD, which is then applied
to the different symptom clusters and clinical observations in
ASD. Because of the developmental nature of the disorder, we
start with a discussion of exploration and development (Devel-
opment and Exploration). Next, we discuss how perceptual and
cognitive alterations in ASD can originate from our theory
(Cognitive Functioning). In the subsequent sections, sensori-
motor and affective consequences are covered (Sensorimotor
Abilities and a Sense of Self and Chronic Unpredictability and
Its Affective Consequences). In the Social Functioning section,
core principles from earlier sections come together to explain
problems in social functioning in ASD. Then, we briefly con-
sider possible neural substrates of the proposed cognitive deficit
(Neurobiological Underpinnings). Before reaching our conclu-
sions, we cover a few related accounts of ASD to discuss
commonalities and indicate the added value of our approach
(Related Approaches).
The Anticipating Brain
Prediction is central for adaptive, intelligent systems (Hawkins
& Blakeslee, 2004). It allows us to efficiently prepare for imping-
ing circumstances that may foster or threaten continued subsis-
tence. However, prediction-based computations can only succeed
when there are in fact reasonably predictable contingencies in the
world. Prediction, therefore, depends upon an animal’s sensitivity
to statistical regularities in the environment and in its interaction
with the environment. Some of this structure is readily available,
and other parts are accessible only through higher order correla-
tions. Our understanding of the role of predictions in shaping
information processing has recently taken a step forward through
the development of predictive coding models (Clark, 2013b). This
computational scheme is heavily inspired by perception-as-
inference (von Helmholtz, 1910/1962) or perception-as-hypothesis
(Gregory, 1980) ideas, which assume that the brain continually
generates predictions on what input comes next based on current
input and learned associations. Predictive coding, however, does
not just stipulate that predictions are generated. It slso stipulates
that these predictions are compared (at many levels of the system)
to incoming sensory input and that the comparison leads to the
computation and representation of an error signal. These prediction
errors are important, because they signal that the current generative
model of the world—the one used to generate the currently best
prediction—is not up to the task of explaining (predicting) the
world. Once a prediction error has been signaled, the system still
has to employ some degree of flexibility in deciding what do to
with that error signal. In an uncertain world, experienced via an
inherently noisy biological processing system, errors will some-
times be spurious and uninformative. Thus, although prediction
errors should sometimes be taken very seriously in updating one’s
predictive model, it is also critical that some prediction errors are
essentially ignored. It is in the imbalance between these options
that we think the symptoms of ASD are to find their cause.
In terms of neural architecture, predictive coding assumes a dual
computational role for every level of processing (Egner, Monti, &
Summerfield, 2010). Representation or prediction units compute
predictions that are fed back, while prediction error units compute
the difference between sensory input and top-down prediction.
These prediction errors then serve as feed-forward input for the
next level. The biological plausibility of this specific architecture
is still under investigation, but the importance of prediction errors
and predictive processing in the brain in general is well estab-
lished. Predictive coding can account for fundamental stages of
perceptual processing, such as the emergence of extraclassical
receptive field effects measured with single cell recordings in the
primary visual cortex (Rao & Ballard, 1999). It can also account
for the complex dynamics between predictions made and input
received at very different stages of the system (den Ouden,
Daunizeau, Roiser, Friston, & Stephan, 2010). Furthermore, it can
explain neural dynamics such as the apparent adaptation to pre-
dictable stimulus contingencies (Summerfield, Monti, Trittschuh,
Mesulam, & Egner, 2008). Finally, there is some evidence for the
existence of separate representations for input and error signals in
the recent discovery of differential sensitivity to predictable stim-
uli in separate clusters of voxels in the fusiform face area (de
Gardelle, Waszczuk, Egner, & Summerfield, 2012).
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
650 VAN DE CRUYS ET AL.
The computational scheme of predictive coding is assumed to
repeat on every level of the perceptual hierarchy (Diuk, Tsai,
Wallis, Botvinick, & Niv, 2013;Wacongne et al., 2011). Each
higher level can capture a higher order regularity in input, relating
events spanning more time or space, because it can work on the
representational “language” of the previous level. Perceptual in-
ference is guided in a top-down way through higher level, con-
ceptual predictions that can be passed downward, generating a
chain of interdependent predictions to match on different levels,
from complex features to low-level stimulus characteristics.
Formally, predictive coding is equivalent to Bayesian inference with
the priors replaced by predictions and sensory evidence replaced by
prediction errors, reflecting the mismatch between the input and the
predictions. However, the differences between these two related ap-
proaches have important implications. A first distinction from
Bayesian approaches concerns the more specific claims about the
neural implementation of predictive coding. Second, replacing
sensory evidence by prediction errors emphasizes that incoming
information is put in context from the very start. The information
immediately becomes input relative to the organism, its models of
the world, and its current state. This approach also emphasizes that
processing does not start with the onset of a stimulus. Preexisting,
intrinsic activity of the brain is considered formative, as it reflects
the continuous predictive activity of the proactive mind-brain (Bar,
2009). Another advantage of predictive coding is that it allows a
natural connection to other neurobehavioral domains, where pre-
diction errors are known to play a crucial role, like midbrain
dopaminergic processing of reward (Schultz, Dayan, & Montague,
1997), hippocampal processing for contextual memory (Honey,
Watt, & Good, 1998), and amygdalar processing for fear learning
(Boll, Gamer, Gluth, Finsterbusch, & Büchel, 2013). This suggests
we may be a step closer to a general theory of the brain as a
prediction engine in which prediction errors emerge as the lingua
franca of neural information processing (den Ouden, Kok, & de
Lange, 2012).
Critically to our theory of ASD, predictive coding operates on
two time scales (Dayan, 2012;Friston, 2010). Predictions are used
here-and-now to shape one’s online estimation of the state of the
world (albeit through an iterative process), but the resulting pre-
diction errors also shape plasticity and learning over longer time
scales. In this way, today’s prediction errors shape tomorrow’s
predictions (paraphrasing a famous Bayesian dictum). Because the
world is not static, predictable contingencies that used to hold can
change, and predictive coding has to track these dynamics. No two
experiences are ever completely the same; thus, prediction error
will always be present to some degree. However, the brain has no
direct, independent means of differentiating mere noise from ac-
tual changes in the world (Feldman, 2013). It is, therefore, critical
that predictive coding incorporates a mechanism to flexibly alter
the extent to which the prediction errors generated by online
estimation affect future learning and plasticity.
A solution to this can be found in terms of a flexible adjustment
of what Friston (2010) described as the precision of the prediction
errors. To explain precision, one can draw the parallel with the
means comparison in a ttest, in which the numerator represents the
prediction error, which is weighted by the estimated standard error
(precision or confidence; Friston, 2009). As in the ttest, precision
is not given in perceptual inference but has to be estimated as well.
In an optimal system, precision has to increase when there still are
learnable regularities in the environment and decrease when it is
estimated that remaining deviations can be attributed to noise that
is unlikely to repeat in next instances or to other irreducible
uncertainties in input. Distinguishing between irreducible and re-
ducible uncertainties is a fallible process, relying on complex
meta-predictions for a given context. The system, therefore, has to
attribute a value or weight to prediction errors in order to deter-
mine to what extent they should induce new learning. The role of
precision is conceptually the same as that of the learning rate
parameter in Rescorla–Wagner learning (see Courville, Daw, &
Touretzky, 2006;O’Reilly, 2013, for a full discussion on learning
in volatile environments). Setting precision consequently relies on
a form of meta-learning: learning what is learnable (Gottlieb,
2012) or estimating the predictability of new contingencies. It is
clear from all of this that precision should be a context-sensitive
measure, to be flexibly optimized dependent on the current class of
input and the state of an organism. Indeed, precision is assumed to
be the mechanism of attention within predictive coding. At its
core, attention is the process of deciding where to look next in
order to allocate resources to that information with the highest
value, understood precisely as input containing reducible uncer-
tainty (Dayan, Kakade, & Montague, 2000;Gottlieb, 2012). Neu-
rally, precision is assumed to be represented by the gain of
bottom-up neural units representing the prediction errors, probably
mediated by neuromodulators (Friston, 2009; see the Neurobio-
logical Underpinnings section).
From this brief overview it should be apparent that predictive
coding provides a framework that allows us to go beyond unidi-
rectional views of information processing. Bottom-up information
streams (predictions errors) are inherently dependent on top-down
influences (predictions), which in their turn are shaped by previous
prediction errors. This complex interplay also means that the
dysfunction of one will automatically have consequences for the
other. Disturbances in the relative contribution of top-down versus
bottom-up information flow have been at the heart of two influ-
ential cognitive theories of ASD, representing apparently diamet-
rically opposing positions (WCC and EPF). A predictive coding
approach provides a principled and refined view on the influence
of top-down versus bottom-up processes and their complex inter-
play.
Predictive Coding in ASD
To bring into focus what we believe is the core processing
deficit in ASD, we have to emphasize again the distinction be-
tween reducible and irreducible uncertainty (prediction errors).
Irreducible uncertainty is due to the inherent stochastic nature of
the world and the inherently noisy biological apparatus with which
we sample from that world. Differentiating between reducible and
irreducible uncertainty requires an estimation of expected uncer-
tainty based on previous prediction errors (Preuschoff & Bossaerts,
2007;Yu & Dayan, 2005). If, through learning, you estimate the
outcomes of a stochastic process to vary with 3 (hypothetical)
units, a prediction error of 2 should not surprise you and therefore
should not urge you to update your model (prediction). When the
size of a prediction error is smaller than the expected variability
(based on past prediction errors) for this event, the current predic-
tion error should be scaled down. Reducible uncertainty, on the
other hand, is present when associations in the world (or our
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
651
PREDICTIVE CODING IN AUTISM
interaction with it) are not yet fully learned. The latter is some-
times called unexpected uncertainty. This is the case when previ-
ously predictive cues have changed and become invalid, so a real
update of the model is necessary. More formally, it is about
situations in which correlations between predictions and prediction
errors have changed.
In relatively unambiguous situations, people with ASD can
successfully learn and apply new contingencies (M. Dawson,
Mottron, & Gernsbacher, 2008). Problems arise, however, when
the predictive value of learning cues changes (i.e., in volatile
environments). For that reason, we situate impairments in ASD in
meta-learning: learning that cues of all present stimuli are learn-
able (i.e., can reliably predict future situations relevant for the task
at hand). This meta-capacity, estimating for which cues predictive
progress can be made, allows typically developing (TD) individ-
uals to distinguish random variability in input from actual, learn-
able changes in environmental regularities. Here, we advocate that
individuals with ASD overestimate the amount of changes in
environmental regularities, because they give too much weight to
their prediction errors.
Another way to conceive of this meta-learning capacity is in
terms of knowing where gains can be made in predicting the world.
If you know where predictive progress can be made, you know
which prediction errors matter and, hence, which prediction errors
should be assigned high precision. Precision is the mechanism of
attention in predictive coding because in this way it affects the
further sampling of the sensory world. Atypical attention happens
to be among the earliest signs of ASD, described in terms of the
flexible and appropriate assignment of salience to stimuli (Elison
et al., 2013;Zwaigenbaum et al., 2005). In ASD, the atypical
distribution of attention has been attributed to slower encoding,
which is consistent with the thesis that too many resources are
invested in sensory processing because precise prediction errors
cannot be discounted and thus attract further processing.
Hence, deriving our model from a general theory of information
processing, predictive coding, and our analysis of what could be
the key problems in ASD, we situate the core deficit in the high,
inflexible precision of prediction errors in autism (HIPPEA). Low-
level sensory prediction errors are generally set at a level of
precision that is too high and independent of context (Palmer,
Paton, Hohwy, & Enticott, 2013;Van de Cruys, de-Wit, Evers,
Boets, & Wagemans, 2013). As mentioned before, it is useful to
consider the consequences with regard to online inference versus
those regarding learning separately. If prediction errors during
online inferences get an unduly high precision, these will urge new
learning for every new event. The predictions that result from this
learning will be shaped by noise that is unlikely to repeat in the
future; hence, these predictions will almost never be applicable. In
neural network learning studies, overfitting takes place when er-
rors for the training set are reduced to an exceedingly low level
(Bakouie, Zendehrouh, & Gharibzadeh, 2009). It is a suboptimal
form of learning because new data (acquired with each new
experience) will generate large errors, meaning that there is little or
no generalization. If errors are always deemed important, every
new instance will be handled as an exception, different from
previous experiences. In the long run, however, those affected by
this dysfunction may succumb to a sort of learned helplessness: too
much learning with no fruits. This may have an especially demo-
tivating effect on particularly noisy interactions, such as those
involved in social situations (see the Social Functioning section).
With regard to the consequences for online inference and be-
havior, we have to distinguish situations in which an exact match
from cue to prediction exists and is functional from situations in
which exact matches will rarely happen or are even dysfunctional.
In the case of exact matching, it is well known that people with
ASD cope incredibly well (Mottron et al., 2013). They often excel
in rigid, exact associations (rote learning). Here, their overfitted
predictions serve them perfectly well, precisely because they suffer
less from interference from similar instances. They seem to trade
off the ability to generalize with a more accurate memory. Hence,
according to HIPPEA, the core processing deficits in ASD become
most evident when some disregard for details and some general-
ization are needed. Generalized inferences are required in situa-
tions where exact matches are not present, which is the rule rather
than the exception in natural situations, especially those involving
social interactions.
In everyday life, multiple cues impinge simultaneously on an
individual. At first exposure this may cause sensory overload,
because selectivity is lost when the informational (predictive)
value of cues cannot be estimated immediately. Predictions are
tested but violated because they are based on spurious correlations.
Individuals with ASD may cope with perceived repeated changes
in contingencies by executing prepotent, impulsive, or model-free
behaviors, described as repetitive, stereotyped behaviors in the
ASD symptomatology (for a discussion on the role of precision in
arbitrating between model-free and model-based behavior, see
Clark, 2013a;Daw, Niv, & Dayan, 2005). In a second stage,
individuals with ASD may “give up” and select cues just to evade
and cope with prediction errors. On their own scale, these cues
may be highly predictable, even though they are not functional in
the situation at hand. Thus, attention and behavior become domi-
nated by one or a few cues (cf. stimulus overselectivity; Lovaas,
Koegel, & Schreibman, 1979), singled out seemingly arbitrarily.
Note that computing prediction errors as such is not impaired in
ASD according to this view. Individuals with ASD can still com-
pare their predictions with actual input. These prediction errors,
however, have to be weighed in accordance to an estimation of
their reliability; that is, the extent to which they are caused by
learnable (changes in) regularities. Attesting to the fact that pre-
diction error computations are intact in people with ASD, their
detection and discrimination performance seems to be similar to
that in typically developing individuals, if not superior (see the
Perceptual Processing section).
Of importance, one can distinguish between two mechanisms
that both can result in inflexibly high precision of prediction errors.
First, it is possible that the neural mechanism for precision is
directly affected in ASD, fixing precision at a high level and
preventing meta-learning (which may take place anyway) from
having an effect on perception and learning. Aberrant neuromodu-
latory mechanisms of precision, as discussed in the Neurobiolog-
ical Underpinnings section, may be responsible here. Second, the
meta-learning prior to the setting of precision may be deficient in
ASD and hence does not provide the needed basis for proper,
context-dependent estimation of precision. Neural regions and
mechanisms that may be central for this capacity are discussed in
the Neurobiological Underpinnings section.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
652 VAN DE CRUYS ET AL.
In a nutshell, HIPPEA consists of the following basic premises.
The starting point is a high and inflexible estimation of precision
of prediction errors in ASD. This indicates that meta-learning is
deficient or short-circuited. Indiscriminately high precision will
mean that unrepeated, accidental variations in the input receive
disproportionate weight. This in turn prevents abstract representa-
tions from being formed, because matching will continue on a
more specific level, closer to the input. Indiscriminately high
precision also induces superfluous learning, leading to narrowly
defined, lower level predictions and incomplete hierarchical mod-
els. Finally, indiscriminately high precision entails a loss of au-
tonomous, flexible attentional selection based on informativeness
(deciding what information to sample based on the different types
of uncertainty in input).
HIPPEA thus situates problems in ASD at the intersection of
perception, attention, learning, and executive functioning (adaptive
control). Further key symptoms of ASD may emerge from this
impairment, but this will be fleshed out in the sections below. We
argue for an impairment in general information processing rather
than in one single domain (e.g., social cognition), supported by the
fact that problems in ASD are not limited to one such domain but
are pervasive. However, this also puts the burden of explanation
with us as to why some domains (specifically the social) would be
affected more than others (see the Social Functioning section).
Development and Exploration
The meta-learning deficit in HIPPEA is very consistent with the
developmental nature of the disorder. The very process of moving
from one simpler developmental stage to the next, more complex
one is impaired when an organism cannot estimate where predic-
tive progress can be made. If any prediction error is deemed as
valuable as the next, an inappropriate lingering on stimuli is
expected to occur. As a result, the kind of exploration that opti-
mizes learning is lacking, because estimating where predictive
progress can be made helps an organism to avoid the large regions
of input space that cannot be learned (fully) and those that are too
difficult at this stage of development. In short, this principle gives
a rationale for the importance of intermediate levels of complexity
in development (Berlyne, 1966;Gibson, 1969;Oudeyer, Baranes,
& Kaplan, 2010). If predictive gain can be properly estimated,
exploration can be guided such that it is aimed at regions with a
difficulty just above current ability, which leads to discernible
progressive stages of increasing complexity, as modeled in devel-
opmental robotics (Oudeyer, Kaplan, & Hafner, 2007). Particu-
larly in noisy, variable environments the mechanism can be ex-
pected to realize more efficient learning. It is easy to see that if this
capacity for active exploration is missing, as we think is the case
in ASD, an individual has to rely much more on the scaffolding
provided by caregivers, explicitly guiding progression from simple
to more naturalistic situations. Apart from prenatal genetic and
neural components, differing degrees of this environmental scaf-
folding may account for heterogeneity in symptom severity and
developmental trajectories in ASD.
The link between prediction violation and exploration is ele-
gantly illustrated in a study by Legare (2012). The study investi-
gated how TD children explain evidence violating their predictions
and illustrated how this mechanism may shape development. Dif-
ferent shapes were put on top of boxes that could light up,
depending on the shape, and those shapes that caused the box to
light up were subsequently labeled as a “blicket.” Children were
then confronted with a violation of the established prediction (no
light for a blicket), and Legare asked them to explain what had
happened. She could distinguish two main types of explanations;
about half of the children tried to explain why the block did not
light up (e.g., no batteries, block was not placed properly), and a
third of the children explained the situation by referring to the
category membership (e.g., “It’s not really a blicket; it only looks
like one”). Most interesting, however, the kind of explanation
children gave predicted the way they played with the objects later
on. Although children who gave a causal explanation explored the
objects more thoroughly, testing different combinations and ex-
perimenting with the placement and orientation of objects to find
out what would happen, children who explained inconsistency in
terms of the categories primarily went about sorting the objects
into two categories based on what had happened when they first
placed the objects on the box. This sorting behavior was a less
sophisticated form of exploration and was less likely to foster deep
understanding of the underlying sources of inconsistencies.
Arguably, the difference hinges on the ability to model uncer-
tainty in associations in the input. This modeled uncertainty be-
comes a handle to dissect underlying causes. The precision of
low-level inconsistency can, with a proper model of uncertainty,
be down-regulated such that the general rule (prediction) is not
violated and so does not have to be abandoned. Rather, modeling
uncertainties in the task opens the door to contextual modulations
of the general rule, which are always at play in practice. When
uncertainty is not accounted for and precision is continuously high,
every minor violation will induce new learning. An inconsistent
finding is categorized anew or is considered a special case unlike
previous instances. The latter is what HIPPEA proposes to be the
case in ASD. Though Legare’s (2012) study included only TD
children, the sorting behavior found in spontaneous play for the
subset of children that gave noncausal explanations is reminiscent
of what is observed in autistic play. Her results show that whether
and how people explain prediction error is linked to the kind of
exploration in which they will engage. In our line of thinking, the
difference already emerges in the way people process perceptual
input that runs counter their predictions, and this may have far-
reaching consequences for exploration and further development,
notably with regard to finding out about why the world functions
as it does.
Considering this change to the nature of exploration in ASD, it
is informative to revisit the so-called dark-room problem within
predictive coding (Friston, Thornton, & Clark, 2012;Froese &
Ikegami, 2013). This problem arises because if, as the fundamental
thesis of predictive coding has it, an organism acts to minimize the
prediction errors it experiences, the simplest solution would be to
seek out a dark room, devoid of prediction errors. Nevertheless,
most organisms venture out of their black boxes and explore the
world. The most obvious way to counter this is by noting that
generalized predictive coding involves not only learned mental
models and perceptual predictionsbut also bodily predictions, pre-
dictions embodied by the very structure of the body, homeostasis,
biomechanics, and the “gross initial neural architecture of the
agent” (Friston et al., 2012). Evolution equips organisms with a
limited set of expected states (cf. homeostasis) that is compatible
with their continued existence (survival). A dark room will not
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
653
PREDICTIVE CODING IN AUTISM
remain a low-prediction-error environment, for instance, when
food is not available.
One means of ensuring that organisms venture out to meet their
needs for survival is to equip them with an ability to tolerate the
prediction errors associated with new, unpredictable environments.
Organisms can adjust the precision of prediction errors based on
the expected volatility of their environment. If precision of low-
level prediction errors is overly high, however, the organism may
very well prefer to stay in a dark, room-like environment. In fact,
the typical autistic state of stereotypic self-stimulation and indif-
ferent withdrawal from the world and from others can be regarded
as “abnormal yet effective ways of reducing prediction errors”
(Froese & Ikegami, 2013, p. 213). Caregivers of children with ASD
often describe them as detached from the world, as if they are living
in their own walled world. This is not because they are unhappy or are
unable to move or sense, but because they are satisfied with the
current level of complexity of the environment. The prediction error
minimization principle says that “we harvest sensory signals that we
can predict” (Friston et al., 2012, para. 15). Hence, it seems that
children with ASD, because they (initially) cannot predict more com-
plex environments, are perfectly content to stick to the confined space
and motion they know.
Cognitive Functioning
In the following sections we review the most relevant literature
illustrating the implications of prediction errors with chronically
high precision in cognitive and perceptual domains. At the end, we
describe the special perceptual and cognitive skills that some
individuals with ASD have developed (savant skills), which can
result from the potential benefits of high-precision prediction
errors when applied to certain domains. The problems in reasoning
about mental states (mentalizing), which are also a central aspect
of cognition in ASD, are covered in the Social Functioning section.
Attention and Executive Functioning
An interesting pattern of findings has emerged from attention
studies in ASD, comprising both superior performance in certain
tasks and severe deficits in others. Below we substantiate that the
specific pattern is largely consistent with HIPPEA. We start by
considering visual search tasks, then move to more complex at-
tention tasks with a larger executive component, and finally make
new predictions based on our account and propose suitable designs
to test these.
Visual search studies demonstrate that performance on some
attentional tasks can be intact or even enhanced in ASD. Superior
visual search has been found both for a target defined by a single
feature and for conjunctive targets (Keehn et al., 2009;O’Riordan,
Plaisted, Driver, & Baron-Cohen, 2001;Plaisted, O’Riordan, &
Baron-Cohen, 1998). Moreover, performance seems to correlate
with symptom severity (Joseph, Keehn, Connolly, Wolfe, &
Horowitz, 2009). Group differences are especially present in con-
junction search tasks or tasks with higher difficulty. A predictive
coding account of visual search would start from the predictability
within search displays. When every item in a display reinforces a
particular prediction, an anomaly (the odd one out) causes an
“error” that becomes salient. Heightened precision of this predic-
tion error means enhanced salience of this oddball, which facili-
tates quick detection. Thus, individuals with ASD seem to be just
as good or even better at exploiting predictability in a display.
In more complex attentional settings, however, performance
usually declines substantially in autism. As we saw, precision (or
weight) of prediction errors should be flexibly adapted based on
meta-learning (learning which features in a task are relevant).
When precision of prediction errors is uniformly high, the selective
force is lost when processing a context with multiple cues. Hence,
difficulties in allocating attention may be expected. Phenomenally,
this may manifest itself as attention to irrelevant features, on the
one hand, and as lack of disengagement or perseverative attention,
on the other hand. Yet, this problem occurs only when multiple
cues compete. If only one cue is present, ensuring that the selection
process is clearly imposed by the task itself, performance remains
intact (Burack, 1994).
A study by Pierce, Glad, and Schreibman (1997) confirms this.
When children with ASD, TD children, and mentally disabled
children were presented with video fragments of social interactions
containing one or more cues, children with ASD performed worse
than the other two groups when asked to answer a set of questions
right after but only in the multiple cue conditions. We argue the
problem is one of autonomous selection; the relevance or redun-
dancy of the cues is not recognized. Consistent with this, task
performance in ASD is expected to suffer most when it is depen-
dent on autonomous exploration and efficiently probing of avail-
able cues rather than fixed instruction (clear top-down selection).
Others before us (e.g., Elsabbagh et al., 2009;Keehn, Müller, &
Townsend, 2013) have situated the origin of problems in ASD in
attentional difficulties, more specifically in disengaging attention.
However, we consider these disengagement problems not as pri-
mary but as an effect of the lack of adaptive precision of prediction
errors. This kind of overselective (perseverative) attention does not
stand in contradiction with what we said before on the lack of
autonomous selectivity in ASD. It is the flexible adapting of
selectivity in a task-dependent way that is lacking in ASD. Uni-
formly high precision will create a prolonged processing of all
stimuli (and an associated deficit in disengaging). This is also
apparent in studies by Sasson and colleagues (Sasson, Elison,
Turner-Brown, Dichter, & Bodfish, 2011;Sasson, Turner-Brown,
Holtzclaw, Lam, & Bodfish, 2008), demonstrating perseverative
attention (longer fixation times per image explored) and more
intensive, detail-oriented exploration of a limited number of im-
ages in ASD. Hohwy and Palmer (2014) noted that increased
precision could lead to longer sampling of incoming signal in order
to attain the precise signal people with ASD expect before making
a decision. If so, such longer sampling may as well help to explain
the larger reaction times for diverse tasks that are often reported in
ASD. In any case, we surmise that lacking disengagement is not
the core mechanism but rather one of the consequences of HIP-
PEA. However, perseveration and overselectivity may often be
strategically replaced by avoidance and underreactivity.
It is clear that the proposed difficulties in autonomous cue
selection will cause broader problems in executive functioning, in
particular with regard to cognitive flexibility or set shifting. Ac-
cording to the executive functioning theory, the latter functions are
impaired in ASD, and this impairment is assumed to underlie the
restricted, repetitive, and stereotyped pattern of behavior and in-
terests (B. R. Lopez, Lincoln, Ozonoff, & Lai, 2005). Though
problems with flexibility have clearly been found in daily life
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
654 VAN DE CRUYS ET AL.
(Gioia, Isquith, Kenworthy, & Barton, 2002;Mackinlay, Charman,
& Karmiloff-Smith, 2006), studies measuring cognitive flexibility
in a clinical or research setting have yielded less consistent evi-
dence. Overall, studies using the Wisconsin Card Scoring Task
report clear deficits, reflected by a higher number of perseverative
responses when a rule switch should occur, and more controlled
task-switching paradigms generally fail to find cognitive flexibility
problems in ASD (Geurts, Corbett, & Solomon, 2009;Van Eylen
et al., 2011).
Several researchers recently suggested that these inconsistencies
may be due to differences in the extent to which explicit task
instructions are given, denoted as open endedness (Van Eylen et
al., 2011;White, 2013). When task instructions contain no explicit
indication of the rules to be applied nor an indication that a rule
switch will occur (as in the Wisconsin Card Scoring Task), results
show rather consistent cognitive flexibility deficits in ASD. In this
case, participants have to be able to autonomously filter out and
focus on relevant information in a situation where multiple cues
compete. There is evidence that individuals with ASD have diffi-
culties doing so and overly focus on irrelevant, often low-level
details (Stoet & López, 2011). In contrast, when a cue explicitly
indicates which rule to apply and when to switch, all studies report
intact performance in ASD. Hence, the act of switching does not
seem to be a problem per se (Poljac & Bekkering, 2012).
All this is very compatible with our interpretation of ASD in
terms of an overweighing of prediction errors. As we saw, the
informativeness of cues has to be derived from meta-models,
which should adjust the precision with which errors based on these
cues are weighed. A loss of this capacity would lead to a deficit in
the ability to autonomously select cues that have predictive value
in situations where multiple cues compete. Learning a new unam-
biguous contingency in itself is not a problem, but individuals with
ASD struggle with spontaneously noticing that the predictive value
of particular information has changed. This leads to cognitive
flexibility deficits on open-ended tasks but not on tasks were
explicit instructions are provided about what is informative and
when. Testing a range of executive functions in ASD, White,
Burgess, and Hill (2009) corroborated that all open-ended tasks
generated group differences but none of the more constrained tasks
did. Hence, this reasoning might also explain some of the incon-
sistencies in studies of other executive functions (Gioia et al.,
2002;White et al., 2009).
Open-ended, generative sorting tasks provide converging evi-
dence. For example, in a free sorting task with childrens books
(Ropar & Peebles, 2007), children with ASD relied less on cate-
gory labels (games vs. sports) and more on purely perceptual
features (color and size) than did TD children. More one-
dimensional sorting was found in free sorting of shapes by children
with ASD, especially in more complex stimulus sets (D. J. Ed-
wards, Perlman, & Reed, 2012). In a 20-questions game, children
with ASD consistently generated questions (predictions in our
context) of lower quality, especially more concrete ones that
eliminated fewer items at a time (Alderson-Day & McGonigle-
Chalmers, 2011). Analyses indicated that difficulties in managing
relevant and irrelevant information were likely sources of the
problems of children with ASD. This cognitive control problem,
which is at the heart of HIPPEA, also explains why individuals
with ASD are particularly slower in early blocks of categorization
learning, when flexible switching of the focus of attention from
one dimension to another dimension is needed (e.g., Soulières,
Mottron, Giguère, & Larochelle, 2011).
To clearly summarize our hypothesis: When real environmental
changes go together with random changes, disentangling the two is
particularly difficult for people with ASD. They seem to be able to
learn changes in contingencies, when these are clearly indicated, as
in some set shifting tasks. Similarly, they can learn fixed contin-
gencies, even in probabilistic environments and without explicit
instructions, as implicit learning studies show (J. Brown, Aczel,
Jiménez, Kaufman, & Grant, 2010;Nemeth et al., 2010;Pruett et
al., 2011). However, these two combined create the clearest defi-
cits. Therefore, we hypothesize that adding noise by using a
probabilistic switching task would increase their flexibility impair-
ments. This has indeed been observed by D’Cruz et al. (2013) in
a reversal learning task with intermittent nonreinforcement. More-
over, these switching problems correlated with severity of repeti-
tive and restrictive behaviors. From our perspective, this kind of
task will be most sensitive in picking up deficits in executive
functioning for ASD.
Although these findings are largely compatible with the predic-
tion derived from HIPPEA, future attentional studies should test
our hypothesis more directly. A modified version of Posner’s
attention cuing task as developed previously (Vossel et al., 2014;
Yu & Dayan, 2005) could contribute to this. In the typical Posner
cuing task, a simple cue (a briefly presented flash) indicates the
actual (valid) location of the target to children in only a certain
percentage of trials (e.g., 75%). Participants typically will learn to
use the cue information to improve their detection speed to the
extent that the cue is reliable. This improvement may also be
present for individuals with ASD, but we predict that things will go
awry in ASD when the probabilistic structure changes unexpect-
edly during the experiment; for instance, when the predictability of
a cue changes across blocks. In such a volatile environment, the
validity of the cue (the extent to which it predicts the target
location) varies over the course of the experiment. Prediction
errors usually lead to the updating of beliefs (predictions) about the
environment, but the impact of these prediction errors should be
tuned to whether additional learning is expected to be still possible.
In a fully learned stable phase, new prediction errors are probabi-
listic noise that should lead to little or no update of predictions.
However, when new learning is estimated to be possible (e.g.,
when probability structure has changed), recent prediction errors
should significantly update current predictions. This task shows
the importance of contextual, flexible setting of precision.
Another variation of the Posner task that could provide a useful
test of our theory has been developed by Yu and Dayan (2005).In
this version, a set of cues (e.g., differently colored arrows pointing
left or right) precedes the target. For any one trial, one particular
cue (color) from the set predicts the target location with a certain
probability (e.g., .5). This cue type and validity remain active for
a considerable amount of time, creating a stable environment.
Then, unbeknownst to the participant, this context is suddenly
changed: A different cue now predicts the target location with a
different cue validity. Note the similarity with traditional set
switching tasks, although the rules there usually are deterministic
rather than probabilistic. Participants with ASD will have distinct
problems with this task, again because two forms of uncertainty
are pitted against each other, as described above. An added benefit
of these tasks is that a hierarchical Bayesian model can be used to
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
655
PREDICTIVE CODING IN AUTISM
quantify precision (or learning rate) on a trial-by-trial basis (Beh-
rens, Woolrich, Walton, & Rushworth, 2007;Yu & Dayan, 2005)
and on a subject-dependent basis (Mathys, Daunizeau, Friston, &
Stephan, 2011;Vossel et al., 2014), pin-pointing exactly whether
and how the learning style of individuals with ASD differs from
that of TD individuals.
From the above it should be clear that HIPPEA has a natural
way of explaining the discrepancy between the experimental data
in contrived laboratory contexts and the clinical observations in
daily life. As most ASD researchers know, it is surprisingly
difficult to find statistically significant group differences in the lab
that should occur, according to everyone’s expectations, based on
the major problems that people with ASD experience every day
(J. L. Amaral, Collins, Bohache, & Kloos, 2012). Natural circum-
stances are often much more unpredictable and open ended, with
lots of accidental variability, and hence lead to clear deficits in
those with ASD (Kenworthy, Yerys, Anthony, & Wallace, 2008).
The lab, in contrast, usually provides a well-controlled environ-
ment, in which it is made very clear what is expected (explicit
instruction, practice trials) and with multiple instances of the same
(often simple) task (repeated trials). Little autonomous control is
needed here. Whereas many TD children easily get bored in such
a context and start talking to the experimenter, kids with ASD
usually like these repetitive, computerized tasks and are motivated
to do well in them.
Perceptual Processing
Research on visual processing in ASD has been dominated by
two related theoretical frameworks that each emphasized a differ-
ent side of the coin: WCC theory emphasized reduced global
processing (Frith & Happé, 1994), and EPF theory emphasized
enhanced local processing (Mottron & Burack, 2001). More recent
accounts describe the peculiar aspects of visual processing in ASD
more in terms of a bias or perceptual style, a disinclination for
global or a preference for local processing (Happé & Booth, 2008;
Happé & Frith, 2006;Mottron et al., 2006). Despite a vast amount
of research on visual perception in ASD, the atypical profile of
visual processing is only partly understood, and the empirical
evidence for the original ideas is mixed (for recent reviews, see
Behrmann, Thomas, & Humphreys, 2006;Dakin & Frith, 2005;
Simmons et al., 2009).
HIPPEA is compatible with both EPF and WCC, but it offers a
more specific foundation, describes dynamics in learning and
inference, and hence has different implications. According to
HIPPEA, precision of bottom-up information is uniformly ampli-
fied. This idea is consistent with EPF, but we can better specify
how and why perception is enhanced. The detectable size of
prediction errors is not smaller; rather, the weights (precision)
these errors receive are higher. HIPPEA does not reduce problems
to a purely bottom-up way of perceptual processing. Because it is
embedded in the inherently bidirectional predictive coding frame-
work, the mutual, constructive interaction of bottom-up and top-
down information flows is central. Specifically, increased preci-
sion of prediction errors will have important consequences with
regard to the kind of predictions that will be formed based on
prediction errors with unusually high precision. Perceptual infer-
ence and learning will not progress to higher level, more abstract
representations because of the emphasis given to violations to
those higher level representations at lower levels of processing.
Learning will result in predictions tuned sharply to exact percep-
tual input cues. As a result, primarily low-level predictions, which
will have limited applicability, will be formed. Higher level pre-
dictions will be triggered less automatically by incoming informa-
tion, an idea that is consistent with WCC.
In ASD, stimuli are treated in an idiosyncratic manner, because
slight deviations are perceived as informative and all experiences
are thus more readily treated as new instead of as belonging to a
known category. More concretely, the focus on prediction errors at
lower levels causes individuals with ASD to focus on concrete but
irrelevant changes in viewpoint or illumination, which impede the
ability to progress to the more relevant, abstract levels of descrip-
tion in terms of shape or object identity. Note, however, that the
predictive machinery in ASD is not deficient in our view: Predic-
tions are still formed and prediction error is computed correctly.
Hence, global interpretations are not necessarily lost in ASD; they
just require more experience, and they will appear only under more
constrained conditions. So, although a familiar representation may
not pop up automatically when a related stimulus appears, top-
down activation of holistic, Gestalt-like templates and global pro-
cessing are often still possible but as a conscious strategy, when
task instructions require it and enough time is available. For
individuals with ASD, it is not the default, automatic processing
mode. This accords nicely with the recent move in the field toward
differences in default preference or bias (often measured by initial
choice responses or reaction times) rather than in distinct inabili-
ties (measured by error rates). This interpretation receives support
from a recent meta-analysis of the mixed evidence from a variety
of local–global perceptual processing tasks, which demonstrates
that global processing takes time in individuals with ASD (Van der
Hallen, Evers, Brewaeys, Van den Noortgate, & Wagemans,
2014). Moreover, the inconsistencies in the literature also make
sense in this perspective. Laboratory tasks mostly use standardized
stimuli and often do not incorporate the noise that is usually
present in real-life stimuli. In these constrained circumstances,
individuals with ASD can actually perform on a typical level.
Low-level perception. According to HIPPEA, low-level dif-
ferences will get boosted and sent upward, influencing behavior
and learning. Setting precision high by default may give an ad-
vantage for lower level processing (but impedes building and using
of a hierarchy of predictions for generalization). In the auditory
domain this is reflected in the frequently reported enhanced pitch
perception in children and in a subgroup of adolescents and adults
with ASD, especially those with early developmental language
delay and language-related difficulties (for reviews, see Haesen,
Boets, & Wagemans, 2011;O’Connor, 2012). Superior pitch pro-
cessing has been established regardless of stimulus complexity
(i.e., pure tones, complex tones, speech sounds, nonwords, words)
using a variety of psychophysical tasks (e.g., identification, dis-
crimination, memory; e.g., Bonnel et al., 2010,2003;Jones et al.,
2009). Also relevant in this context is the increased prevalence of
absolute pitch and musical savants in the ASD population (e.g.,
Heaton, Williams, Cummins, & Happé, 2008).
In visual perception, findings are more mixed. Most studies
have found little or no group differences for visual acuity (Sim-
mons et al., 2009). One study observed a small group difference
indicating superior contrast sensitivity in individuals with ASD
(Bertone, Mottron, Jelenic, & Faubert, 2005). Another study found
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
656 VAN DE CRUYS ET AL.
evidence for superior visual acuity (Ashwin, Ashwin, Rhydderch,
Howells, & Baron-Cohen, 2009), but this finding has been dis-
puted on methodological grounds (M. Bach & Dakin, 2009), and
replication attempts have failed (Bölte et al., 2012;Kéïta, Mottron,
& Bertone, 2010). Based on HIPPEA, however, there is still
potential for well-controlled studies to find detection differences.
In particular, it may be relevant to look at classic effects of
perceptual gain control (Hillyard, Vogel, & Luck, 1998) in ASD,
because precision is thought to rely on gain control of the output
of neural units representing the perceptual prediction errors (Fris-
ton, 2009). Foss-Feig, Tadin, Schauder, and Cascio (2013) very
recently found that detection of motion direction of a single clearly
visible grating can be done based on significantly shorter presen-
tation times in ASD compared to controls. The improvement was
not present for low-contrast gratings, for which gain control is
negligible. Hence, it seems it is caused by reduced contrast satu-
ration of high-contrast gratings in ASD, consistent with a deficient
perceptual gain control.
Local versus global perception. A common paradigm to
study a more locally focused processing style in ASD is to exam-
ine susceptibility to visual illusions. Overall, these studies yielded
mixed to positive effects. Although some authors did not find a
difference in performance for ASD (C. Brown, Gruber, Boucher,
Rippon, & Brock, 2005;Rouse, Donnelly, Hadwin, & Brown,
2004), most others showed a diminished illusion susceptibility in
ASD (e.g., Bölte, Holtmann, Poustka, Scheurich, & Schmidt,
2007;Mitchell, Mottron, Soulieres, & Ropar, 2010). This dimin-
ished susceptibility has been taken to imply that individuals with
ASD are, in general, less influenced by contextual or prior infor-
mation, remaining closer to the actual sensory input, an idea that is
perfectly consistent with HIPPEA. For instance, when Ropar and
Mitchell (2002) asked participants to estimate the shape of an
illuminated disc presented at a slanted angle in a darkened room,
control participants reported a more circular shape (closer to the
inferred distal stimulus, discounting the slant), and participants
with ASD reported a more elliptic shape (closer to the proximal
stimulus, not discounting the slant).
The global–local processing issue is usually studied with the
block design task and the embedded figures task. The first study
showed enhanced performance in both of these tasks in individuals
with ASD (Shah & Frith, 1993), which was interpreted as evidence
for reduced interference by the automatic processing of the global
level. Later studies, however, yielded mixed results (e.g., Bölte,
Hubl, Dierks, Holtmann, & Poustka, 2008;Ropar & Mitchell,
2001). Collectively, these results point to a difference in degree of
efficiency or ease with which the task is performed, rather than a
discrete performance difference (Van der Hallen et al., 2014).
Another domain in which the visual abilities in ASD have
received a lot of attention is the perception of motion. A study by
Bertone, Mottron, Jelenic, and Faubert (2003) revealed intact
first-order (luminance-defined) motion processing but impaired
second-order (texture-defined) motion processing. Motion coher-
ence studies, in which observers have to track the presence or
direction of coherently moving (luminance-defined) dots among
differing proportions of randomly moving dots, generally yielded
higher motion coherence thresholds in individuals with ASD (e.g.,
Milne et al., 2002;Pellicano, Gibson, Maybery, Durkin, & Bad-
cock, 2005;Spencer et al., 2000), although there are also excep-
tions (De Jonge et al., 2007;Del Viva, Igliozzi, Tancredi, &
Brizzolara, 2006;Saygin, Cook, & Blakemore, 2010). A recent
study may explain this inconsistency (Robertson, Martin, Baker, &
Baron-Cohen, 2012) by reporting a deficit in perception of motion
coherence at short exposure durations, which lessens with increas-
ing exposure durations.
The finding that added noise is especially detrimental for global
motion perception in ASD is one that follows directly from
HIPPEA. Distinguishing noise and signal is particularly important
in these paradigms. As explained before, people with ASD attri-
bute unduly high value to noise that is unlikely to repeat, in an
attempt to properly fit the input. Global motion will more readily
break down for them, because they end up with errors that are too
important to fit with an abstracted, global pattern. When the noise is
absent, as in the plaid motion stimuli in Vandenbroucke, Scholte, van
Engeland, Lamme, and Kemner (2008), global motion perception
seems to be intact in ASD.
Research with bistable figures suggests that people with ASD
can generate and maintain top-down predictions, because when
guided to do so, they easily succeed in making the different
interpretations of ambiguous figures (Ropar, Mitchell, & Ackroyd,
2003). However, we would advise the use of binocular rivalry in
future studies (rather than the pen-and-paper type face–vase or
duck–rabbit tests used so far), because it has been proposed to be
explained by predictive coding (Hohwy, Roepstorff, & Friston,
2008). Indeed, input related to the suppressed image in binocular
rivalry can be considered prediction error, because it is unex-
plained by the currently dominant percept. Only two studies have
been performed so far, with one showing unaltered binocular
rivalry in ASD (Said, Egan, Minshew, Behrmann, & Heeger,
2013) and the other finding lower switch rates and more mixed
percepts (Robertson, Kravitz, Freyberg, Baron-Cohen, & Baker,
2013). Mixed percepts could be the preferred way to minimize
prediction errors in ASD (i.e., less explaining away through higher
level constructs and hence staying closer to the input). Note that
care should be put into finding the right stimuli for use in autism,
because availability of top-down templates evidently also influ-
ences rivalry. For example, the first study uses gratings and the
second uses familiar objects. The less familiar (or semantically
high-level) the stimulus, the better it is for use in studies of ASD,
at least when the focus is really on switching dynamics. Future
binocular rivalry studies in ASD should specifically look at mixed
percepts and fusion, because this is the expected result according
to predictive coding if precise prediction errors are present for both
hypotheses (Hohwy et al., 2008). Another, yet to be tested predic-
tion from HIPPEA would be that adding noise (prediction error) to
the input has a stronger effect on the breaking of one percept (and
possibly inducing a switch) in those with ASD than in controls.
Face and speech perception. Face and speech perception are
crucial for smooth and successful social interactions and are there-
fore prominent targets of ASD research. Deep difficulties here can
go a long way in explaining communication problems so central in
ASD. Of interest, face and speech perception are also prime
examples of the hierarchical “analysis by synthesis” approach
inherent to predictive coding. This would normally provide infer-
ences on high-level semantic sources of incoming sensory infor-
mation (a generative model) that can cascade into multiple levels
of predictions for activity in regions below, suppressing or ex-
plaining away new input, as long as it is sufficiently well pre-
dicted. Yet, what is sufficient has to be learned as well (meta-
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
657
PREDICTIVE CODING IN AUTISM
learning), given the requirements of speech or face understanding.
If precision of prediction errors is invariably high, individuals with
ASD will have more difficulties in “abstracting away” the short-
term, contingent, low-level features of the stimuli. Behaviorally,
this is expressed in individuals with ASD as a superior access to
the underlying low-level visual or acoustic representations. On the
other hand, they cannot fully exploit the higher level predictions.
This disadvantage is particularly felt in naturalistic face or speech-
in-noise perception. The problem of which variations to encode in
a given situation and which to disregard comes to the forefront in
both speech and faces, which is exactly one of the problems for
individuals with ASD, according to HIPPEA. Not only does the
brain have to pick up and learn small auditory or visual differ-
ences, it also has to learn which ones are informative, in the sense
of predictive for different kinds of social goals, and which differ-
ences to discard.
In speech, invariant phonetic cues are embedded within a variety
of acoustic cues (e.g., fundamental frequency, accent, intonation,
timbre) and can only be extracted by integrating and interpolating
information, a process that is supported by higher level linguistic
guidance through phonotactic, semantic, and syntactic constraints
(predictions). During development very young TD children learn
to generalize consonants, vowels, and words across voices (e.g., of
different gender), disregarding irrelevant absolute pitch cues in
favor of more complex relative distances. However, in those with
ASD we see increased access to fine-grained acoustic features of
complex sounds (e.g., disembedding tones from musical chords;
Heaton, 2003;Mottron, Peretz, & Ménard, 2000) and superior
perceptual processing of acoustic features of speech (e.g., Heaton
et al., 2008;Jarvinen-Pasley, Peppe, King-Smith, & Heaton,
2008). Consistent with HIPPEA, it has been suggested that these
individuals generate overly specific categories of sounds that im-
pede learning of higher level abstract patterns (Crespi, 2013)
needed for speech development. Early developmental language
delays as well as broader linguistic impairments later in life are
indeed prevalent in the individuals who show superior acoustic
processing of pitch (Bonnel et al., 2010;Jones et al., 2009).
Additionally, noise with similar characteristics as the signal
(speech), substantially hinders performance in those with ASD
(E. G. Smith & Bennetto, 2007), because these “errors” are not
easily ignored.
A similar challenge is posed by faces, characterized by a very
high intraclass similarity, with small and rather subtle differences
among many dimensions distinguishing two human faces from
each other. Countless transformations of an individual face among
several dimensions should be ignored. A face has to be recognized
despite variability in, for instance, lighting conditions, face orien-
tation, changeable facial features (e.g., facial hair, spots, wrinkles,
freckles), and extrafacial features (e.g., hair style, hats). Due to
their meta-learning problems, individuals with ASD may fail to
make this distinction between relevant and irrelevant variability
and hence get lost in nonfunctional characteristics. This may
explain their poorer face memory and their face identity recogni-
tion problems (Weigelt, Koldewyn, & Kanwisher, 2012).
Paralleling evidence on global–local processing in general, there
is no strong evidence for a reduced global or enhanced local face
processing style in ASD. For example, no reduced face inversion
effect, no attenuated composite face illusion, no diminished part–
whole effect, and no decreased susceptibility to the Thatcher
illusion were found in ASD (for a review, see Weigelt et al., 2012).
More implicit measures, which are less prone to compensatory
strategies, do find differences in face processing, contrary to most
behavioral studies. For example, children with ASD fail to show
the typical longer looking times (van der Geest, Kemner, Verbaten,
& van Engeland, 2002) and the typical larger pupil dilation (Falck-
Ytter, 2008) for upright than for inverted faces. Moreover, ERP
studies demonstrated that the typical differential response to up-
right versus inverted faces is not present in adults with ASD
(McPartland, Dawson, Webb, Panagiotides, & Carver, 2004;
Webb et al., 2012). These findings all point to less efficient face
processing, because selection and emphasis of predictive cues are
missing, throwing face perception back to processes similar to
those used for inverted faces. It also fits with HIPPEA that when
global face processing deficits are found, they will disappear if
participants with ASD are explicitly cued (e.g., “look at the eyes”),
as shown by B. Lopez, Donnelly, Hadwin, and Leekam (2004).
Studies finding disturbed formation of face prototypes in ASD
may also confirm our account (Gastgeb, Rump, Best, Minshew, &
Strauss, 2009;Gastgeb, Wilkinson, Minshew, & Strauss, 2011).
Forming a face prototype typically requires the use of the central
tendency in all encountered exemplar faces to arrive at an implicit,
average representation, ignoring the within-category variability
(Valentine, 1991). In ASD, however, the emergence of a familiar,
broad face prototype will not occur automatically. For categoriza-
tion to work, new instances have to be recognized as similar to
previously experienced examples. The chronically high precision
of prediction errors will impede this ability by overemphasizing
the extent to which new input deviates from previous examples or
learned templates. Plaisted et al. (1998) consistently found that
high-functioning adults with ASD learned to discriminate between
configurations of colored disks to higher levels of accuracy than
controls; when tested with slightly different exemplars of the same
overall configurations, normal controls showed transfer from the
learned exemplars to the novel ones, and individuals with ASD did
not. As a result, individuals with ASD may be slower at catego-
rization learning (e.g., Klinger & Dawson, 2001;Soulières et al.,
2011), and they may be less spontaneous in extracting a prototype
from a series of exemplars (e.g., Gastgeb, Dundas, Minshew, &
Strauss, 2012;Vladusich, Olu-Lafe, Kim, Tager-Flusberg, &
Grossberg, 2010).
Finally, impaired formation of a familiar, broad face prototype
can also be seen in the reduced face adaption aftereffects (e.g.,
Pellicano, Jeffery, Burr, & Rhodes, 2007;Rutherford, Troubridge,
& Walsh, 2012). Though these findings may mean that perception
is less influenced by prior knowledge (in this case the shifted
prototype; Pellicano & Burr, 2012), we would propose that it is the
consequence of an abnormal updating of representations (proto-
types). An adapting exemplar may not update the main prototype,
because it contains important enough differences for individuals
with ASD to deserve creation of a novel, narrow prototype. Future
studies of lower level feature adaptation, currently lacking in ASD,
may be able to resolve this debate.
Mismatch negativity. Although the predictive coding ac-
count has originally been conceptualized in the visual domain, a
growing number of studies has also investigated predictive coding
phenomena in the auditory modality (Arnal & Giraud, 2012;
Winkler, Denham, & Nelken, 2009). In this regard, auditory mis-
match negativity (MMN) research has been particularly informa-
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
658 VAN DE CRUYS ET AL.
tive. Here, presentation of an unexpected oddball stimulus within
a sequence of repeated predictable stimuli, elicits a novelty re-
sponse in the event-related potential. The MMN was originally
interpreted as reflecting change detection on the basis of a passive,
bottom-up process of adaptation to the repeated stimuli (May &
Tiitinen, 2010). Recent evidence, however, has shown that the
MMN does not reflect release of repetition suppression but is the
result of a violated prediction rather than a physical stimulus
change (e.g., Todorovic, van Ede, Maris, & de Lange, 2011;
Wacongne, Changeux, & Dehaene, 2012). A series of studies
further made plausible that the perceptual cortex indeed imple-
ments a hierarchy of predictions and prediction errors, with repe-
tition suppression attenuating neural responses in a very early time
window (4060 ms), stimulus expectation on the basis of uncon-
scious local predictions attenuating the intermediate stage of pro-
cessing (100–200 ms; i.e., the typical MMN which originates in
sensory areas), and stimulus expectations on the basis of more
global, integrative and conscious predictions modulating the later
P3b novelty response (300600 ms, originating from a broader
frontoparietal predictive network; Todorovic & de Lange, 2012;
Wacongne et al., 2011). With regard to the MMN, a number of
studies observed larger amplitudes and/or earlier latencies to in-
frequent pitch changes in tones and vowels in those with ASD
relative to TD controls, thus suggesting hypersensitivity and su-
perior recognition of pitch change (e.g., Ferri et al., 2003;Gomot,
Giard, Adrien, Barthelemy, & Bruneau, 2002;Lepisto et al., 2005;
but see Dunn, Gomes, & Gravel, 2008). Of interest, Gomot et al.
(2011) showed that these electrophysiological abnormalities were
significantly more pronounced in children who displayed greater
difficulties in tolerating change. The MMN response to infrequent
phonemic changes in vowels or consonants, however, is typically
smaller and/or delayed in ASD and thus suggestive of impaired
recognition of the more global phonetic characteristics of speech
(e.g., Kujala, Lepisto, Nieminen-von Wendt, Naatanen, & Naa-
tanen, 2005;Lepisto et al., 2006). Finally, the later P3b compo-
nent, presumably characterizing more global and integrative vio-
lations of expectations, exhibits smaller amplitudes in those with
ASD relative to controls (e.g., G. Dawson, Finley, Phillips, Galp-
ert, & Lewy, 1988;Kemner, Verbaten, Cuperus, Camfferman, &
Van Engeland, 1995). Comparing neurophysiological findings per-
taining to MMN versus P3b processing suggests that the brains of
individuals with ASD are tuned to register low-level local changes
in transition probabilities (enhanced and earlier MMN sensory
responses toward simple stimuli) but have difficulty picking up
changes in the broader frontoparietal predictive system, which is
tuned toward more global, higher level patterns. This is at least
compatible with the view that increased low-level precision hin-
ders the formation of appropriate predictions higher up.
Savant Skills
Autistic savants are individuals with ASD with co-occurring
excellence in an isolated skill (i.e., an “island of genius” that
contrasts with the individual’s general lower-than-average abili-
ties). Savantism has been identified in a wide range of neurological
and neurodevelopmental disorders but is most frequently reported
in ASD. Savant skills are estimated to be present in one out of 10
autistic individuals, with males outnumbering females (approx.
6:1; Howlin, Goode, Hutton, & Rutter, 2009;Treffert, 2009).
Savant skills usually fall within one of five general categories (i.e.,
musical abilities, calendar calculating, mathematics, art and me-
chanical or spatial skills; Treffert, 2009). Although the savant skill
of an individual may evolve over the years, the skill should not
fade or disappear over time but should remain a peak in perfor-
mance.
Several scholars have attempted to explain the mechanism be-
hind the savant skills. Plaisted (2001) suggested a reduced ability
to process similarity at the perceptual and attentional level, which
results in a reduced tendency to generalize information. Baron-
Cohen (2006) postulated an increased drive to construct or ana-
lyze, which he referred to as “hyper-systemizing.” The alleged
adaptive function of the systemizing mechanism is to serve as a
law detector and a change-predicting mechanism. He argued that
people with ASD prefer either no change or systems that change in
highly lawful or predictable ways (i.e., systems with rule-bound
change, such as mathematics, physics, objects that spin or recur,
music, machines, collections), and this is why they become dis-
abled or change-resistant when faced with systems characterized
by complex change (such as social interaction). Mottron et al.
(2006) and Mottron et al. (2013) emphasized the putative role of
enhanced feed-forward low-level perception and suggested that
individuals with ASD have a developmental predisposition to
veridical mapping of data and information. Although these ac-
counts provide insight into the origin of such a skill, HIPPEA
makes more specific claims about the underlying mechanisms.
Our predictive coding approach explains why similarity is not
processed in the same way in those with ASD, consistent with
Plaisted (2001). It also elucidates why complex change is chal-
lenging (Baron-Cohen, 2006): This is where meta-learning should
lead to distinguishing mere noise from actual environmental
changes. Finally, the veridical mapping can also be seen as a
consequence of the constant drive to reduce even irrelevant pre-
diction errors (Mottron et al., 2013). Although predictions shaped
by noise and irrelevant details will often result in impaired or slow
processing, doing this for a specific, limited topic of interest can be
quite possible and, above all, rewarding. Developing such a skill
becomes extrinsically motivating (e.g., getting praise and atten-
tion) but also intrinsically, as making successful predictions in this
particular domain will result in feelings of reward and the notion
that the generally unpredictable world is more controllable. For
example, phone numbers have an exact but arbitrary mapping
(Mottron et al., 2013). All known examples of savant skills—for
instance, 3D drawings or musical play from memory—combine
two factors: an exquisite discriminative sensory ability and an
exceptional (rote) memory capacity (A. L. Hill, 1978;Treffert,
2009). The first is a general feature of ASD, we would argue,
originating from high-precision low-level prediction errors. A lack
of abstraction is actually an advantage here. Clearly, this discrim-
inative ability can only fully be put to use in the case of high
memory capacity. This may be the feature that is specific to
savants, but even then resource constraints may seriously limit the
savant domain.
Sensorimotor Abilities and a Sense of Self
Within the predictive coding theory, actions also entail predic-
tions; namely, of their proprioceptive and exteroceptive conse-
quences. According to M. J. Edwards, Adams, Brown, Pareés, and
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
659
PREDICTIVE CODING IN AUTISM
Friston (2012), movement is defined by “what we want to see (or
feel), rather than what we want to do” (p. 3498). In this view,
actions can be regarded as being aimed at fulfilling predictions
(reducing prediction errors) of perceptual input. Several ASD
symptoms can be readily interpreted from this perspective. Given
that actions generate prediction errors, those actions that reduce
these prediction errors to extreme minima should be preferred.
Accordingly, some of the most characterizing symptoms in ASD
are the stereotypical, repetitive (predictable) behavior patterns
(Turner, 1999). These patterns establish controllable and thus very
predictable proprioceptive (kinesthetic) feedback that helps indi-
viduals with ASD to better cope with their environment (Ornitz,
1974). In a similar vein, the repetitive handling of lighting and
spinning objects and the repetitive tactile self-stimulation can be
regarded as manners of creating a predictive environment to re-
duce and cope with prediction error. Unpredictable surroundings in
particular may be expected to elicit this kind of behavior, with the
aim of reestablishing predictability and reducing stress (see the
Chronic Unpredictability and Its Affective Consequences section).
Ornitz (1974) observed that “in their spontaneous activity autistic
children are continually spirting, twirling, flicking, tapping, or
rubbing objects. Furthermore, they repetitively flap, writhe, wig-
gle, or oscillate their extremities while regarding them intently” (p.
204). This latter part is significant because it indicates that al-
though TD children might progress to more complex kinds of
“play” (learning), children with ASD continue to be engaged in
and learn from these simpler patterns.
According to HIPPEA, atypical behavior has the aim of regu-
lating excessive amounts of prediction errors. At first sight, this
seems very similar to the explanation invoked by the EPF theory;
namely, reducing excessive perceptual input (Mottron et al., 2006).
However, in our view, individuals with ASD aim only to reduce
that part of the perceptual input that cannot be predicted and
moreover actively attempt to create predictability to compensate.
Reports of autistic children screaming all day, despite being hy-
persensitive to noise themselves, might be understood as a way of
dealing with prediction errors by making the sensory environment
more predictable. The active desire for predictable sensory expe-
rience is brought even more clearly into light by Temple Grandin,
an autistic woman who built a mechanic body squeeze machine
because she liked the feeling of being touched and hugged but
wanted it to be a perfectly controlled (i.e., predictable) act instead
of the unpredictable overstimulating human touch (Edelson, Edel-
son, Kerr, & Grandin, 1999;Grandin, 1992). In a similar vein, the
“high systemizing” concept used by Baron-Cohen, Ashwin, Ash-
win, Tavassoli, and Chakrabarti (2009) to characterize the cogni-
tive style of individuals with ASD can underscore that predictable
patterns are formed and are important in their minds. The obses-
sion with regularity can be seen as borne of an overweighing of
deviations.
The sense of self and of agency has also been related to (intero-
ceptive) predictive coding (Apps & Tsakiris, 2013;Seth, Suzuki,
& Critchley, 2012). It is through the tightly cross-modally corre-
lated proprioceptive, tactile, and visual input of self-induced
movements that we construct the sense of a self that acts in the
world. The high-level concept of the self is the most plausible
prediction explaining low-level regularities in cross-modal input.
This view of the emergence of the self via the observed correla-
tions between proprioceptive, tactile, and visual modalities can
also explain why artificially created correlations can create the
illusion that extracorporeal objects are part of our own body (e.g.,
rubber hand illusion; Apps & Tsakiris, 2013;Botvinick & Cohen,
1998). Awareness of self and body as distinct from the world is
thus dependent upon a certain degree of tolerance derived from the
active, successful suppression of interoceptive prediction errors
(Seth et al., 2012). The presence of repetitive, stereotyped move-
ments in ASD during early development suggests that an abnor-
mally large amount of correlated input is needed to establish a
sense of self as separated from the surroundings (see also Brincker
& Torres, 2013).
Two recent studies using the rubber hand illusion, an illusion of
perceived arm position induced by correlated (synchronized) stim-
ulation (Palmer et al., 2013;Paton, Hohwy, & Enticott, 2012),
support this view. Both in individuals with ASD and in those with
high but nonclinical ASD traits, the consequences of experiencing
the illusion (on drift and movement) were reduced. A higher
estimated precision of prediction errors may indeed lead to a
reduced illusory percept, requiring more tightly correlated input
(than is usually provided in this rubber hand procedures) for the
illusory percept to fully establish itself. More generally, motor
coordination problems, often noted in ASD (Fournier, Hass, Naik,
Lodha, & Cauraugh, 2010;M. L. Matson, Matson, & Beighley,
2011), may be another consequence of overprecision of movement
prediction errors in contexts that actually have a considerable
amount of uncertainty (Palmer et al., 2013).
The observation that the repetitive, self-focused behaviors often
decrease during development (Richler, Huerta, Bishop, & Lord,
2010) suggests that extensive exposure may eventually lead to a
more stable sense of self. However, the typical “insistence on
sameness” (Kanner, 1943) remains or increases with age, indicat-
ing that exteroceptive prediction errors generally remain precise.
This insistence on routine or rituals and resistance to trivial
changes in the surroundings again demonstrate that children with
ASD do develop clear predictions on what should happen next in
the current situation, in contrast to theories positing a uniformly
weaker application of predictions in those with ASD (Pellicano &
Burr, 2012). Therefore, insistence on sameness may be considered
a hallmark of HIPPEA: It signals a clear grasp (prediction) on how
the world should behave, while assigning too much importance to
incidental changes.
Chronic Unpredictability and Its
Affective Consequences
One of the most prominent clinical observations in individuals
with ASD is their unusual reactivity to sensory stimuli. Numerous
clinical and personal reports describe the presence of both hyper-
and hyposensitivity to sensory stimulation. Hypersensitivity has
been described in various modalities (Blakemore et al., 2006;Kern
et al., 2006;Khalfa et al., 2004). Enhanced sensitivity to loud and
unexpected sounds is particularly evident in children with ASD
(e.g., Grandin, 1995;Tomchek & Dunn, 2007) and appears to
decrease with age, with adults with ASD becoming more similar to
TD adults (Kern et al., 2006). Yet, feelings of stimulus overload
and hypersensitivity to noise are also common in adults on the
autistic spectrum (in particular in social situations, like receptions
or parties) and can cause great distress and anxiety. Enhanced
sensitivity to visual stimuli is less common in ASD but does occur
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
660 VAN DE CRUYS ET AL.
(e.g., under the form of enhanced discomfort to bright light; Kern
et al., 2001). When the gain of the neural units representing the
prediction errors is fixed at a high level, it is easy to see that
hypersensitivity becomes very likely, especially for unexpected
input, as is the case in ASD. Overweighting of irrelevant predic-
tion errors causes sensory overload.
Seeing that unpredictability is at the core of the sensory over-
load, we can also attempt to explain its negative affective impact.
Uncertainty has long been identified as a factor that intensifies
stress and anxiety (Herry et al., 2007;Miller, 1981). In addition to
leading to increased stress and anxiety, persistent significant pre-
diction errors may actually by themselves generate negative affect
(Huron, 2006;Van de Cruys & Wagemans, 2011). When predic-
tions are invoked, there is actually something at stake; namely, the
success of current internal models of the environment. When
prediction errors signal the need for extra resources, aimed at
updating the internal model, they may have negative affective
value. For example, supposedly neutral perceptual prediction er-
rors activate the habenula, a region known to code prediction
errors of negative valence (Schiffer, Ahlheim, Wurm, & Schubotz,
2012;Schiffer & Schubotz, 2011). Originating from the cognitive
dissonance tradition, recent frameworks in social psychology cen-
ter precisely on the link between expectation violation (or uncer-
tainty) and anxiety, with much of human cognition and behavior
interpreted as efforts to reestablish a coherent, predictable world
model (Hirsh, Mar, & Peterson, 2012;Proulx, Inzlicht, & Harmon-
Jones, 2012).
The taxing, negative experience described in ASD as sensory
overload or oversensitivity is, according to HIPPEA, a logical
consequence of a brain continuously signaling that prediction
errors merit the recruitment of more resources for learning. The
proactive (predictive) investment of the system makes this a par-
ticularly aversive experience. Conversely, making progress in pre-
dicting the world (reducing prediction errors) may genuinely feel
rewarding. Note that not the static state of low prediction error but
rather the transition (change) from a state of high prediction errors
to a state of low errors may induce positive affect (Joffily &
Coricelli, 2013;Oudeyer et al., 2010;Van de Cruys & Wagemans,
2011). This kind of reward arguably is the driving force for further
exploration and learning (cf. the Development and Exploration
section). However, difficulties in estimating where predictive
progress can be made could largely rob a person from experiencing
this type of reward, with detrimental implications for intrinsic
motivation. Indeed, problems in general motivation and explora-
tion are reported in ASD (Koegel & Mentis, 1985;Ozonoff et al.,
2008), from very early on in development (Zwaigenbaum et al.,
2005).
The combination of increased uncertainty-related anxiety and
decreased reward of exploration may have particularly incapaci-
tating and far-reaching effects in the longer term. We have referred
to learned helplessness to indicate the anxious avoidance and lack
of motivation caused by repeated frustration in predicting one’s
surroundings. By caregivers this may be interpreted as hyporeac-
tivity (Ben-Sasson et al., 2009;Tomchek & Dunn, 2007). Social
interactions might suffer most from this lack of motivation (Che-
vallier, Grezes, Molesworth, Berthoz, & Happé, 2012), with ob-
vious consequences with regard to the willingness to engage in
social relations. We do not consider social motivation problems to
be the origin of ASD, but our account agrees with social motiva-
tion theories (Chevallier et al., 2012) that this is an important
aggravating factor in the syndrome. Indeed, social interactions are
not perceived to be that enjoyable or rewarding in individuals with
ASD (Chevallier et al., 2012). Unsurprisingly, a lot of interven-
tions focus on increasing the reward of social interactions. If social
situations are avoided from early on in life, the number of social
learning experiences decreases, and so, in a vicious circle, even
more social impairments ensue.
Taken together, these factors arguably make individuals with
ASD more vulnerable to mood and anxiety problems, which are
indeed overrepresented in ASD (Kim, Szatmari, Bryson, Streiner,
& Wilson, 2000). Hence, mood problems, anxiety, and anxious
avoidance should in our view be considered as secondary symp-
toms, originating from accumulated experience with (irreducible)
prediction errors and from repeated frustration in learning. Con-
sistent with this, anxiety and mood problems seem to increase
during childhood in ASD (Kim et al., 2000).
Social Functioning
Social interaction problems are among the first described symp-
toms of ASD (Asperger, 1944;Kanner, 1943) and are crucial
pieces in the Diagnostic and Statistical Manual of Mental Disor-
ders classification (American Psychiatric Association, 2013). So-
cial impairments stand out strongly in the clinical phenotype
(demonstrated by the existence of ASD questionnaires focusing
merely on the social symptoms; e.g., Constantino, 2002), and
retrospective studies often report early signals in the social domain
(Volkmar, Chawarska, & Klin, 2005). The phenomenal and clin-
ical prominence of social deficits spurred a wealth of evidence on
social impairments. Therefore, a central challenge for core infor-
mation processing dysfunction theories of ASD is to explain why
abnormalities manifest themselves most clearly in the social do-
main.
What sets social situations apart from nonsocial situations? Or
better, what distinguishes social tasks in the lab from the tasks used
for other (lower level) domains? Like Simmons et al. (2009),we
wonder whether social may just be a synonym of complex here.
However, our approach allows us to pinpoint exactly what this
complexity may entail with regard to the difficulties in ASD. Most
ingredients have been provided in the previous sections, but in the
social domain they come together and are expressed to the fullest.
Social Complexity
Our brief overview of face and speech processing impairments
in ASD did not strongly speak for a special status of faces or
speech as such. Here, too, we do not want to treat social judge-
ments differently from other processing. It is more fitting, we
argue, to view them as just another kind of inference, in this case
inference about other people’s emotions or intentions from their
facial expressions, gaze, bodily postures, and so on (Hohwy &
Palmer, 2014;Zaki, 2013). Therefore, the same mix of accidental
uncertainty and informative changes determines the social prob-
lems in ASD. No two social scenarios are identical. Numerous
accidental properties in the rich social environment are mostly
uninformative and should be ignored. This is ideally what tuning
down precision should accomplish. Individuals should (meta-)
learn which aspects are informative and which are irrelevant to the
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
661
PREDICTIVE CODING IN AUTISM
social rules governing the current situation. This is particularly diffi-
cult when these noisy social contingencies are changing and context
bound, which they mostly are (Barrett, Mesquita, & Gendron, 2011).
There is rarely a one-to-one mapping between social signals and their
meaning. For example, happiness can be expressed with an obvious
loud laughter, but an enigmatic Mona Lisa smile is possible too. A
similar laugh can signify consent (humor) or rejection (irony). Sub-
cultures (e.g., youth culture) invent new meanings for old signals
(e.g., words) or new signals for old meanings. In addition, low-level
input can be dramatically different while the same social rules apply.
Instead of flexibly adjusting the precision of prediction errors based
on previous and current experiences, individuals with ASD will get
flooded by the wealth of available information in a social situation.
Generalizing what we said about face perception, people with
ASD fail to discriminate between informative and irrelevant prop-
erties when making social judgments (cf. the lack of autonomous
selection in attention). The result is that social information does
not seem to be particularly salient for them or at least not more so
than nonsocial stimuli. This deficit is most clearly illustrated by
eye-movement studies. Individuals with ASD show a reduced
attention to faces but more attention toward bodies and objects in
the background of a social scene (e.g., Klin, Jones, Schultz, Volk-
mar, & Cohen, 2002;Rice, Moriuchi, Jones, & Klin, 2012). Within
faces too, differences in information selection are noticeable. They
also do not seem to have learned the typical informativeness of the
eyes region, crucial for face and emotion recognition. Instead,
studies reveal a bias for the mouth region and scanning patterns
toward the outer face characteristics (such as hair; Harms, Martin,
& Wallace, 2010). From early childhood on, children with ASD do
not show the usual preference for social stimuli (Klin, 1991,1992).
Two-year-old children with ASD rather attend to nonsocial phys-
ical contingencies instead of socially relevant biological motion
(Klin, Lin, Gorrindo, Ramsay, & Jones, 2009). We think this
should be explained by the steadier, lower level predictability of
the former.
The fact that the atypical viewing patterns and the emotion
recognition deficits are most apparent when using complex stim-
ulus material (Chevallier et al., 2012;Harms et al., 2010) also
speaks for our hypothesis. Although the distinction between rele-
vant and irrelevant information may be rather clear-cut in simple
social stimuli (e.g., isolated, well-controlled, and reused faces),
using ecologically valid stimuli (e.g., noisy, dynamic social
scenes) implies more competition from distracting (irrelevant)
information.
Gradually, TD children form “social scripts”: abstracted and
broadly applicable knowledge structures, representing an orga-
nized sequence of actions, causes and consequences within a
certain social context (e.g., making friends). In children with ASD
this capacity to generate adequate social scripts is found to be
impaired (Loth, Happé, & Gomez, 2010). It is easy to see that
indiscriminate precision of social and nonsocial cues results in
narrow and specific social scripts (e.g., making friends when I’m
playing soccer), wrought with spurious, concrete features. Inter-
ventions that try to remedy social script deficits select and describe
the relevant cues for a given script, linking it with possible appro-
priate responses (for a meta-analysis, see Reynhout & Carter,
2011).
Multisensory Integration
Adequate social understanding heavily relies on integration of
multiple sources of information, both within modality and across
modalities. The same facial expressions can receive completely
opposite meanings depending on the bodily context in which they
appear (Aviezer, Trope, & Todorov, 2012). In other situations
different modalities provide complementary information, to be
used to figure out emotions and intentions from face-to-face-
communication. In such cases, additional information of another
modality helps the interpretation. For instance, visual articulatory
information aids speech perception, especially under noisy circum-
stances. Again we note that uncertainty of the different sources has
to be taken into account in order to determine which information
should have more say in the eventual social judgment. Indeed, this
can be formalized with Bayes’ theorem (Zaki, 2013), which is
already widely used in (nonsocial) perceptual cue integration stud-
ies. For optimal inference, the expected uncertainty (precision) of
the different sensory sources should determine differential reliance
(weight) on those sources.
Individuals with ASD are known to have difficulties with such
multisensory integration (Iarocci & McDonald, 2006), for instance
with the detection of intermodal correspondence of facial and
vocal affect (e.g., Loveland et al., 1995). If precision is fixed at a
similarly high level for all sources, as HIPPEA maintains, optimal
integration will not take place, because all cues, even redundant or
very uncertain ones, will be weighed equally. Moreover, the spa-
tiotemporal contiguity of two inputs required to perceive them as
belonging to the same distal cause would be more strictly defined
for people with high precision. Any minor spatiotemporal mis-
match between two cues (e.g., visual-auditory in the ventriloquist
effect or visual-haptic in the rubber hand illusion) will render it
more likely that these will be experienced as distinct unimodal
events rather than an integrated, multimodal event (Palmer et al.,
2013). The attenuated McGurk effect found in ASD could be
similarly explained (Mongillo et al., 2008;Taylor, Isaac, & Milne,
2010).
Mentalizing
Theory of mind or “mentalizing,” refers to the ability to read the
(facial) expressions of other people; to understand their feelings,
intentions, wishes and thoughts; and to use this—mostly implic-
it—knowledge to understand another individual’s actions and
guide one’s own actions (Premack & Woodruff, 1978). A vast
amount of research in ASD has focused on the theory-of-mind
problems in individuals with ASD, arguing that individuals with
ASD have difficulties in placing themselves into the mental world
of others and themselves (sometimes described as “mindblind-
ness”; e.g., Baron-Cohen, 2001;Frith & Frith, 2003). The discov-
ery of mirror neurons (in monkeys) that are active both during the
action observation or imagination (offline processing) and during
the online execution of an action (e.g., Kohler et al., 2002) led to
the conjecture that action-understanding and even mentalizing
crucially rely on this class of neurons. All discussions on the
precise role and distribution of mirror neurons in the brain aside,
this finding conclusively showed that action execution and action
perception are closely intertwined. Predictive coding offers a new
perspective on the implementation of goal and intention inference
in the mirror system (e.g., Friston, Lawson, & Frith, 2013;Kilner,
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
662 VAN DE CRUYS ET AL.
Friston, & Frith, 2007;Koster-Hale & Saxe, 2013). As mentioned
before, actions could be conceived of as a series of hierarchical
predictions (Hamilton, Brindley, & Frith, 2007), going from longer
term intentions and goals (e.g., to splash water in your friend’s
face) over short-term goals (e.g., to grasp a glass of water) and
motor plans (movement sequences), down to the muscle com-
mands and kinesthetics. At all levels, predictions will be matched
with input, resulting in prediction errors that drive and guide
proper action execution. Of importance, in predictive coding this
hierarchical model is used for forward action generation and also
serves inverse inference: figuring out goals from observed actions
(Kilner et al., 2007). Observed actions will both automatically
generate expectations on the kinematics and muscle activation
linked to it and create discrepancies that can be explained away
only by inferring an appropriate intention on the highest levels.
How can this system distinguish then between own actions and
another’s actions? Put differently, observed action creates predic-
tion errors because motor plans and goals are generated, but
muscle and kinesthetics are inactive. How does the brain avoid
automatically executing (mimicking) observed actions to reduce
those low-level prediction errors?
The assumed mechanism is, again, precision (Clark, 2013a).
The prediction errors have a high expected precision, which makes
sure actions you initiate yourself are properly executed. These
prediction errors will be suppressed by your own accurate predic-
tions (goals), often inciting a sense of confidence or agency (see
Sensorimotor Abilities and a Sense of Self). For action observa-
tion, however, estimated precision of motor prediction errors
should be tuned down, such that they receive low weight and the
thrust of processing moves to higher level inference of goals and
intentions. In this way, precision becomes the mechanism that
allows organisms to exploit the learned hierarchical models for
action execution also for mentalizing and offline planning (Clark,
2013a).
Following this reasoning, a deficit in the flexible tuning of
precision of prediction errors, resulting in an overly high estima-
tion of precision, as HIPPEA assumes to be the case in ASD, may
give rise to a couple of related problems. First, it may contribute to
offline (motor) planning problems (Booth, Charlton, Hughes, &
Happé, 2003;Hughes, 1996), with high precision preventing in-
dividuals from transcending the immediate input, as noted earlier.
Second, failure to lower the precision of low-level prediction
errors during action observation may automatically lead to precise
proprioceptive prediction errors, because the action is not exe-
cuted. A possible strategy to reduce these errors is the mimicking
of (formal aspects of) others’ behavior. Indeed, hyperimitation of
formal aspects of behavior (Bird, Leighton, Press, & Heyes, 2007;
Spengler, Bird, & Brass, 2010) and echolalia and echopraxia, the
automatic copying of others’ speech or behavior, occur more
frequently in the ASD population. We are cautious in pointing to
this possibility, because precision of motor errors may be deter-
mined by a different neurotransmitter (dopamine) than perceptual
errors (see the Neurobiological Underpinnings section), and not
every child with ASD shows this automatic mimicking.
A third possible problem of inflexible tuning of precision links
back to our discussion on visual and auditory perception. We noted
there that top levels of hierarchical models may not get properly
built (learned), because processing is stuck in low-level matching
due to the high precision of low-level prediction errors. If, for
motor execution and planning too, individuals with ASD end up
with incomplete hierarchical models, they may be unable to reach
the higher levels of conceptual inferences of goal and intention.
Consequentially, these individuals will experience difficulties in
inferring emotions from their own bodily states and expressions
(Seth et al., 2012). Indeed, alexithymia is often found in ASD and
has recently been shown to better predict poor recognition of
emotional expressions than ASD as such (Cook, Brewer, Shah, &
Bird, 2013). From the predictive coding standpoint, where one
model is used for both recognizing emotion and inferring own
emotion, this makes a lot of sense. Brain responses related to empathy are
also modulated by alexithymia rather than ASD (Bird et al., 2010).
If these findings are corroborated, it may turn out the empathy and
emotion recognition problems in ASD (see Harms et al., 2010;
Uljarevic & Hamilton, 2013, for a literature review and a meta-
analysis, respectively) are not primary symptoms but are inher-
ently linked to alexithymia. The processing profile of ASD as we
sketched may predispose patients to alexithymia, because high-
precision interoceptive prediction errors prevent adequate emo-
tional inferences (Seth et al., 2012).
Neurobiological Underpinnings
In this paper we primarily wanted to articulate the cognitive,
computational foundation of our account and its behavioral con-
sequences. We do want to briefly survey plausible neurobiological
underpinnings of the proposed mechanism as well, without giving
an exhaustive review of the neurobiology of ASD (see, e.g., D. G.
Amaral, Schumann, & Nordahl, 2008;Bauman & Kemper, 2005).
A more systematic, extensive discussion of ASD neurobiology in
light of HIPPEA will be needed in the future.
Using HIPPEA, we can tentatively divide neurobiological find-
ings into three parts: first, studies directly targeting the neural
regulation of precision; second, studies on the neural basis of
models of uncertainty and meta-learning that feed into regulation
of precision; and third, downstream consequences of high preci-
sion for neural plasticity and connectivity. We consider the first
two here and leave the last part for later work.
Precision Regulation
In Friston’s predictive coding model, precision is regulated by
neuromodulators that control the gain of the units representing
prediction errors (Friston, 2010). This gain determines the impact
of prediction errors on units that encode the predictions. Neuro-
modulators such as acetylcholine (ACh) and norepinephrine (NE)
have long been known to influence attention and learning, so they
are likely candidates for this role. In particular, the neuromodulator
ACh is assumed to enhance precision of perceptual prediction
errors (Friston, 2010). Indeed, a pharmacological agent that in-
creases ACh availability in cholinergic synapses increases the
event-related response to deviations of predictions (Moran et al.,
2013) and attenuates the decrease in responses with repeated
stimulation (repetition suppression). However, Yu and Dayan
(2005) proposed a different, complementary role of ACh and NE,
in which only expected uncertainty, linked to the known stochas-
ticity (lack of reliability) of a predictive relationship, is coded by
ACh. NE, on the other hand, tracks unexpected uncertainty; that is,
the actual, important changes in the regularities governing the
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
663
PREDICTIVE CODING IN AUTISM
relationships in the world (Duzel & Guitart-Masip, 2013;Payzan-
LeNestour, Dunne, Bossaerts, & O’Doherty, 2013). A context-
dependent modulation of the balance between these two must
ensure that learning is enabled when learning is due (for actual
changes).
The findings on (nicotinic) cholinergic signaling in ASD are
very inconclusive at this stage, but a few studies have reported
abnormalities (Lam, Aman, & Arnold, 2006), including in the
main source of ACh, the basal forebrain (Bauman & Kemper,
1994;Perry et al., 2001). Raised NE signaling in ASD is suggested
by elevated blood plasma levels (Lam et al., 2006) and by a
tonically high arousal system as shown by a tonically elevated
heart rate in autistic children, with reduced phasic response (Kootz
& Cohen, 1981). For pupil size the same pattern has been reported:
increased tonic pupil size and increased latency, smaller constric-
tion amplitude, and lower constriction velocity for the pupillary
light reflex compared to TD children (Anderson & Colombo,
2009;Fan, Miles, Takahashi, & Yao, 2009). This is noteworthy,
because of the known coupling of pupil size with the NE system,
more specifically with activity in the principal source of NE
projections, the locus coeruleus (Rajkowski, Kubiak, & Aston-
Jones, 1993). Finally, prenatal overstimulation of the
2
-
adrenergic receptor by an agonist is associated with increased risk
of ASD (Connors et al., 2005).
Hence, available evidence already seems to point to some loss in
the dynamic range of ACh and NE neuromodulation, but direct
tests await. Pharmacological studies applying an agent that in-
creases central cholinergic signaling should verify whether the
ERP or behavioral response to expectation violation is modulated
similarly in individuals with and without ASD (cf. Moran et al.,
2013). If cholinergic signaling is already at ceiling in ASD, an
additional boost of this system may not make a difference. Alter-
natively, a cholinergic antagonist may, in ASD, lead to “normal”
performance on tasks that benefit from disregarding smaller dif-
ferences (on which ASD subjects are usually worse). With regard
to NE, there may be considerable potential in measuring pupil
dynamics in ASD. Nassar et al. (2012) demonstrated that learning
dynamics can be tracked by pupil size measurements, suggesting
that NE arousal systems indeed can regulate learning. Their pre-
dictive inference task required adjusting of precision (learning
rate), because predictive relationships changed at certain points
(“change points”) in the course of the task, as explained before (see
Attention and Executive Functioning). Apparently, pupil diameter
change is monotonically related to change point probability, where
prediction errors should indeed receive high weight. Additionally,
average pupil size reflects “uncertainty that arises after change
points and signals the need for rapid learning” (Nassar et al., 2012,
p. 1043). Recall that this uncertainty was called reducible. If ASD
is linked to increased precision of prediction errors across the
board, as HIPPEA maintains, this should be apparent both in
average learning rate and pupil metrics in this sort of task.
Finally, there is evidence that these neuromodulators can act as
meta-plastic signals regulating the potential of synapses to undergo
activity-dependent long-term potentiation (e.g., Inoue et al., 2013).
This provides another link with precision as a meta-learning signal
that should be explored more. Indeed, several of the genetic
mutations linked to ASD have an important role in the regulation
of plasticity (e.g., Delorme et al., 2013;Ebert & Greenberg, 2013;
Hutsler & Zhang, 2010). Relatedly, the valproic acid rat model of
ASD shows twice the amount of long term potentiation of controls
(Markram & Markram, 2010).
Models of Uncertainty
We emphasized before that precision of prediction errors does
not appear out of the blue. The brain builds meta-models, predic-
tions of prediction errors, to estimate precision. These meta-
models are formally not that different from regular predictive
models assumed to take place across the perceptual hierarchy.
Arguably then, these meta-models may be represented also in a
distributed manner across the cortex. However, there is evidence
that some regions are more involved than others in the processing
of uncertainty.
Two regions that are good candidates for this and that have
recently attracted researchers’ interest in ASD are the insula and
the anterior cingulate cortex (ACC). Both are thought to be central
parts of the so-called salience network, the circuit involved in
responding to behaviorally important stimuli and in cognitive
control. Indeed, we could replace the somewhat vague term sa-
lience with precision, because in se they have similar intent;
namely, determining value or relevance of input for behavior and
learning. The salience network is closely connected to the motor
system, suggesting a role in generating exploratory actions (Rush-
worth, Behrens, Rudebeck, & Walton, 2007), as we discussed in
the Development and Exploration section on exploration in ASD.
Also, it is deemed to be crucial in judging whether to persist in or
switch the current attentional set (Dosenbach et al., 2006). Evi-
dently, models of uncertainty in input are vital in such decisions.
Finally, the ACC innervates the locus coeruleus-NE system
(Aston-Jones & Cohen, 2005), perhaps allowing it to modulate
gain (precision) of prediction errors in the (sensory) cortex.
A recent study found hyperactivation in dorsal ACC in response
to visual oddball stimuli in ASD (i.e., infrequently presented,
deviant stimulus; Dichter, Felder, & Bodfish, 2009), consistent
with the idea that expectation violations are more salient. In
healthy participants, ACC activity is found for behaviorally rele-
vant prediction errors (Ide, Shenoy, Yu, & Li, 2013;Metereau &
Dreher, 2013). Others have found evidence that the cingulate
cortex not only represents the prediction errors but also performs
the computations underlying the adaptive regulation of precision
(D. R. Bach, Hulme, Penny, & Dolan, 2011;Behrens et al., 2007).
The insula too is known to be involved in prediction under
uncertainty. Activity in ACC and insula is strongly coupled, and,
critically, this coupling is modulated by prediction errors (Li-
mongi, Sutherland, Zhu, Young, & Habib, 2013). Using a gam-
bling game, Preuschoff, Quartz, and Bossaerts (2008) showed that
activity in the anterior insula can code that part of uncertainty that
cannot be reduced, due to the stochasticity of the associations at
hand, also called known or expected uncertainty. There is evidence
from dynamic causal modeling analyses that anterior insula is the
entry point of the salience network and drives ACC activity (Ham,
Leff, de Boissezon, Joffe, & Sharp, 2013;Limongi et al., 2013). If
true, a possible hypothesis is that the insula corrects incoming
prediction errors for known stochasticity and thereby helps ACC
and further regions to properly attribute salience (precision) of the
prediction errors. In any case, insula, ACC and possibly neighbor-
ing frontal regions may cooperate to dissect uncertainty, with the
aim of estimating where predictive progress can be made and
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
664 VAN DE CRUYS ET AL.
setting precision accordingly (attending to the right aspects of
input; see also Karlsson, Tervo, & Karpova, 2012). When, as is the
case for ASD, there is abnormal connectivity and activity of ACC
and insula (Di Martino et al., 2009;Uddin & Menon, 2009), this
estimation process may go awry, leading to unadaptive and pos-
sibly chronically high precision. Much more work is needed be-
cause existing neuroimaging work in ASD mostly uses task con-
trasts (often using faces or other complex stimuli) that are hardly
informative in relation to our proposal. Fortunately, the needed
paradigms have already been applied in nonclinical participants.
Related Approaches
Several important theoretical frameworks of ASD can be use-
fully compared to ours. Some have been emphasized in previous
sections, but here we more closely look at those that were not
discussed before and that are most akin to our theorizing, in
postulating a broader information-processing account. In the sec-
ond part of this section, we address the question whether a unified
account is possible at all, in view of the heterogeneous nature of
ASD. We close this section with a discussion of a recent theory of
schizophrenia, which is closely related to our theory of ASD.
Other Information Processing Accounts of ASD
A straightforward, Bayesian way to conceptualize problems in
ASD could be to assume broader (high uncertainty) priors or
predictions that therefore have a weaker influence on the outcome
of perceptual inference. Indeed, this road has recently been taken
by Pellicano and Burr (2012) in a thought-provoking article (for a
related approach, see Gomot & Wicker, 2012). Pellicano and Burr
argued that this may cause perceptual outcomes to remain closer to
the perceptual input, minimally biased by top-down, prior knowl-
edge, an idea that is consistent with the WCC theory. Hence, this
account explains why individuals with ASD may be less suscep-
tible to visual illusions that are caused by prior knowledge or
contextual interactions (see the Perceptual Processing section). In
other words it would, according to Pellicano and Burr, result in a
more accurate or “real” perception.
In addition to spurring an interesting discussion (Brock, 2012;
Friston et al., 2013;Teufel, Subramaniam, & Fletcher, 2013;van
Boxtel & Lu, 2013;Van de Cruys et al., 2013), this stance has been
criticized on theoretical and empirical grounds. Teufel et al. (2013)
reminded us that “a perceptual system that refines sensory infor-
mation by prior knowledge provides a better estimate of real but
hidden causes than perception that is based on the ambiguous
sensory information on its own, because the former system ex-
ploits all the relevant information available.” In this regard,
broader priors would lead to less accurate perception because the
actual input is always noisy and ambiguous. Even in the case of
visual illusions, it is not “priors per se [that] render perception less
accurate; rather, it is the application of the wrong prior that leads
to the illusory percept” (Teufel et al., 2013). Furthermore, Brock
(2012) noted that perception (the posterior) can move closer to
perceptual input (likelihood) for two different reasons: either, as
Pellicano and Burr (2012) argued, the prior is broader (higher
uncertainty, lower precision) or the likelihood is sharper (lower
uncertainty, higher precision). It should be clear that the proposal
in HIPPEA is more akin the second option. Finally, there is
evidence that individuals with ASD are very well capable of
building precise expectations from experience (see the Attention
and Executive Functioning section). Indeed that may be the reason
why they are so perturbed by information that deviates from this
information. The problems, we argue, arise because these devia-
tions receive too much salience. Instead of a lack of precision in
predictions, there may be a heightened precision of prediction
errors in ASD.
It is also interesting to distinguish our view from approaches
locating the core problem in ASD in a reduced signal-to-noise ratio
in neural processing (Belmonte et al., 2004;Simmons et al., 2009).
Although increasing noise usually impairs psychophysical perfor-
mance, it can improve detection under restricted conditions, a
phenomenon called stochastic resonance (Goris, Wagemans, &
Wichmann, 2008). Though speculative at this stage, increased
internal noise in neural communication may in this way be able to
explain both improved performance in a limited number of tasks
and impaired performance on more complex, high-level tasks
(Simmons et al., 2009). HIPPEA in contrast, does not necessarily
assume increased internal noise in neural signaling but rather a
higher weighing of external and internal “noise” (accidental fea-
tures), causing the system to attempt to capture this irrelevant,
nonrepeating noise. We believe that this view is more readily
compatible with the broad range of behavioral peculiarities in
ASD.
Unifying Theories of ASD, in the Face of Its Genetic
and Phenotypic Heterogeneity
Several scholars have lamented the overgrowth of unifying
theories on ASD, seeing that they fail to deliver a convincing
account for every ASD symptom cluster. Heterogeneity in under-
lying genetics similarly seems to suggest that there is not one but
rather a multitude of deficits underlying the ASD pathology
(Happé & Ronald, 2008). Finally and most important, phenotypic
variability is notorious in ASD (Rommelse, Geurts, Franke, Bu-
itelaar, & Hartman, 2011). This causes but may also be caused by
difficulties in diagnosing ASD. Bringing the view of ASD as
singular entity even further into question is the fact that “virtually
every symptom characteristic of ASD can be observed in children
who do not fit this diagnostic category” (Bishop, 1989, The Bor-
derlands of Autism section, para. 2). This of course, does not
necessarily imply that these symptoms when they appear together
in ASD are just the result of the “worst of luck.” Still, these
observations have led Happé and Ronald (2008) to describe ASD
as a fractionable triad, with three independent components (com-
munication problems, social interaction deficits, and repetitive and
restricted behaviors and interests) coincidentally co-occurring.
Only when the three conspire do subclinical signs become clinical
symptoms meriting a diagnosis.
Naturally, we agree with Happé and Ronald (2008) that a better
characterization of the subcomponents of ASD is much needed,
but an intrinsic coherence of the components may shine through
only when the appropriate level of description has been found. As
we progress toward more realistic models of the mind-brain, we
may be able to formulate more fitting explanations of ASD within
these broader models. HIPPEA can be considered a first step in
that direction.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
665
PREDICTIVE CODING IN AUTISM
Furthermore, there may be more coherence in the ASD symp-
tom clusters than these critical authors assume. For example,
although executive functioning and attentional deficits may not be
specific to ASD (cf. attention-deficit/hyperactivity disorder), the
specific pattern of executive capacities impaired and intact may be
distinguishable from other disorders and may have a privileged
relationship with social or emotional symptoms of ASD. It is no
doubt a challenge to connect social and communicative symptoms
to more basic processing differences, due to divergence in the
pathways leading to such high-level dysfunctions and to possible
compensatory mechanisms saving these capacities for others. In-
deed, a truly developmental account such as HIPPEA will predict
quite some variability in the unfolding of clinical symptoms de-
pending on interactions with the environment.
Finally, heterogeneity in underlying (epi)genetic and molecular
paths toward the syndrome does not preclude the possibility that
one main cognitive mechanism is impaired. There is little reason to
expect a one-to-one mapping from cognitive processing to neuro-
biology. The previous section provided possible ways HIPPEA
links up with neurobiological evidence.
ASD in Relation to Schizophrenia
Increasing evidence suggests that ASD has common genetic risk
factors and neuroanatomical overlap with schizophrenia (Carroll &
Owen, 2009;Cheung et al., 2010;Serretti & Fabbri, 2013). In-
triguingly, a recent theory of schizophrenia (Adams, Stephan,
Brown, Frith, & Friston, 2013;Fletcher & Frith, 2009) invoked
undue high precision of prediction errors to explain positive symp-
toms in schizophrenia (hallucinations and delusions). The authors
proposed that high-precision prediction errors cannot be reduced
and are propagated to higher levels, where they induce radical
updates of beliefs to somehow make sense of them. Hence, they
result in the strange worldviews and delusions.
It seems to us that inflexible, high-precision prediction errors are
a better fitting explanation for ASD than they are for schizophre-
nia. Overprecise prediction errors as a fundamental, indeed devel-
opmental, characteristic would be present from very early on in
life. Hence, the relatively late onset of schizophrenia needs ex-
plaining. Also, overly high precise prediction errors arguably do
not sufficiently explain the specific, improbable, and utterly bi-
zarre contents of delusional beliefs (Silverstein, 2013). Other
things that may be important to consider are the specific level of
origin of the prediction errors (conceptual or action vs perceptual
prediction errors; Adams et al., 2013;Fletcher & Frith, 2009) and
the subjective confidence level (precision) that top-down beliefs
can take on (to explain their fervor).
Although the cognitive commonality of schizophrenia and ASD
may match their genetic and neuroanatomical overlap, it also
highlights a central challenge for predictive coding theories of
mental illnesses: If they want to provide more than overaccom-
modating just-so stories for mental disorders, these theories should
be able to give good, constraining explanations for the cognitive
and neural specificities of each disorder. More work is clearly
needed in this respect.
Conclusions
Although one core deficit is unlikely to explain all heterogeneity
in ASD, it is quite remarkable that our approach can accommodate
a broad range of reported deficits and peculiarities. This also
makes sense because meta-learning is central in development
across domains. Meta-cognition, conceptualized as the ability to
monitor and adaptively use uncertainty, is generally fragile, costly,
and conclusively demonstrated in only a few, cognitively higher
developed species (Carruthers, 2008;J. D. Smith, 2009;J. D.
Smith, Coutinho, Church, & Beran, 2013). Dysfunction of this
capacity may impact higher level functions such as emotion pro-
cessing and social cognition, but it also has a pervasive effect on
attention, cognitive control, perception, and learning. Hence, HIP-
PEA is broader than earlier single-deficit accounts of ASD, be-
cause it is not linked to a certain symptom cluster. At the same
time, however, HIPPEA is more specific than those accounts,
homing in on the disturbed mechanism.
Every existing neurocognitive theory is criticized for not being
universal and not being specific for ASD. How does how HIPPEA
fare on those accounts? First, does HIPPEA maintain that every
individual with ASD shows inflexibly high precision of prediction
errors (universality)? We argue that this is indeed the case but
leave room for two ways to arrive at this high precision: a direct,
possibly neuromodulatory, deficit in the precision mechanism or a
deficit in the extraction from experience of information that should
be used to estimate precision (meta-learning). Second, does every
individual with chronically high-precision prediction errors suffer
from ASD (specificity)? Again, we answer positively but with the
important qualification that HIPPEA is consistent with the exis-
tence of a spectrum of ASD traits. It distinguishes different per-
ceptual, cognitive, emotional, and social processes according to the
extent to which they can be affected by chronically high-precision
errors. This naturally leads to the notion of a spectrum. Just how
high and how fixed precision is determines whether normal func-
tioning is still possible. Indeed, some people may be able to turn
their “deficit” into an asset in tasks that benefit greatly from their
specific processing style (Gonzalez, Martin, Minshew, &
Behrmann, 2013).
Evidence-based treatments and psychoeducation for ASD that
focus on early learning (such as applied behavioral analysis,
Lovaas, 1987;Rogers & Vismara, 2008) could take inspiration
from HIPPEA, which also has learning at its core but demarcates
the circumstances under which problems in ASD arise. Animal
models of ASD-related diseases show that environmental enrich-
ment can reduce risk of developmental disorders (G. Dawson,
2008). We also remarked that people with ASD may be able to
learn and use high-level predictions, given extensive exposure to
more and different situations. However, most of all, our approach
reaffirms the importance of more scaffolding during learning (e.g.,
Bellon, Ogletree, & Harn, 2000;Odom et al., 2003). Our section
on exploration (Development and Exploration) made it clear that
children with ASD need more support with the gradual progression
from simple to naturalistic stimuli (e.g., using virtual environ-
ments), taking into account uncertainty and its causes. Finally and
slightly counterintuitively, reducing intense concentration on
learning experiences, thus preventing individuals from trying to
match all details (“early stopping”), has also been proposed to be
beneficial (Bakouie et al., 2009).
Although we consider HIPPEA to be a rich and promising
theory, much of what we have offered here is post hoc. The
specific theory of ASD we proposed in this paper is based on
predictive coding in normal functioning, but so far most of the
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
666 VAN DE CRUYS ET AL.
explanatory power is in our selective (albeit broad) synthesis of the
literature on ASD. Future research will have to corroborate its
unique predictive power. In the preceding sections, we have often
added comments about shortcomings in the current literature as
well as specific hypotheses derived from our theory that remain to
be tested. With a very general theory like predictive coding, there
is always a risk of nonfalsifiability (see also the extensive discus-
sion sparked by Clark, 2013b), but we are convinced that our
theory of predictive coding in ASD is specific enough to be
testable. Although we mainly addressed the functional (psycho-
logical) level in this paper, we are optimistic that HIPPEA is at
least compatible with an explanation at the neural level. We hope
the progress that is currently being made in filling in the neural
mechanisms behind predictive coding will help answer the ques-
tion of precisely why individuals with ASD end up with high,
inflexible precision.
In sum, our intent with this paper was to sketch the breadth of
implications of HIPPEA with regard to aberrant development and
to point to new empirical questions for ASD research flowing from
this view. Ultimately, this will give us a better handle on ASD,
connecting clinical to neurobiological descriptions and providing a
firmer foundation for treatment.
References
Adams, R. A., Stephan, K. E., Brown, H. R., Frith, C. D., & Friston, K. J.
(2013). The computational anatomy of psychosis. Frontiers in Psychi-
atry, 4, Article 47. doi:10.3389/fpsyt.2013.00047
Alderson-Day, B., & McGonigle-Chalmers, M. (2011). Is it a bird? Is it a
plane? Category use in problem-solving in children with autism spec-
trum disorders. Journal of Autism and Developmental Disorders, 41,
555–565. doi:10.1007/s10803-010-1077-9
Amaral, D. G., Schumann, C. M., & Nordahl, C. W. (2008). Neuroanatomy
of autism. Trends in Neurosciences, 31, 137–145. doi:10.1016/j.tins
.2007.12.005
Amaral, J. L., Collins, S., Bohache, K. T., & Kloos, H. (2012). Beyond the
black-and-white of autism: How cognitive performance varies with
context. In H. Kloos, B. J. Morris, & J. L. Amaral (Eds.), Current topics
in children’s learning and cognition (pp. 145–170). doi:10.5772/53937
American Psychiatric Association. (2013). Diagnostic and statistical man-
ual of mental disorders (5th ed.). Arlington, VA: American Psychiatric
Publishing.
Anderson, C. J., & Colombo, J. (2009). Larger tonic pupil size in young
children with autism spectrum disorder. Developmental Psychobiology,
51, 207–211. doi:10.1002/dev.20352
Apps, M. A. J., & Tsakiris, M. (2013). The free-energy self: A predictive
coding account of self-recognition. Neuroscience & Biobehavioral Re-
views, 41, 85–97. doi:10.1016/j.neubiorev.2013.01.029
Arnal, L. H., & Giraud, A. L. (2012). Cortical oscillations and sensory
predictions. Trends in Cognitive Sciences, 16, 390–398. doi:10.1016/j
.tics.2012.05.003
Ashwin, E., Ashwin, C., Rhydderch, D., Howells, J., & Baron-Cohen, S.
(2009). Eagle-eyed visual acuity: An experimental investigation of en-
hanced perception in autism. Biological Psychiatry, 65, 17–21. doi:
10.1016/j.biopsych.2008.06.012
Asperger, H. (1944). Die “autistischen psychopathen” im kindesalter [Au-
tistic psychopaths in childhood]. Archiv für Psychiatrie und Nervenk-
rankheiten, 117, 76–136.
Aston-Jones, G., & Cohen, J. D. (2005). An integrative theory of locus
coeruleus-norepinephrine function: Adaptive gain and optimal perfor-
mance. Annual Review of Neuroscience, 28, 403–450. doi:10.1146/
annurev.neuro.28.061604.135709
Aviezer, H., Trope, Y., & Todorov, A. (2012, November 30). Body cues,
not facial expressions, discriminate between intense positive and nega-
tive emotions. Science, 338, 1225–1229. doi:10.1126/science.1224313
Bach, D. R., Hulme, O., Penny, W. D., & Dolan, R. J. (2011). The known
unknowns: Neural representation of second-order uncertainty, and am-
biguity. Journal of Neuroscience, 31, 4811–4820. doi:10.1523/
JNEUROSCI.1452-10.2011
Bach, M., & Dakin, S. C. (2009). Regarding “Eagle-Eyed Visual Acuity:
An Experimental Investigation of Enhanced Perception in Autism.”
Biological Psychiatry, 66, e19e20. doi:10.1016/j.biopsych.2009.02
.035
Baird, G., Simonoff, E., Pickles, A., Chandler, S., Loucas, T., Meldrum,
D., & Charman, T. (2006, July 15). Prevalence of disorders of the autism
spectrum in a population cohort of children in South Thames: The
Special Needs and Autism Project (SNAP). Lancet, 368, 210–215.
doi:10.1016/S0140-6736(06)69041-7
Bakouie, F., Zendehrouh, S., & Gharibzadeh, S. (2009). Does a kind of
over-fitting occur in the brain of autistic patients? Journal of Neuropsy-
chiatry and Clinical Neurosciences, 21, Article 343. doi:10.1176/appi
.neuropsych.21.3.343
Bar, M. (2009). The proactive brain: Memory for predictions. Philosoph-
ical Transactions of the Royal Society B: Biological Sciences, 364,
1235–1243. doi:10.1098/rstb.2008.0310
Baron-Cohen, S. (2001). Theory of mind and autism: A review. Interna-
tional Review of Research in Mental Retardation, 23, 169–184.
Baron-Cohen, S. (2006). The hyper-systemizing, assortative mating theory
of autism. Progress in Neuro-Psychopharmacology and Biological Psy-
chiatry, 30, 865–872. doi:10.1016/j.pnpbp.2006.01.010
Baron-Cohen, S., Ashwin, E., Ashwin, C., Tavassoli, T., & Chakrabarti, B.
(2009). Talent in autism: Hyper-systemizing, hyper-attention to detail
and sensory hypersensitivity. Philosophical Transactions of the Royal
Society B: Biological Sciences, 364, 1377–1383. doi:10.1098/rstb.2008
.0337
Baron-Cohen, S., Ring, H., Bullmore, E., Wheelwright, S., Ashwin, C., &
Williams, S. (2000). The amygdala theory of autism. Neuroscience &
Biobehavioral Reviews, 24, 355–364. doi:10.1016/S0149-
7634(00)00011-7
Barrett, L. F., Mesquita, B., & Gendron, M. (2011). Context in emotion
perception. Current Directions in Psychological Science, 20, 286–290.
doi:10.1177/0963721411422522
Bauman, M. L., & Kemper, T. L. (1994). Neuroanatomic observations of
the brain in autism. In M. L. Bauman & T. L. Kemper (Eds.), The
neurobiology of autism (pp. 119–145). Baltimore, MD: Johns Hopkins
Press.
Bauman, M. L., & Kemper, T. L. (Eds.). (2005). The neurobiology of
autism (2nd ed.). Baltimore, MD: Johns Hopkins Press.
Behrens, T. E. J., Woolrich, M. W., Walton, M. E., & Rushworth, M. F. S.
(2007). Learning the value of information in an uncertain world. Nature
Neuroscience, 10, 1214–1221. doi:10.1038/nn1954
Behrmann, M., Thomas, C., & Humphreys, K. (2006). Seeing it differently:
Visual processing in autism. Trends in Cognitive Sciences, 10, 258–264.
doi:10.1016/j.tics.2006.05.00
Bellon, M. L., Ogletree, B. T., & Harn, W. E. (2000). Repeated storybook
reading as a language intervention for children with autism: A case study
on the application of scaffolding. Focus on Autism and Other Develop-
mental Disabilities, 15, 52–58. doi:10.1177/108835760001500107
Belmonte, M. K., Cook, E. H., Anderson, G. M., Rubenstein, J. L. R.,
Greenough, W. T., Beckel-Mitchener, A.,...Tierney, E. (2004). Autism
as a disorder of neural information processing: Directions for research
and targets for therapy. Molecular Psychiatry, 9, 646663. doi:10.1038/
sj.mp.4001499
Ben-Sasson, A., Hen, L., Fluss, R., Cermak, S., Engel-Yeger, B., & Gal, E.
(2009). A meta-analysis of sensory modulation symptoms in individuals
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
667
PREDICTIVE CODING IN AUTISM
with autism spectrum disorders. Journal of Autism and Developmental
Disorders, 39, 1–11. doi:10.1007/s10803-008-0593-3
Berlyne, D. E. (1966, July 1). Curiosity and exploration. Science, 153,
25–33. doi:10.1126/science.153.3731.25
Bertone, A., Mottron, L., Jelenic, P., & Faubert, J. (2003). Motion percep-
tion in autism: A “complex” issue. Journal of Cognitive Neuroscience,
15, 218–225. doi:10.1162/089892903321208150
Bertone, A., Mottron, L., Jelenic, P., & Faubert, J. (2005). Enhanced and
diminished visuo-spatial information processing in autism depends on
stimulus complexity. Brain, 128, 2430–2441. doi:10.1093/brain/awh561
Bird, G., Leighton, J., Press, C., & Heyes, C. (2007). Intact automatic
imitation of human and robot actions in autism spectrum disorders.
Proceedings of the Royal Society B: Biological Sciences, 274, 3027–
3031. doi:10.1098/rspb.2007.1019
Bird, G., Silani, G., Brindley, R., White, S., Frith, U., & Singer, T. (2010).
Empathic brain responses in insula are modulated by levels of alexithy-
mia but not autism. Brain, 133, 1515–1525. doi:10.1093/brain/awq060
Bishop, D. V. (1989). Autism, Asperger’s syndrome and semantic-
pragmatic disorder: Where are the boundaries? British Journal of Dis-
orders of Communication, 24, 107–121.
Blakemore, S., Tavassoli, T., Calò, S., Thomas, R., Catmur, C., Frith, U.,
& Haggard, P. (2006). Tactile sensitivity in Asperger syndrome. Brain
and Cognition, 61, 5–13. doi:10.1016/j.bandc.2005.12.013
Boll, S., Gamer, M., Gluth, S., Finsterbusch, J., & Büchel, C. (2013).
Separate amygdala subregions signal surprise and predictiveness during
associative fear learning in humans. European Journal of Neuroscience,
37, 758–767. doi:10.1111/ejn.12094
Bölte, S., Holtmann, M., Poustka, F., Scheurich, A., & Schmidt, L. (2007).
Gestalt perception and local–global processing in high-functioning au-
tism. Journal of Autism and Developmental Disorders, 37, 1493–1504.
doi:10.1007/s10803-006-0231-x
Bölte, S., Hubl, D., Dierks, T., Holtmann, M., & Poustka, F. (2008). An
fMRI-study of locally oriented perception in autism: Altered early visual
processing of the block design test. Journal of Neural Transmission,
115, 545–552. doi:10.1007/s00702-007-0850-1
Bölte, S., Schlitt, S., Gapp, V., Hainz, D., Schirman, S., Poustka, F.,...
Walter, H. (2012). A close eye on the eagle-eyed visual acuity hypoth-
esis of autism. Journal of Autism and Developmental Disorders, 42,
726–733. doi:10.1007/s10803-011-1300-3
Bonnel, A., McAdams, S., Smith, B., Berthiaume, C., Bertone, A., Ciocca,
V.,...Mottron, L. (2010). Enhanced pure-tone pitch discrimination
among persons with autism but not Asperger syndrome. Neuropsycho-
logia, 48, 2465–2475. doi:10.1016/j.neuropsychologia.2010.04.020
Bonnel, A., Mottron, L., Peretz, I., Trudel, M., Gallun, E., & Bonnel, A.
(2003). Enhanced pitch sensitivity in individuals with autism: A signal
detection analysis. Journal of Cognitive Neuroscience, 15, 226–235.
doi:10.1162/089892903321208169
Booth, R., Charlton, R., Hughes, C., & Happé, F. (2003). Disentangling
weak coherence and executive dysfunction: Planning drawing in autism
and attention-deficit/hyperactivity disorder. Philosophical Transactions
of the Royal Society B: Biological Sciences, 358, 387–392. doi:10.1098/
rstb.2002.1204
Botvinick, M., & Cohen, J. (1998, February 19). Rubber hands “feel” touch
that eyes see. Nature, 391, 756. doi:10.1038/35784
Brincker, M., & Torres, E. B. (2013). Noise from the periphery in autism.
Frontiers in Integrative Neuroscience, 7, Article 34. doi:10.3389/fnint
.2013.00034
Brock, J. (2012). Alternative Bayesian accounts of autistic perception:
Comment on Pellicano and Burr. Trends in Cognitive Sciences, 16,
573–574; author reply 574–575. doi:10.1016/j.tics.2012.10.005
Brown, C., Gruber, T., Boucher, J., Rippon, G., & Brock, J. (2005).
Gamma abnormalities during perception of illusory figures in autism.
Cortex, 41, 364–376. doi:10.1016/S0010-9452(08)70273-9
Brown, J., Aczel, B., Jiménez, L., Kaufman, S. B., & Grant, K. P. (2010).
Intact implicit learning in autism spectrum conditions. Quarterly Jour-
nal of Experimental Psychology, 63, 1789–1812. doi:10.1080/
17470210903536910
Burack, J. A. (1994). Selective attention deficits in persons with autism:
Preliminary evidence of an inefficient attentional lens. Journal of Ab-
normal Psychology, 103, 535–543. doi:10.1037/0021-843X.103.3.535
Carroll, L. S., & Owen, M. J. (2009). Genetic overlap between autism,
schizophrenia and bipolar disorder. Genome Medicine, 1, 1–7. doi:
10.1186/gm102
Carruthers, P. (2008). Meta-cognition in animals: A skeptical look. Mind &
Language, 23, 5889. doi:10.1111/j.1468-0017.2007.00329.x
Cheung, C., Yu, K., Fung, G., Leung, M., Wong, C., Li, Q.,...McAlonan,
G. (2010). Autistic disorders and schizophrenia: Related or remote? An
anatomical likelihood estimation. PLoS ONE, 5(8), e12233. doi:
10.1371/journal.pone.0012233
Chevallier, C., Grezes, J., Molesworth, C., Berthoz, S., & Happé, F. (2012).
Brief report: Selective social anhedonia in high functioning autism.
Journal of Autism and Developmental Disorders, 42, 1504–1509. doi:
10.1007/s10803-011-1364-0
Clark, A. (2013a). The many faces of precision. Frontiers in Psychology,
4, Article 270. doi:10.3389/fpsyg.2013.00270
Clark, A. (2013b). Whatever next? Predictive brains, situated agents, and
the future of cognitive science. Behavioral and Brain Sciences, 36,
181–204. doi:10.1017/S0140525X12000477
Connors, S. L., Crowell, D. E., Eberhart, C. G., Copeland, J., Newschaffer,
C. J., Spence, S. J., & Zimmerman, A. W. (2005).
2
-adrenergic receptor
activation and genetic polymorphisms in autism: Data from dizygotic
twins. Journal of Child Neurology, 20, 876884. doi:10.1177/
08830738050200110401
Constantino, J. N. (2002). The Social Responsiveness Scale. Los Angeles,
CA: Western Psychological Services.
Cook, R., Brewer, R., Shah, P., & Bird, G. (2013). Alexithymia, not
autism, predicts poor recognition of emotional facial expressions. Psy-
chological Science, 24, 723–732. doi:10.1177/0956797612463582
Courville, A. C., Daw, N. D., & Touretzky, D. S. (2006). Bayesian theories
of conditioning in a changing world. Trends in Cognitive Sciences, 10,
294–300. doi:10.1016/j.tics.2006.05.004
Crespi, B. (2013). Developmental heterochrony and the evolution of au-
tistic perception, cognition and behavior. BMC Medicine, 11, Article
119. doi:10.1186/1741-7015-11-119
Dakin, S., & Frith, U. (2005). Vagaries of visual perception in autism.
Neuron, 48, 497–507. doi:10.1016/j.neuron.2005.10.018
Daw, N. D., Niv, Y., & Dayan, P. (2005). Uncertainty-based competition
between prefrontal and dorsolateral striatal systems for behavioral con-
trol. Nature Neuroscience, 8, 1704–1711. doi:10.1038/nn1560
Dawson, G. (2008). Early behavioral intervention, brain plasticity, and the
prevention of autism spectrum disorder. Development and Psychopa-
thology, 20, 775–803. doi:10.1017/S0954579408000370
Dawson, G., Finley, C., Phillips, S., Galpert, L., & Lewy, A. (1988).
Reduced P3 amplitude of the event-related brain potential: Its relation-
ship to language ability in autism. Journal of Autism and Developmental
Disorders, 18, 493–504.
Dawson, M., Mottron, L., & Gernsbacher, M. A. (2008). Learning in
autism. In J. H. Byrne Series (Ed.) & H. Roediger (Vol. Ed.), Learning
and memory: A comprehensive reference: Cognitive psychology of mem-
ory (pp. 759–772). London, United Kingdom: Elsevier.
Dayan, P. (2012). Twenty-five lessons from computational neuromodula-
tion. Neuron, 76, 240–256. doi:10.1016/j.neuron.2012.09.027
Dayan, P., Kakade, S., & Montague, P. R. (2000). Learning and selective
attention. Nature Neuroscience, 3, 1218–1223. doi:10.1038/81504
D’Cruz, A., Ragozzino, M. E., Mosconi, M. W., Shrestha, S., Cook, E. H.,
& Sweeney, J. A. (2013). Reduced behavioral flexibility in autism
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
668 VAN DE CRUYS ET AL.
spectrum disorders. Neuropsychology, 27, 152–160. doi:10.1037/
a0031721
de Gardelle, V., Waszczuk, M., Egner, T., & Summerfield, C. (2012).
Concurrent repetition enhancement and suppression responses in extra-
striate visual cortex. Cerebral Cortex, 23, 2235–2244. doi:10.1093/
cercor/bhs211
de Jonge, M. V., Kemner, C., de Haan, E. H., Coppens, J. E., van den Berg,
T. J. T. P., & van Engeland, H. (2007). Visual information processing in
high-functioning individuals with autism spectrum disorders and their
parents. Neuropsychology, 21, 65–73. doi:10.1037/0894-4105.21.1.65
Delorme, R., Ey, E., Toro, R., Leboyer, M., Gillberg, C., & Bourgeron, T.
(2013). Progress toward treatments for synaptic defects in autism. Na-
ture Medicine, 19, 685–694. doi:10.1038/nm.3193
Del Viva, M. M., Igliozzi, R., Tancredi, R., & Brizzolara, D. (2006).
Spatial and motion integration in children with autism. Vision Research,
46, 1242–1252. doi:10.1016/j.visres.2005.10.018
den Ouden, H. E., Daunizeau, J., Roiser, J., Friston, K. J., & Stephan, K. E.
(2010). Striatal prediction error modulates cortical coupling. Journal of
Neuroscience, 30, 3210–3219. doi:10.1523/JNEUROSCI.4458-09.2010
den Ouden, H. E., Kok, P., & de Lange, F. P. (2012). How prediction errors
shape perception, attention, and motivation. Frontiers in Psychology, 3,
Article 548. doi:10.3389/fpsyg.2012.00548
Dichter, G. S., Felder, J. N., & Bodfish, J. W. (2009). Autism is charac-
terized by dorsal anterior cingulate hyperactivation during social target
detection. Social Cognitive and Affective Neuroscience, 4, 215–226.
doi:10.1093/scan/nsp017
Di Martino, A., Ross, K., Uddin, L. Q., Sklar, A. B., Castellanos, F. X., &
Milham, M. P. (2009). Functional brain correlates of social and nonso-
cial processes in autism spectrum disorders: An activation likelihood
estimation meta-analysis. Biological Psychiatry, 65, 63–74. doi:
10.1016/j.biopsych.2008.09.022
Diuk, C., Tsai, K., Wallis, J., Botvinick, M., & Niv, Y. (2013). Hierarchical
learning induces two simultaneous, but separable, prediction errors in
human basal ganglia. Journal of Neuroscience, 33, 5797–5805. doi:
10.1523/JNEUROSCI.5445-12.2013
Dosenbach, N. U., Visscher, K. M., Palmer, E. D., Miezin, F. M., Wenger,
K. K., Kang, H. C.,...Petersen, S. E. (2006). A core system for the
implementation of task sets. Neuron, 50, 799812. doi:10.1016/j.neuron
.2006.04.031
Dunn, M., Gomes, H., & Gravel, J. (2008). Mismatch negativity in children
with autism and typical development. Journal of Autism and Develop-
mental Disorders, 38, 52–71. doi:10.1007/s10803-007-0359-3
Duzel, E., & Guitart-Masip, M. (2013). Not so uncertain at last: Locus
coeruleus and decision making. Neuron, 79, 9–11. doi:10.1016/j.neuron
.2013.06.023
Ebert, D. H., & Greenberg, M. E. (2013, January 17). Activity-dependent
neuronal signalling and autism spectrum disorder. Nature, 493, 327–
337. doi:10.1038/nature11860
Edelson, S., Edelson, M., Kerr, D., & Grandin, T. (1999). Behavioral and
physiological effects of deep pressure on children with autism: A pilot
study evaluating the efficacy of Grandin’s Hug Machine. American
Journal of Occupational Therapy, 53, 145–152.
Edwards, D. J., Perlman, A., & Reed, P. (2012). Unsupervised categori-
zation in a sample of children with autism spectrum disorders. Research
in Developmental Disabilities, 33, 1264–1269. doi:10.1016/j.ridd.2012
.02.021
Edwards, M. J., Adams, R. A., Brown, H., Pareés, I., & Friston, K. J.
(2012). A Bayesian account of “hysteria”. Brain, 135, 3495–3512.
doi:10.1093/brain/aws129
Egner, T., Monti, J. M., & Summerfield, C. (2010). Expectation and
surprise determine neural population responses in the ventral visual
stream. Journal of Neuroscience, 30, 16601–16608. doi:10.1523/
JNEUROSCI.2770-10.2010
Elison, J. T., Paterson, S. J., Wolff, J. J., Reznick, J. S., Sasson, N. J., Gu,
H.,...Piven, J. (2013). White matter microstructure and atypical visual
orienting in 7-month-olds at risk for autism. American Journal of Psy-
chiatry, 170, 899–908. doi:10.1176/appi.ajp.2012.12091150
Elsabbagh, M., Divan, G., Koh, Y., Kim, Y. S., Kauchali, S., Marcín, C.,
. . . Fombonne, E. (2012). Global prevalence of autism and other
pervasive developmental disorders. Autism Research, 5, 160–179. doi:
10.1002/aur.239
Elsabbagh, M., Volein, A., Holmboe, K., Tucker, L., Csibra, G., Baron-
Cohen, S.,...Johnson, M. H. (2009). Visual orienting in the early
broader autism phenotype: Disengagement and facilitation. Journal of
Child Psychology and Psychiatry, 50, 637–642. doi:10.1111/j.1469-
7610.2008.02051.x
Falck-Ytter, T. (2008). Face inversion effects in autism: A combined
looking time and pupillometric study. Autism Research, 1, 297–306.
doi:10.1002/aur.45
Fan, X., Miles, J. H., Takahashi, N., & Yao, G. (2009). Abnormal transient
pupillary light reflex in individuals with autism spectrum disorders.
Journal of Autism and Developmental Disorders, 39, 1499–1508. doi:
10.1007/s10803-009-0767-7
Feldman, J. (2013). Tuning your priors to the world. Topics in Cognitive
Science, 5, 13–34. doi:10.1111/tops.12003
Ferri, R., Elia, M., Agarwal, N., Lanuzza, B., Musumeci, S. A., & Pennisi,
G. (2003). The mismatch negativity and the P3a components of the
auditory event-related potentials in autistic low-functioning subjects.
Clinical Neurophysiology, 114, 1671–1680. doi:10.1016/S1388-
2457(03)00153-6
Fletcher, P. C., & Frith, C. D. (2009). Perceiving is believing: A Bayesian
approach to explaining the positive symptoms of schizophrenia. Nature
Reviews Neuroscience, 10, 48–58. doi:10.1038/nrn2536
Foss-Feig, J. H., Tadin, D., Schauder, K. B., & Cascio, C. J. (2013). A
substantial and unexpected enhancement of motion perception in autism.
Journal of Neuroscience, 33, 8243–8249. doi:10.1523/JNEUROSCI
.1608-12.2013
Fournier, K. A., Hass, C. J., Naik, S. K., Lodha, N., & Cauraugh, J. H.
(2010). Motor coordination in autism spectrum disorders: A synthesis
and meta-analysis. Journal of Autism and Developmental Disorders, 40,
1227–1240. doi:10.1007/s10803-010-0981-3
Friston, K. J. (2009). The free-energy principle: A rough guide to the
brain? Trends in Cognitive Sciences, 13, 293–301. doi:10.1016/j.tics
.2009.04.005
Friston, K. J. (2010). The free-energy principle: A unified brain theory?
Nature Reviews Neuroscience, 11, 127–138. doi:10.1038/nrn2787
Friston, K. J., Lawson, R., & Frith, C. D. (2013). On hyperpriors and
hypopriors: Comment on Pellicano and Burr. Trends in Cognitive Sci-
ences, 17, 1. doi:10.1016/j.tics.2012.11.003
Friston, K. J., Thornton, C., & Clark, A. (2012). Free-energy minimization
and the dark-room problem. Frontiers in Psychology, 3, Article 130.
doi:10.3389/fpsyg.2012.00130
Frith, U., & Frith, C. D. (2003). Development and neurophysiology of
mentalizing. Philosophical Transactions of the Royal Society of London
B: Biological Sciences, 358, 459473. doi:10.1098/rstb.2002.1218
Frith, U., & Happé, F. (1994). Autism: Beyond “theory of mind.” Cogni-
tion, 50, 115–132.
Froese, T., & Ikegami, T. (2013). The brain is not an isolated “black box,”
nor is its goal to become one. Behavioral and Brain Sciences, 36,
213–214. doi:10.1017/S0140525X12002348
Gastgeb, H., Dundas, E. M., Minshew, N. J., & Strauss, M. S. (2012).
Category formation in autism: Can individuals with autism form cate-
gories and prototypes of dot patterns? Journal of Autism and Develop-
mental Disorders, 42, 1694–1704. doi:10.1007/s10803-011-1411-x
Gastgeb, H., Rump, K. M., Best, C. A., Minshew, N. J., & Strauss, M. S.
(2009). Prototype formation in autism: Can individuals with autism
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
669
PREDICTIVE CODING IN AUTISM
abstract facial prototypes? Autism Research, 2, 279–284. doi:10.1002/
aur.93
Gastgeb, H., Wilkinson, D., Minshew, N., & Strauss, M. (2011). Can
individuals with autism abstract prototypes of natural faces? Journal of
Autism and Developmental Disorders, 41, 1609–1618. doi:10.1007/
s10803-011-1190-4
Geschwind, D. H. (2011). Genetics of autism spectrum disorders. Trends in
Cognitive Sciences, 15, 409416. doi:10.1016/j.tics.2011.07.003
Geurts, H. M., Corbett, B., & Solomon, M. (2009). The paradox of
cognitive flexibility in autism. Trends in Cognitive Sciences, 13, 7482.
doi:10.1016/j.tics.2008.11.006
Gibson, E. (1969). Principles of perceptual learning and development. East
Norwalk, CT: Appleton-Century-Crofts.
Gioia, G. A., Isquith, P. K., Kenworthy, L., & Barton, R. M. (2002).
Profiles of everyday executive function in acquired and developmental
disorders. Child Neuropsychology, 8, 121–137. doi:10.1076/chin.8.2
.121.8727
Gomot, M., Blanc, R., Clery, H., Roux, S., Barthelemy, C., & Bruneau, N.
(2011). Candidate electrophysiological endophenotypes of hyper-
reactivity to change in autism. Journal of Autism and Developmental
Disorders, 41, 705–714. doi:10.1007/s10803-010-1091-y
Gomot, M., Giard, M. H., Adrien, J. L., Barthelemy, C., & Bruneau, N.
(2002). Hypersensitivity to acoustic change in children with autism:
Electrophysiological evidence of left frontal cortex dysfunctioning. Psy-
chophysiology, 39, 577–584. doi:10.1111/1469-8986.3950577
Gomot, M., & Wicker, B. (2012). A challenging, unpredictable world for
people with autism spectrum disorder. International Journal of Psycho-
physiology, 83, 240–247. doi:10.1016/j.ijpsycho.2011.09.017
Gonzalez, C., Martin, J. M., Minshew, N. J., & Behrmann, M. (2013).
Practice makes improvement: How adults with autism out-perform oth-
ers in a naturalistic visual search task. Journal of Autism and Develop-
mental Disorders, 43, 2259–2268. doi:10.1007/s10803-013-1772-4
Goris, R. L. T., Wagemans, J., & Wichmann, F. A. (2008). Modelling
contrast discrimination data suggest both the pedestal effect and sto-
chastic resonance to be caused by the same mechanism. Journal of
Vision, 8(15), Article 17. doi:10.1167/8.15.17
Gottlieb, J. (2012). Attention, learning, and the value of information.
Neuron, 76, 281–295. doi:10.1016/j.neuron.2012.09.034
Grandin, T. (1992). Calming effects of deep touch pressure in patients with
autistic disorder, college students, and animals. Journal of Child and
Adolescent Psychopharmacology, 2, 63–72. doi:10.1089/cap.1992.2.63
Grandin, T. (1995). Thinking in pictures and other reports of my life with
autism. New York, NY: Vintage Books.
Gregory, R. L. (1980). Perceptions as hypotheses. Philosophical Transac-
tions of the Royal Society B: Biological Sciences, 290, 181–197. doi:
10.1098/rstb.1980.0090
Haesen, B., Boets, B., & Wagemans, J. (2011). A review of behavioural
and electrophysiological studies on auditory processing and speech
perception in autism spectrum disorders. Research in Autism Spectrum
Disorders, 5, 701–714. doi:10.1016/j.rasd.2010.11.006
Ham, T., Leff, A., de Boissezon, X., Joffe, A., & Sharp, D. J. (2013).
Cognitive control and the salience network: An investigation of error
processing and effective connectivity. Journal of Neuroscience, 33,
7091–7098. doi:10.1523/JNEUROSCI.4692-12.2013
Hamilton, A. F. C., Brindley, R. M., & Frith, U. (2007). Imitation and
action understanding in autistic spectrum disorders: How valid is the
hypothesis of a deficit in the mirror neuron system? Neuropsychologia,
45, 1859–1868. doi:10.1016/j.neuropsychologia.2006.11.022
Happé, F., & Booth, R. (2008). The power of the positive: Revisiting weak
coherence in autism spectrum disorders. Quarterly Journal of Experi-
mental Psychology, 61, 5063. doi:10.1080/17470210701508731
Happé, F., & Frith, U. (2006). The weak coherence account: Detail-focused
cognitive style in autism spectrum disorders. Journal of Autism and
Developmental Disorders, 36, 5–25. doi:10.1007/s10803-005-0039-0
Happé, F., & Ronald, A. (2008). The “fractionable autism triad”: A review
of evidence from behavioural, genetic, cognitive and neural research.
Neuropsychology Review, 18, 287–304. doi:10.1007/s11065-008-9076-8
Harms, M., Martin, A., & Wallace, G. L. (2010). Facial emotion recogni-
tion in autism spectrum disorders: A review of behavioral and neuro-
imaging studies. Neuropsychology Review, 20, 290–322. doi:10.1007/
s11065-010-9138-6
Hawkins, J., & Blakeslee, S. (2004). On intelligence. New York, NY:
Times Books.
Heaton, P. (2003). Pitch memory, labelling and disembedding in autism.
Journal of Child Psychology and Psychiatry, 44, 543–551. doi:10.1111/
1469-7610.00143
Heaton, P., Williams, K., Cummins, O., & Happé, F. (2008). Autism and
pitch processing splinter skills: A group and subgroup analysis. Autism,
12, 203–219. doi:10.1177/1362361307085270
Herry, C., Bach, D. R., Esposito, F., Di Salle, F., Perrig, W. J., Scheffler,
K.,...Seifritz, E. (2007). Processing of temporal unpredictability in
human and animal amygdala. Journal of Neuroscience, 27, 5958–5966.
doi:10.1523/JNEUROSCI.5218-06.2007
Hill, A. L. (1978). Savants: Mentally retarded individuals with special
skills. International Review of Research in Mental Retardation, 9, 277–
298.
Hill, E. L. (2004). Executive dysfunction in autism. Trends in Cognitive
Sciences, 8, 26–32. doi:10.1016/j.dr.2004.01.001
Hillyard, S. A., Vogel, E. K., & Luck, S. J. (1998). Sensory gain control
(amplification) as a mechanism of selective attention: Electrophysiolog-
ical and neuroimaging evidence. Philosophical Transactions of the
Royal Society B: Biological Sciences, 353, 1257–1270. doi:10.1098/rstb
.1998.0281
Hirsh, J. B., Mar, R. A., & Peterson, J. B. (2012). Psychological entropy:
A framework for understanding uncertainty-related anxiety. Psycholog-
ical Review, 119, 304–320. doi:10.1037/a0026767
Hohwy, J., & Palmer, C. (2014). Social cognition as causal inference:
Implications for common knowledge and autism. In M. Gallotti & J.
Michael (Eds.), Perspectives on social ontology and social cognition
(pp. 167–189). New York, NY: Springer.
Hohwy, J., Roepstorff, A., & Friston, K. (2008). Predictive coding explains
binocular rivalry: An epistemological review. Cognition, 108, 687–701.
doi:10.1016/j.cognition.2008.05.010
Honey, R. C., Watt, A., & Good, M. (1998). Hippocampal lesions disrupt
an associative mismatch process. Journal of Neuroscience, 18, 2226
2230.
Howlin, P., Goode, S., Hutton, J., & Rutter, M. (2009). Savant skills in
autism: Psychometric approaches and parental reports. Philosophical
Transactions of the Royal Society B: Biological Sciences, 364, 1359
1367. doi:10.1098/rstb.2008.0328
Hughes, C. (1996). Brief report: Planning problems in autism at the level
of motor control. Journal of Autism and Developmental Disorders, 26,
99–107. doi:10.1007/BF02276237
Huron, D. (2006). Sweet anticipation: Music and the psychology of expec-
tation. Cambridge, MA: MIT Press.
Hutsler, J. J., & Zhang, H. (2010). Increased dendritic spine densities on
cortical projection neurons in autism spectrum disorders. Brain Re-
search, 1309, 83–94. doi:10.1016/j.brainres.2009.09.120
Iarocci, G., & McDonald, J. (2006). Sensory integration and the perceptual
experience of persons with autism. Journal of Autism and Developmen-
tal Disorders, 36, 77–90. doi:10.1007/s10803-005-0044-3
Ide, J. S., Shenoy, P., Yu, A. J., & Li, C. R. (2013). Bayesian prediction and
evaluation in the anterior cingulate cortex. Journal of Neuroscience, 33,
2039–2047. doi:10.1523/JNEUROSCI.2201-12.2013
Inoue, W., Baimoukhametova, D. V., Füzesi, T., Cusulin, J. I. W., Ko-
blinger, K., Whelan, P. J.,...Bains, J. S. (2013). Noradrenaline is a
stress-associated metaplastic signal at GABA synapses. Nature Neuro-
science, 16, 605–612. doi:10.1038/nn.3373
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
670 VAN DE CRUYS ET AL.
Jarvinen-Pasley, A., Peppe, S., King-Smith, G., & Heaton, P. (2008). The
relationship between form and function level receptive prosodic abilities
in autism. Journal of Autism and Developmental Disorders, 38, 1328
1340. doi:10.1007/s10803-007-0520-z
Joffily, M., & Coricelli, G. (2013). Emotional valence and the free-energy
principle. PLoS Computational Biology, 9(6), e1003094. doi:10.1371/
journal.pcbi.1003094
Jones, C. R., Happé, F., Baird, G., Simonoff, E., Marsden, A. J., Tregay, J.,
. . . Charman, T. (2009). Auditory discrimination and auditory sensory
behaviours in autism spectrum disorders. Neuropsychologia, 47, 2850
2858. doi:10.1016/j.neuropsychologia.2009.06.015
Joseph, R. M., Keehn, B., Connolly, C., Wolfe, J. M., & Horowitz, T. S.
(2009). Why is visual search superior in autism spectrum disorder?
Developmental Science, 12, 1083–1096. doi:10.1111/j.1467-7687.2009
.00855.x
Kanner, L. (1943). Autistic disturbances of affective contact. Nervous
Child, 2, 217–250.
Karlsson, M. P., Tervo, D. G. R., & Karpova, A. Y. (2012, October 5).
Network resets in medial prefrontal cortex mark the onset of behavioral
uncertainty. Science, 338, 135–139. doi:10.1126/science.1226518
Keehn, B., Brenner, L., Ramos, A., Lincoln, A., Marshall, S., & Muller, R.
(2009). Brief report: Eye-movement patterns during an embedded fig-
ures test in children with ASD. Journal of Autism and Developmental
Disorders, 39, 383–387. doi:10.1007/s10803-008-0608-0
Keehn, B., Müller, R., & Townsend, J. (2013). Atypical attentional net-
works and the emergence of autism. Neuroscience & Biobehavioral
Reviews, 37, 164–183. doi:10.1016/j.neubiorev.2012.11.014
Kéïta, L., Mottron, L., & Bertone, A. (2010). Far visual acuity is unre-
markable in autism: Do we need to focus on crowding? Autism Re-
search, 3, 333–341. doi:10.1002/aur.164
Kemner, C., Verbaten, M., Cuperus, J., Camfferman, G., & Van Engeland,
H. (1995). Auditory event-related brain potentials in autistic children
and three different control groups. Biological Psychiatry, 38, 150–165.
doi:10.1016/0006-3223(94)00247-Z
Kenworthy, L., Yerys, B. E., Anthony, L. G., & Wallace, G. L. (2008).
Understanding executive control in autism spectrum disorders in the lab
and in the real world. Neuropsychology Review, 18, 320–338. doi:
10.1007/s11065-008-9077-7
Kern, J. K., Miller, V. S., Cauller, L., Kendall, R., Mehta, J., & Dodd, M.
(2001). Effectiveness of N,N-dimethylglycine in autism and pervasive
developmental disorder. Journal of Child Neurology, 16, 169–173.
doi:10.1177/088307380101600303
Kern, J. K., Trivedi, M., Garver, C., Grannemann, B., Andrews, A., Savla,
J.,...Schroeder, J. (2006). The pattern of sensory processing abnor-
malities in autism. Autism, 10, 480494. doi:10.1177/1362
361306066564
Khalfa, S., Bruneau, N., Rogé, B., Georgieff, N., Veuillet, E., Adrien, J.,
. . . Collet, L. (2004). Increased perception of loudness in autism.
Hearing Research, 198, 87–92. doi:10.1016/j.heares.2004.07.006
Kilner, J. M., Friston, K. J., & Frith, C. D. (2007). Predictive coding: An
account of the mirror neuron system. Cognitive Processing, 8, 159–166.
doi:10.1007/s10339-007-0170-2
Kim, J. A., Szatmari, P., Bryson, S. E., Streiner, D. L., & Wilson, F. J.
(2000). The prevalence of anxiety and mood problems among children
with autism and Asperger syndrome. Autism, 4, 117–132. doi:10.1177/
1362361300004002002
Klin, A. (1991). Young autistic childrens’ listening preferences in regard to
speech: A possible characterization of the symptom of social with-
drawal. Journal of Autism and Developmental Disorders, 21, 2942.
doi:10.1007/BF02206995
Klin, A. (1992). Listening preferences in regard to speech in four children
with developmental disabilities. Journal of Child Psychology and Psy-
chiatry, 33, 763–769. doi:10.1111/j.1469-7610.1992.tb00911.x
Klin, A., Jones, W., Schultz, R., Volkmar, F., & Cohen, D. (2002). Visual
fixation patterns during viewing of naturalistic social situations as pre-
dictors of social competence in individuals with autism. Archives of
General Psychiatry, 59, 809816.
Klin, A., Lin, D., Gorrindo, P., Ramsay, G., & Jones, W. (2009, May 14).
Two-year-olds with autism orient to non-social contingencies rather than
biological motion. Nature, 459, 257–261. doi:10.1038/nature07868
Klinger, L., & Dawson, G. (2001). Prototype formation in autism. Devel-
opment and Psychopathology, 13, 111–124. doi:10.1017/
S0954579401001080
Koegel, R. L., & Mentis, M. (1985). Motivation in childhood autism: Can
they or won’t they? Journal of Child Psychology and Psychiatry, 26,
185–191. doi:10.1111/j.1469-7610.1985.tb02259.x
Kohler, E., Keysers, C., Umilta
`, M. A., Fogassi, L., Gallese, V., &
Rizzolatti, G. (2002, August 2002). Hearing sounds, understanding
actions: Action representation in mirror neurons. Science, 297, 846
848. doi:10.1126/science.1070311
Kootz, J. P., & Cohen, D. J. (1981). Modulation of sensory intake in
autistic children: Cardiovascular and behavioral indices. Journal of the
American Academy of Child Psychiatry, 20, 692–701. doi:10.1097/
00004583-198102000-00002
Koster-Hale, J., & Saxe, R. (2013). Theory of mind: A neural prediction
problem. Neuron, 79, 836848. doi:10.1016/j.neuron.2013.08.020
Kujala, T., Lepisto, T., Nieminen-von Wendt, T., Naatanen, P., & Naa-
tanen, R. (2005). Neurophysiological evidence for cortical discrimina-
tion impairment of prosody in Asperger syndrome. Neuroscience Let-
ters, 383, 260–265. doi:10.1016/j.neulet.2005.04.048
Lam, K. S., Aman, M. G., & Arnold, L. E. (2006). Neurochemical corre-
lates of autistic disorder: A review of the literature. Research in Devel-
opmental Disabilities, 27, 254–289. doi:10.1016/j.ridd.2005.03.003
Legare, C. H. (2012). Exploring explanation: Explaining inconsistent ev-
idence informs exploratory, hypothesis-testing behavior in young chil-
dren. Child Development, 83, 173–185. doi:10.1111/j.1467-8624.2011
.01691.x
Lepisto, T., Kujala, T., Vanhala, R., Alku, P., Huotilainen, M., & Naa-
tanen, R. (2005). The discrimination of and orienting to speech and
non-speech sounds in children with autism. Brain Research, 1066,
147–157. doi:10.1016/j.brainres.2005.10.052
Lepisto, T., Silokallio, S., Nieminen-von Wendt, T., Alku, P., Naatanen,
R., & Kujala, T. (2006). Auditory perception and attention as reflected
by the brain event-related potentials in children with Asperger syn-
drome. Clinical Neurophysiology, 117, 2161–2171. doi:10.1016/j.clinph
.2006.06.709
Limongi, R., Sutherland, S. C., Zhu, J., Young, M. E., & Habib, R. (2013).
Temporal prediction errors modulate cingulate–insular coupling. Neu-
roImage, 71, 147–157. doi:10.1016/j.neuroimage.2012.12.078
Lopez, B., Donnelly, N., Hadwin, J. A., & Leekam, S. R. (2004). Face
processing in high-functioning adolescents with autism: Evidence for
weak central coherence. Visual Cognition, 11, 673–688. doi:10.1080/
13506280344000437
Lopez, B. R., Lincoln, A. J., Ozonoff, S., & Lai, Z. (2005). Examining the
relationship between executive functions and restricted, repetitive symp-
toms of autistic disorder. Journal of Autism and Developmental Disor-
ders, 35, 445–460. doi:10.1007/s10803-005-5035-x
Loth, E., Happé, F., & Gomez, J. C. (2010). Variety is not the spice of life
for people with autism spectrum disorders: Frequency ratings of central,
variable and inappropriate aspects of common real-life events. Journal
of Autism and Developmental Disorders, 40, 730–742. doi:10.1007/
s10803-009-0929-7
Lovaas, O. I. (1987). Behavioral treatment and normal educational and
intellectual functioning in young autistic children. Journal of Consulting
and Clinical Psychology, 55, 3–9. doi:10.1037/0022-006X.55.1.3
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
671
PREDICTIVE CODING IN AUTISM
Lovaas, O. I., Koegel, R. L., & Schreibman, L. (1979). Stimulus overse-
lectivity in autism: A review of research. Psychological Bulletin, 86,
1236–1254. doi:10.1037/0033-2909.86.6.1236
Loveland, K., Tunali-Kotoski, B., Chen, R., Brelsford, K., Ortegon, J., &
Pearson, D. (1995). Intermodal perception of affect in persons with
autism or Down syndrome. Development and Psychopathology, 7, 409
418.
Mackinlay, R., Charman, T., & Karmiloff-Smith, A. (2006). High func-
tioning children with autism spectrum disorder: A novel test of multi-