RESEARCH - Ricerca sull'intelligenza artificiale specializzata al MIT

RESEARCH

Bridging physiological and perceptual views of
autism by means of sampling-based
Bayesian inference

Rodrigo Echeveste1

, Enzo Ferrante1*

, Diego H. Milone1*

, and Inés Samengo2*

1Research Institute for Signals, Sistemi, and Computational Intelligence sinc(io) (FICH-UNL/CONICET), Santa Fe, Argentina
2Medical Physics Department and Balseiro Institute (CNEA-UNCUYO/CONICET), Bariloche, Argentina
*These authors contributed equally to this work.

a n o p e n a c c e s s

j o u r n a l

Keywords: Autism, Neural circuits, Inhibitory dysfunction, Hypopriors, Sampling-based inference

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

ABSTRACT

Theories for autism spectrum disorder (ASD) have been formulated at different levels, ranging
from physiological observations to perceptual and behavioral descriptions. Understanding the
physiological underpinnings of perceptual traits in ASD remains a significant challenge in the
field. Here we show how a recurrent neural circuit model that was optimized to perform
sampling-based inference and displays characteristic features of cortical dynamics can help
bridge this gap. The model was able to establish a mechanistic link between two descriptive levels
for ASD: a physiological level, in terms of inhibitory dysfunction, neural variability, E
oscillations, and a perceptual level, in terms of hypopriors in Bayesian computations. We took
two parallel paths—inducing hypopriors in the probabilistic model, and an inhibitory dysfunction
in the network model—which lead to consistent results in terms of the represented posteriors,
providing support for the view that both descriptions might constitute two sides of the same coin.

AUTHOR SUMMARY

Two different views of autism, one regarding altered probabilistic computations, and one
regarding inhibitory dysfunction, are brought together by means of a recurrent neural network
model trained to perform sampling-based inference in a visual setting. Inoltre, the model
captures a variety of experimental observations regarding differences in neural variability and
oscillations in subjects with autism. By linking neural connectivity, dynamics, and function,
this work contributes to the understanding of the physiological underpinnings of perceptual
traits in autism spectrum disorder.

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

INTRODUCTION

Autism spectrum disorder (ASD) refers to a complex neurodevelopmental condition involving
persistent challenges in social interaction and communicative skills, and restricted/repetitive
behaviors (American Psychiatric Association, 2013). While some recent studies suggest that
ASD could be detected during the first year of life in some children, early signs seem to be
nonspecific, with group differences more robustly found after a child’s first birthday (Vedere
Ozonoff, Heung, Byrd, Hansen, & Hertz-Picciotto, 2008, for a review).

Citation: Echeveste, R., Ferrante, E.,
Milone, D. H., & Samengo, IO. (2022).
Bridging physiological and perceptual
views of autism by means of sampling-
based Bayesian inference. Network
Neuroscience, 6(1), 196–212. https://doi
.org/10.1162/netn_a_00219

DOI:
https://doi.org/10.1162/netn_a_00219

Supporting Information:
https://doi.org/10.1162/netn_a_00219
https:// bitbucket.org/ RSE_1987/ssn
_inference_numerical_experiments
https:// bitbucket.org/ RSE_1987
/inhibitory_dysfunction

Received: 27 agosto 2021
Accepted: 1 Dicembre 2021

Competing Interests: The authors have
declared that no competing interests
exist.

Corresponding Author:
Rodrigo Echeveste
recheveste@sinc.unl.edu.ar

Handling Editor:
Olusola Ajilore

Copyright: © 2021
Istituto di Tecnologia del Massachussetts
Pubblicato sotto Creative Commons
Attribuzione 4.0 Internazionale
(CC BY 4.0) licenza

The MIT Press

Bridging physiological and perceptual views of autism

Almost two decades ago, John Rubenstein and Michael Merzenich suggested that many of
the symptoms related to ASD might reflect an abnormal ratio between excitation and inhibi-
tion leading to hyperexcitability of cortical circuits in ASD subjects (Rubenstein & Merzenich,
2003). Since then, a variety of studies have linked reduced inhibitory signaling in the brain
with ASD symptoms, either observing how behavior typically associated with ASD emerges
in animals when inhibitory pathways are altered, or measuring gamma-aminobutyric acid
(GABA) concentration or GABA receptors in several brain regions (see Cellot & Cherubini,
2014, for a detailed review). Further support for this view comes from the fact that ASD
patients suffer from epilepsy with a prevalence up to 25 times that of the neurotypical popu-
lation (Bolton et al., 2011).

Establishing a direct link between ASD and impaired inhibition in specific circuits in
humans has not been easy. Infatti, two recent in vivo studies in humans have shown puz-
zling results (Horder et al., 2018; Robertson, Ratai, & Kanwisher, 2016). In these studies
inhibition was assessed both behaviorally (in visual tasks where inhibition is widely
believed to play a key role in neurotypical behavior) and by measuring either GABA con-
centration (Robertson et al., 2016) or number of GABA receptors (Horder et al., 2018) in the
brains of ASD and control subjects. È interessante notare, while ASD subjects showed a marked def-
icit in binocular rivalry, characteristic of a disruption in inhibitory signaling, GABA concen-
trations in the visual cortex were normal (Robertson et al., 2016). Tuttavia, while GABA
concentration was predictive of rivalry dynamics in controls, the same was not true within
the ASD population, evidencing a disruption of inhibitory action. Allo stesso modo, while ASD sub-
jects show an altered performance in the paradoxical motion perception task (a proxy mea-
sure of GABA signaling), GABA receptor availability in the brain of those participants
showed no significant difference from controls (Horder et al., 2018). Both studies suggest
an impairment in inhibitory signaling that cannot be explained by coarse differences in
GABA concentration or receptor availability at the level of brain areas, and that might affect
specific circuits instead. To complicate matters further, there is evidence for not only inhib-
itory but also excitatory dysfunction in ASD, and it has been hypothesized that homeostatic
principles might be the reason behind this seemingly contradictory result (Nelson & Valakh,
2015). The idea is that if, for instance, inhibition is reduced, excitatory synapses might be
then adjusted to try to compensate for the overall change in neural activity that reduction
would ensue. Computational modeling of local cortical circuits expressed in terms of exci-
tation and inhibition might therefore provide a fruitful avenue of research to guide future
esperimenti.

From the point of view of perception in ASD, a variety of theories have been put forward
over the past two decades. Highly influential descriptive theories include the weak central
coherence theory (Happé & Frith, 2006) and the enhanced perceptual functioning theory
(Mottron, Dawson, Soulieres, Hubert, & Burack, 2006). Here we will focus on computational
accounts of perception in ASD, and in particular on a Bayesian view of perception (Palmer,
Lawson, & Hohwy, 2017). We will later also make connections to another influential compu-
tational theory formulated in terms of predictive coding (Van Boxtel & Lu, 2013; Van de Cruys
et al., 2014).

Within the Bayesian framework, inference about the external world proceeds by multipli-
catively combining preexistent knowledge (expressed in terms of a prior probability distribu-
zione) and current sensory evidence (represented in terms of a likelihood function), to form a
posterior distribution that encapsulates our belief about the state of the world after having
observed a given stimulus (Knill & Richards, 1996). Rather than expressing that belief as a
single point estimate of what is most probable, the posterior distribution provides a richer

197

GABA:
Main inhibitory neurotransmitter.

Prior:
Probability distribution encapsulating
an observer’s knowledge about the
latent variables before observing the
stimulus.

Likelihood function:
Function describing the conditional
probability of an observation for each
state of the latent variables.

Posterior:
Conditional probability over the
latent variables after observing a
given stimulus.

Network Neuroscience

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Bridging physiological and perceptual views of autism

Hypoprior:
A chronically attenuated prior,
whose uncertainty is higher than
implied by the statistics of stimuli.

Transient overshoot:
Excursion in neural responses that
exceeds mean responses over a brief
period of time after the onset of the
stimulus.

Gamma oscillations:
Rhythmic patterns of activity with a
frequency between 20 E 80 Hz.

Divisive normalization:
Process by which the responses of
single neurons are divisively
modulated by the responses of other
neurons.

description, naturally incorporating the associated uncertainty that remains after the observa-
zione. A growing body of evidence indicates that, at least in some settings, the brain is able to
operate with probability distributions in this way to perform approximate Bayesian inference
(see Fiser, Berkes, Orbán, & Lengyel, 2010, for a review). In recent years it has been proposed
that in ASD subjects these forms of Bayesian computations are carried out abnormally, Sopra-
weighting sensory evidence with respect to prior information (Palmer et al., 2017; Pellicano &
Burr, 2012). Concretely, the authors in Pellicano and Burr (2012) proposed that this is a con-
sequence of chronically attenuated priors (termed hypopriors), characterized by broader dis-
tributions (cioè., higher uncertainty).

The related theoretical framework of predictive coding proposes that the cortex is organized
following a circuit motif where feedback connections from higher to lower order sensory areas
signal predictions of lower level responses, while feedforward connections signal errors
between predictions and actually observed lower level responses (Rao & Ballard, 1999). Pro-
ponents of predictive coding theories have rightfully pointed out that Bayesian theories by
themselves (without specifying a concrete implementation) do not offer a mechanistic expla-
nation for ASD perception (Van Boxtel & Lu, 2013), which is key to understand how physio-
logical observations may be linked to perceptual and behavioral traits in ASD subjects. As has
been observed by Aitchison and Lengyel (2017), Bayesian inference and predictive coding are
not necessarily mutually exclusive: Predictive coding can be seen as a computational motif
that can implement several computational goals (one of which is Bayesian inference), while
Bayesian inference can be seen as a computational objective that can have several implemen-
tations (one of which is predictive coding). Inoltre, as noted in the aforementioned review,
telling apart the use of a Bayesian predictive coding scheme from a direct variable code in an
empirical setting is no trivial matter. Strong transient overshoots at stimulus onset, for instance,
which are a typical signature of predictive coding, can also emerge in direct variable coding
schemes (Aitchison & Lengyel, 2016; Echeveste, Aitchison, Hennequin, & Lengyel, 2020).
Infatti, while weighting predictive errors more strongly by increasing synaptic gains in the
motif could explain sensory hypersensitivity in ASD subjects (Palmer et al., 2017), a competing
explanation can be provided within a direct variable coding scheme, as we show in the pres-
ent study. We note however that while predictive coding schemes can incorporate gamma
oscillations (Bastos et al., 2012), it is not clear how they would account for the contrast-
dependent frequency modulation of these oscillations (Roberts et al., 2013), or the stimulus-
dependent modulations of neural variability (Churchland et al., 2010; Orbán, Berkes, Fiser, &
Lengyel, 2016).

A popular implementation choice for probabilistic inference is that of probabilistic popu-
lation codes (PPCs; Mamma, Beck, Latham, & Pouget, 2006), where the posterior distribution is
encoded in the average rates of a population of neurons. This framework has been used in
the past to link inhibitory deficits and Bayesian computations in an artificial neural network
model consisting of two feedforward layers followed by a stage of divisive normalization
(Rosenberg, Patterson, & Angelaki, 2015). In this work, a probabilistic version of the model
was constructed to capture the “oblique effect.” This term describes the fact that neurotypical
subjects tend to be more sensitive to cardinal than to oblique orientations in a visual orienta-
tion discrimination task (Westheimer & Beard, 1998). Infatti, a modulation of the divisive
normalization factor in this model was shown to account for the observed reduction of the
oblique effect in ASD subjects (Dickinson, Jones, & Milne, 2014). The standard PPC frame-
work requires constant Fano factors (no variability modulation; Ma et al., 2006), and further-
more feedforward network implementations can only capture mean rate responses, but fail to
account for the dynamical properties of neural responses that arise from recurrent connectivity.

Network Neuroscience

198

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Bridging physiological and perceptual views of autism

It is hence unclear in this framework how altered neural variability observed in the ASD pop-
ulation (Haigh, Heeger, Dinstein, Minshew, & Behrmann, 2015; Milne, 2011) and gamma
oscillations (van Diessen, Senders, Jansen, Boersma, & Bruining, 2015) would relate to prob-
abilistic computations in these subjects.

Sampling-based theories for probabilistic inference offer an alternative mechanistic imple-
mentation for Bayesian inference. Within this framework, neural circuits represent posterior
distributions by drawing samples over time from those distributions (Berkes, Orbán, Lengyel,
& Fiser, 2011; Haefner, Berkes, & Fiser, 2016). È interessante notare, sampling-based models for prob-
abilistic inference have recently begun to establish direct links between cortical dynamics and
perception (Echeveste et al., 2020). A neural circuit model of a cortical hypercolumn respecting
Dale’s principle and performing fast sampling-based inference in a visual task displayed a suite
of features that are typically observed in cortical recordings across species and experimental
conditions. The network showed highly variable responses with strong inhibition-dominated
transients at stimulus onset, and stimulus-dependent gamma oscillations, as observed in the
cortex (Haider, Häusser, & Carandini, 2013; Ray & Maunsell, 2010; Roberts et al., 2013).
The model further evidenced stimulus-dependent variability modulations consistent with
experimental findings (Roberts et al., 2013). Divisive normalization of mean responses
(Carandini & Heeger, 2012) was also shown to emerge in this network as a result of its recurrent
dynamics. This is interesting since divisive normalization was precisely the starting point for the
probabilistic model in Rosenberg et al. (2015), and in previous work linking uncertainty and
neural variability via gain modulation (Hénaff, Boundy-Singer, Meding, Ziemba, & Goris,
2020). The computational and dynamical properties of the network make it a viable candidate
to test the link between Bayesian computations and several physiological features observed in
ASD such as inhibitory dysfunction, as well as differences in neural variability and oscillations.

In what follows we will first set the basis for this work by recapitulating some of the key
findings of Echeveste et al. (2020), relating probabilistic inference, and dynamics in a network
model that we will take to describe healthy control subjects. We will then make use of the
connection between perception and physiology established by this model and take two par-
allel routes to explore two different theories for autism: a perceptual theory expressed in terms
of hypopriors, and a physiological theory concerning impaired inhibition. The first path will
involve modifying the probabilistic model under which perception takes place, and more con-
cretely its prior, and observing the consequences of that choice in terms of the observer’s pos-
teriors. The second path will involve inducing an inhibitory deficit in the neural network
whose job is to sample from the corresponding posteriors, and analyzing the effect of that
modification in the posteriors represented by the network. We will then compare the results
of both approaches to determine to what extent these two seemingly unrelated theories are
compatible. Finalmente, we show that the induced inhibitory deficit in the network model pro-
duces changes in the variability and dynamics of the network. We will evaluate these changes
in the context of empirical observations in ASD subjects and other theoretical accounts for
ASD. These include an increase in neural variability, as well as an increase in the power
and frequency of gamma oscillations. The network also becomes hypersensitive to intense
stimuli, displaying stronger transient responses at stimulus onset.

RESULTS

Bayesian Inference of Visual Features Implemented by a Recurrent E-I Neural Circuit

The starting point for perceptual inference within the Bayesian framework is a probabilistic
model that describes one’s assumptions about how observed stimuli relate to variables of

Network Neuroscience

199

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Bridging physiological and perceptual views of autism

interest in the outside world. This forward model is usually referred to as a generative model,
and the role of an ideal Bayesian observer is to invert this probabilistic relationship to obtain
posterior distributions over those variables of interest given the observed stimulus. The gener-
ative model employed here is a Gaussian scale-mixture model (GSM; see Figure 1A and
Methods), which has been shown to capture the statistics of natural images at the level of small
image patches (Wainwright & Simoncelli, 2000). Importantly, inference under this model had
already been shown to explain features of behavior and stationary response distributions in

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Sketches of the generative model, and a neural circuit implementing sampling-based
Figura 1.
probabilistic inference under that model. (UN) The Gaussian scale mixture (GSM) generative model.
Under this model, each image patch is built as a linear combination of local features (projective
fields), whose intensities are drawn from a multivariate Gaussian distribution. This linear combina-
tion is then further scaled by a global contrast level and subject to noise. The features were in this
case a set of localized oriented Gabor filters that differed only in their orientations and were uni-
formly spread between −90° and 90°. The image serving as stimulus in the figure is for illustration
only. Photo credit: Santa Fe Bridge by Enzo Ferrante (https://eferrante.github.io/). (B) A 2D projection
of the posterior distribution for a given a visual stimulus as computed by the Bayesian ideal observer
under the GSM. (C) The recurrent E–I neural network receives an image patch as an input, che è
filtered by feedforward receptive fields matching the projective fields of GSM in panel A. Each latent
variable in the GSM is represented by the activity of one E cell in the network. (D) A 2D projection of
the neural responses of E cells corresponding the same 2 latent variables shown in panel B. Over
time, the network samples from the posterior distribution corresponding to the stimulus it receives.

200

Latent variable:
A variable of interest to which an
observer has no direct access and
hence needs to infer from an
observation of other related
variables.

Network Neuroscience

Bridging physiological and perceptual views of autism

neural data in visual perception (Coen-Cagli, Kohn, & Schwartz, 2015; Orbán et al., 2016;
Schwartz, Sejnowski, & Dayan, 2009). Under this version of the GSM, natural image patches
are constructed as linear combinations of Gabor filters of different orientations, which are then
scaled by a global contrast variable. The goal of the inference process was to estimate the
probability distribution of the intensity with which each Gabor filter (each orientation) partic-
ipated in the observed image. In turn, in order to model cortical neural dynamics, a common
recurrent neural network model is employed: the stabilized supralinear network (SSN; Vedere
Figure 1B and Methods; Ahmadian, Rubin, & Mugnaio, 2013; Hennequin, Ahmadian, Rubin,
Lengyel, & Mugnaio, 2018). Neurons in the network were arranged around a ring, according
to their preferred orientation, under the approximation of the visual inference problem being
rotationally symmetric (though see Discussion). Inoltre, neurons in the network respected
Dale’s principle, with two separate populations for excitatory (E) and inhibitory (IO) cells. IL
SSN thus formulated was then optimized using current machine learning methods to approx-
imate a Bayesian ideal observer under the GSM: When the network receives an image patch
as its input, it produces samples over time with its neural activity so as to represent the corre-
sponding posterior distribution (Figure 1C–D). Examples of the image patches used to train the
rete, as well as sample neural trajectories, are presented in Figures 2A and 2B,

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Inference under the GSM and responses in the original network, here representing healthy
Figura 2.
neurotypical subjects. Replotted from Echeveste et al. (2020). In all panels, shades of green correspond
to the ideal observer, while red corresponds to network responses, as in Figure 1. Line colors in panel B
indicate different contrast levels, which are the same as stimulus frames in panel A, indicating to which
stimulus responses correspond. (UN) Stimuli (shade of frame color indicates contrast level, split green,
blue, and red indicate that the same stimuli were used as input to the ideal observer and to both neural
networks). (B) Covariance ellipses (2 standard deviations) of the ideal observer’s posterior distributions
(green) and of the networks’ corresponding response distributions (red). Red trajectories show sample
500-ms sequences of activities in the networks. As in the sketch of Figure 1, 2D projections corre-
sponding to two representative latent variables / excitatory cells are shown. These two correspond
to projective fields / receptive fields at preferred orientations 42° and 16°. (C) Mean (top) and standard
deviation (bottom) of latent variable intensities ordered by each latent’s orientation, for each stimulus
in the training set. Left: from the ideal observer’s posterior distribution (green). Right: E cell membrane
potentials uE from the networks’ stationary distributions (red). Response moments in panel C were
estimated from n = 20,000 independent samples (taken 200 ms apart).

Network Neuroscience

201

Bridging physiological and perceptual views of autism

rispettivamente. After training, posterior distributions sampled by network responses match those
prescribed by the ideal observer (see Figure 2C, cf. green and red). Once trained, the SSN
model thus establishes a mechanistic link between neural dynamics in terms of an E-I circuit
and perception formulated as sampling-based probabilistic inference. In what follows we
exploit this link to take two complementary paths: inducing simple perturbations to the
GSM to induce hypopriors, and to the SSN to induce an inhibitory dysfunction.

Perturbing the Generative Model: The Effect of Hypopriors

To illustrate and generate intuitions on the effect of hypopriors, we begin by employing a sim-
plified one-dimensional toy example (Methods). Let us assume the “true” prior, correctly
describing the statistics of the world concerning a particular inference process, is a zero-mean
Gaussian. Let us further assume for this toy example that the likelihood is also a Gaussian func-
tion whose precision is modulated by a contrast variable that expresses the degree of reliability
of the sensory stimulus. If we vary the stimulus contrast we can compute a posterior distribution
for each stimulus under this true prior (Figure 3A–B, in green). If however we were to employ a
hypoprior, questo è, a prior with a higher variance, we would obtain posterior distributions that
overweight sensory evidence, in the sense that they more closely resemble the likelihood func-
zione (in both mean and variance) than they should. This in turn results in a higher posterior mean
and in higher uncertainty about the estimate (Figure 3B, cf. green and blue lines).

Let us now turn to the GSM. Also in this case, a global contrast variable regulates the reliability
of the stimulus. Tuttavia, in contrast to the one-dimensional toy example presented before, infer-
ence in this case takes place in a higher dimensional space. We again modify the prior distribution
to induce a hypoprior. We do so in the simplest possible way, by scaling the prior covariance
matrix by a constant factor larger than 1.0 (Methods). In Figure 3C we compare the posterior
distributions calculated under the true prior (in green) with those computed under the hypoprior
(in blue). As expected, we again find that hypopriors result in overweighting of sensory stimuli,
with higher posterior means and higher uncertainty about the estimates (Figure 3D, cf. green and
blue lines), consistent with the postulates of Pellicano and Burr (2012).

Perturbing the Network: The Effect of Inhibitory Deficits

We now turn our attention to the network model. In what follows we will refer to the original
SSN, presented in Figure 2, as the neurotypical (NT) rete. As previously stated, the NT-
network was constructed in terms of separate excitatory and inhibitory populations. Here
we target inhibitory connections by scaling down their efficacy by a global constant value
(Methods). In order to ensure that baseline activity levels are not affected, and following the
ideas of Nelson and Valakh (2015), we also scaled excitatory connections globally in a
homeostatic fashion (see Supplementary Figure 1 and Methods). We will henceforth refer to
the network where inhibitory deficits have been induced as the ASD-network. As we did for
the generative model, we then compared the mean and standard deviation of the posterior
distributions encoded by both networks in terms of their response samples (Figure 3E–F). Nota-
bly, we observed that ASD-network representations of the posteriors also seemed to over-
weight current sensory information. Infatti, posterior means were higher in the ASD- di
in the NT-network (Figure 3F top panel, cf. red and blue lines). In passing, we note that
because of the original approximate inference scheme, the scaling of the mean and standard
deviation with contrast between the original network and the posterior are similar but not iden-
tical. In particular, while mean responses in the generative model saturate at high contrasts,
they only decelerate in the network model, without actually saturating. Infatti, responses in
this type of network models do not saturate. They either continue to grow or “bounce back”

Network Neuroscience

202

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Bridging physiological and perceptual views of autism

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Figura 3. Hypopriors and impaired inhibition. (UN, B) Effect of hypopriors on posterior predictions for a 1D toy example. Priors, likelihoods, E
posteriors are all Gaussian. A contrast variable regulating the likelihood precision plays the role of the perceptual reliability of stimuli. Two exam-
ple inference cases are presented: under the true (well-calibrated) prior (dashed, green) and under a wider hypoprior (dashed, blue). (UN) The prior
(dashed, colore) and likelihood (dashed, black) are multiplicatively combined according to Bayes’ rule to form the posterior (continuous, colore). (B)
Posterior mean (top plot) and standard deviation (bottom plot) under the true prior (green) and the hypoprior (blue), as a function of contrast
(likelihood precision). (C, D) Effect of hypopriors on posterior predictions for the full multivariate GSM model. (C) Mean (top plots) and standard
deviation (bottom plots) of latent variable intensities ordered by each latent’s orientation, for each stimulus in Figure 2. Left: for the well-calibrated
ideal observer’s posterior distribution (green). Right: under a hypoprior (blue). (D) Posterior mean (top) and standard deviation (bottom), averaged
across all latent variables, under the true prior (green) and the hypoprior (blue), as a function of contrast. (E, F) Effect of impaired inhibition on
network responses. (E) Mean (top) and standard deviation (bottom) of latent variable intensities ordered by each latent’s orientation, for each
stimulus in the training set. E cell membrane potentials uE from the stationary response distributions for the NT-network (left, red), and for the
ASD-network (right, blue). (F) Mean (top) and standard deviation (bottom) of neural responses, averaged across all cells, for the NT-network
(red) and the ASD-network (blue), as a function of contrast. Circles, and gray dots on x-axis of panels D and F indicate training contrast levels.

Network Neuroscience

203

Bridging physiological and perceptual views of autism

and begin to decrease (Ahmadian et al., 2013). Allo stesso modo, a slightly higher standard deviation is
observed in the network with respect to the posterior at low contrast, which stems from an
underestimation of the variance of neural responses under the Gaussian approximation during
training of the network (Echeveste et al., 2020).

Higher uncertainty about the estimates was also found in the network (Figure 3F bottom
panel, cf. red and blue lines), just as it happened for the generative model under hypopriors
(compare Figure 3, panels D and F). È interessante notare, we have reached the same qualitative traits
by two very different approaches and following two theories expressed at widely different
levels: one perceptual, one physiological.

It is important to note that sampling-based implementations of Bayesian inference establish
a direct link between uncertainty and neural variability, since the width of the posterior dis-
tribution is directly related to the amount of variability. Infatti, we observe that weaker inhi-
bition leads to higher variability in the neural responses of the ASD-network compared with
the NT-network (Figure 3F, bottom panel, cf. red and blue lines), as had been suggested in
Rubenstein and Merzenich (2003), where the point had been made that a disruption of E-I
balance leading to a hyperexcitable cortex would lead to increased cortical “noise.” Indeed,
higher neural variability has been experimentally reported in ASD subjects both in EEG (Milne,
2011) and in fMRI (Haigh et al., 2015) studies.

An advantage of employing a neural network model such as the SSN, which shows char-
acteristic features of cortical dynamics, such as gamma oscillations and transient overshoots
(including their contrast dependence), is that we can also explore the predictions the model
makes for these features, now for the ASD-network.

Primo, we look at gamma oscillations. To that end we computed the power spectrum from
the local field potential (LFP), from which we extracted the peak gamma frequency for different
contrast levels for both networks (Figure 4A). We note that the overall frequency modulation is
very similar in both networks, with slightly higher peak gamma frequency in the ASD-network
for high contrast stimuli (cf. Figure 4B, left panel, red and blue). Previous work has reported
higher peak gamma frequency in ASD subjects solving a visual task, which was interpreted as
a sign of “increased neural inhibition” (Dickinson, Bruyns-Haylett, Smith, Jones, & Milne,
2016). At first glance, this might seem at odds with the starting point for our work where
we have weakened inhibitory synapses. It is worth noting however that total inputs (both E
and I) result in a balanced recurrent network from a dynamic equilibrium, which may result
in higher inhibitory currents, despite weaker inhibitory synapses. This is precisely the case
here (see Supplementary Figure 1D). Infatti, it has been known for decades that balanced
networks are prone to so-called paradoxical effects (Tsodyks, Skaggs, Sejnowski, & McNaughton,
1997), whereby direct external inhibitory inputs to I cells can actually lead to increased I rates.
This also hints at why seemingly contradictory results are often found regarding inhibition in
ASD depending on what exactly is chosen as a measure of inhibition.

È interessante notare, gamma power is higher for the ASD-network (see sharper gamma peaks in the
spectra of Figure 4A, and in Figure 4B, right plot, blue vs. red). An insight into the functional
interpretation of this effect can be obtained from analyzing neural responses at zero contrast,
representing what is usually termed spontaneous activity in the literature. In sampling-based
models, such as this one, spontaneous activity is postulated to encode the prior distribution
(Berkes et al., 2011). Infatti, when the stimulus is completely uninformative, as is the case at
zero contrast, the posterior matches the prior. The model hence predicts higher gamma power
in spontaneous activity, which is in line with previous reports of higher gamma band power in
resting-state activity of ASD subjects (van Diessen et al., 2015).

Network Neuroscience

204

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Bridging physiological and perceptual views of autism

Figura 4. Transient responses and oscillations. (UN) LFP power as a function of frequency for stimuli of different contrast levels (same stimuli and
colors as in Figure 3) in the NT-network (left), and in the ASD-network (right). Both networks present strong gamma oscillations (see peaks in the
gamma band, indicated by empty circles). (B) Comparison of oscillatory behavior in both networks. On the left, the peak gamma frequency is
presented as a function of stimulus contrast for both networks. Very minimal differences are observed. On the right, the total power within the
gamma band is presented as a function of contrast for both networks. A higher gamma power is observed for the ASD-network at all contrasts,
with strong differences at low contrasts. (C) Across-trial average transient responses for stimuli of different contrast levels in the neurotypical
rete (left) and in the ASD-network (right). Both networks present strong stimulus-dependent transient overshoots. (D) Comparison of over-
shoot sizes. The maximal firing rate is presented as a function of stimulus contrast for both networks. We observe that the ASD-network presents
stronger peak responses at higher contrasts, overreacting to intense stimuli. NT-network results reproduced from Echeveste et al. (2020).

We finally turn our attention to transient responses. We compared the ASD- and NT-
networks in terms of their trial-averaged firing rates around stimulus onset (Figure 4C). IL
model predicts higher maximal firing rates (and not only mean rates) for the ASD-network than
for the NT-network at intermediate and high contrasts (cf. Figure 4D, red and blue), indicating
that the ASD-network has become hypersensitive to intense stimuli. We note that theories of
perception expressed in terms of predictive coding usually interpret peak rates as a measure of
sorpresa, novelty, or unexpectedness (Rao & Ballard, 1999), and indeed a predictive coding
account of ASD perceptual traits, including abnormal sensory sensitivity, has been postulated
by several authors in the past (Van Boxtel & Lu, 2013; Van de Cruys et al., 2014). Results from
the ASD-network, which we here interpret from a Bayesian inference perspective, are then not
inconsistent with a predictive coding view of perceptual differences in the ASD population.

DISCUSSION

Neural network models are increasingly being used as a tool to study how differences in neural
architectures may be linked to symptoms in different disorders (Lanillos et al., 2020). In this
work we have employed a neural network model of a V1 cortical hypercolumn trained to
perform sampling-based probabilistic inference in a visual task to build a mechanistic bridge
between descriptions of ASD formulated at two very different levels: a physiological level, In
terms of inhibitory dysfunction (Rubenstein & Merzenich, 2003), neural variability (Haigh
et al., 2015; Milne, 2011), and gamma oscillations (van Diessen et al., 2015); and a perceptual
level, in terms of hypopriors in Bayesian computations (Pellicano & Burr, 2012). In what fol-
lows we describe merits of this work, limitations, and open questions.

Network Neuroscience

205

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Bridging physiological and perceptual views of autism

Merits

We have taken two parallel paths: in one perturbing the probabilistic generative model in
order to induce hypopriors, and in the other perturbing the neural network model to induce
an inhibitory dysfunction. We observed that both approaches lead to consistent results in terms
of the represented posterior distributions, providing support for the possibility that both views
of ASD might actually constitute two sides of the same coin.

Employing a neural network model such as the SSN, which not only performs inference in a
perceptual task but also displays characteristic features of cortical dynamics while doing so
(Echeveste et al., 2020), allowed us to make further connections between characteristic differences
in these dynamics and inhibitory dysfunction in ASD subjects. Stimulus-dependent variability
modulations in the network, and concretely the direct link between neural variability and uncer-
tainty established by sampling-based implementations of inference, predicted higher variability in
neural responses in the ASD- versus the NT-network. Infatti, increased neural variability has been
reported in ASD subjects both in EEG (Milne, 2011) and in fMRI (Haigh et al., 2015) studies. More-
Sopra, transient overshoots, usually interpreted in predictive coding theories to represent novelty,
sorpresa, or unexpectedness (Rao & Ballard, 1999), are present in the network, with higher
responses for strong stimuli in the ASD-network versus the NT-network, indicating an oversensi-
tivity to intense stimuli, a feature often reported in children with ASD (Kern et al., 2006).

Inoltre, oscillations in the ASD-network displayed higher gamma band oscillatory
power, consistent with observations in resting-state EEG recordings of ASD subjects (van
Diessen et al., 2015). Peak gamma frequencies were also higher in the ASD-network for
high-contrast stimuli, a fact that has indeed been observed in EEG recordings from subjects
performing an orientation discrimination task (Dickinson et al., 2016), and that had been
attributed to increased inhibition. We confirmed that, despite having decreased the efficacy
of inhibitory synapses in our network, mean inhibitory inputs were indeed actually larger
for high-contrast stimuli. This observation is in line with the known fact that balanced E-I net-
works are prone to “paradoxical effects” regarding inhibition (Tsodyks et al., 1997), Dove
average rates result from a dynamic balance of excitation and inhibition, and might explain
apparent contradictions between studies reporting increased/decreased inhibition (Cellot &
Cherubini, 2014; Dickinson et al., 2016). These results also highlight the importance of neural
network simulations to assist in the interpretation of physiological observations regarding the
role of inhibition in cortical recordings.

Limitations and Open Questions

Training recurrent neural networks with expansive nonlinearities beyond mean responses is
currently a challenging and computationally expensive task. These networks are prone to
instabilities and current optimization for second-order moments requires either a large number
of trials, or matrix-matrix operations that scale as n3 in the number of neurons (Hennequin &
Lengyel, 2016). Infatti, the choice of the simple generative model played a key role in order
to make the training problem tractable with currently available optimization techniques, but it
imposes some limitations. The GSM produces multivariate Gaussian posteriors (which enabled
training the network with currently available second-order moment-matching methods), E
was further constructed to be rotationally symmetric (which drastically reduced the number of
network parameters to be optimized, as well as the required number of training examples). UN
model constructed in this way will however not be able to capture features of human behavior
in popular tests of visual perception, such as the “oblique effect,” where neurotypical subjects
seem to be more sensitive to cardinal orientations (Westheimer & Beard, 1998), an effect that is

Network Neuroscience

206

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Bridging physiological and perceptual views of autism

reduced in ASD subjects (Dickinson et al., 2014). Tackling problems like these in a sampling-
based setting will require developing tools to train more flexible networks that can produce
richer posterior distributions. It should be noted that these limitations are, Tuttavia, of a tech-
nical nature, and are not inherent to the sampling-based inference framework.

Secondo, the model employed to explain simple, low-level perceptual computations was
constructed in terms of a single V1 hypercolumn, and is hence only able to capture local
dynamical features, such as locally generated gamma oscillations. Hypothetically, the ideas
presented here can be extended to the representation of other circular variables beyond
orientation of visual stimuli, such as head direction in rodents (Skaggs, Knierim, Kudrimoti,
& McNaughton, 1995), motor intent in primates (Georgopoulos, Taira, & Lukashin, 1993),
physical space in grid cells (McNaughton, Battaglia, Jensen, Moser, & Moser, 2006), or ocu-
lomotor control (Seung, 1998). In all these examples, highly specialized brain areas receive
assorted inputs that carry a noisy, filtered, and distributed representation of a circular variable.
The recurrent activity of the network constitutes a mechanistic implementation of an inference
processi, which could be potentially executed through a sampling-based Bayesian inference
strategy, as explored here. If that were the case, the strong reliance of ASD subjects on the
likelihood could also be broadened beyond the realm of sensory processing. Extensions of
these ideas are also conceivable to other one-dimensional, yet aperiodic, domini, ad esempio
sound pitch (Aronov, Nevers, & Tank, 2017), navigation speed (Kropff, Carmichael, Moser,
& Moser, 2015), or elapsed time (Tsao et al., 2018) Quale, although still fairly narrow in their
semantic content, involve some degree of higher level processing. Tuttavia, as we progress
into still higher cognitive functions, the understanding of how context-dependent modulations
of cortical dynamics emerge during complex perceptual tasks will likely require models where
multiple circuits interact (Simone & Wallace, 2016). In this sense, hierarchical or spatially
extended versions of the SSN model employed here may provide adequate substrates to study
inference of higher level perceptual tasks where longer range aspects of cortical dynamics,
such as gamma synchronization, might emerge.

Third, we have focused on one aspect of probabilistic inference: inferring the state of a set
of latent variables under perceptual uncertainty. The study of other aspects of this problem,
such as inferring temporal transitions (Sinha et al., 2014), or causal relationships (Noel,
Shivkumar, Dokka, Haefner, & Angelaki, 2021), and their link to altered inhibition and neural
dynamics, will require the use of different architectures and generative models and constitute
worthwhile avenues of future research.

Closing Remarks

We have shown how recurrent neural networks optimized for sampling-based inference are
viable candidates to bridge the gap between Bayesian perceptual theories of ASD and their
physiological underpinnings in terms of inhibitory dysfunction, neural variability, and oscilla-
zioni. We believe these results highlight the potential for the use of the emerging body of
function-optimized neural networks (Echeveste et al., 2020; Hennequin, Vogels, & Gerstner,
2014; Orhan & Mamma, 2017; Remington, Narain, Hosseini, & Jazayeri, 2018; Song, Yang, &
Wang, 2016; Yamins et al., 2014) as models to establish mechanistic links between neural
activity and computations in the cortex that go beyond the study of neurotypical perception.

METHODS

In order to link cortical dynamics and probabilistic computations we modified the parameters
of the probabilistic and network models employed in Echeveste et al. (2020). In what follows

Network Neuroscience

207

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Bridging physiological and perceptual views of autism

we describe those changes and refer the reader to the original paper for a more detailed
description of the models and of the original model parameters.

The Generative Model

In this work the Gaussian scale mixture model (GSM; Wainwright & Simoncelli, 2000) È
employed as the generative model of natural images (at the level of small patches) under
which inference is carried out in the primary visual cortex ( V1; Coen-Cagli et al., 2015; Orbán
et al., 2016). Under the GSM an image patch x is obtained by linearly combining a number of
local features (given by the columns of a matrix A), which are weighted by a corresponding
number of feature coefficients given by y, further scaled by a single contrast variable z, E
finally corrupted by additive white Gaussian noise. This forward generative model can then be
summarized in terms of the likelihood function given by
(cid:3)

(cid:1)
xjy; z ∼ N z A y; σ2
xI

;

(1)

together with the priors for the feature coefficients and the contrast variable z. Local features
were assumed to be drawn from a multivariate Gaussian:

y ∼ N 0; Cð

Þ;

(2)

and the contrast was assumed to be drawn from a gamma prior. To induce hypopriors we
modified the overall scale of the prior covariance matrix C, by taking CHP = αHPC, with αHP =
1.5. Other values were explored without qualitative differences (not shown). We note that
taking αHP > 1 results in wider priors, as required for a hypoprior.

The one-dimensional toy example model of Figure 3A–B, corresponds to a one-
x = 100. As in the full GSM, we

dimensional GSM with prior variance C = 4, A = 10, and σ2
took αHP = 1.5.

Network Dynamics and Architecture
The circuit model consisted of a nonlinear, stochastic network respecting Dale’s principle,
with NE excitatory and NI inhibitory neurons. The evolution of the membrane potential ui of
each neuron i in this model is described by Hennequin et al. (2018):

τi

dui
dt

¼ −ui tð Þ þ hi tð Þ þ

Wijrj tð Þ þ η

i tð Þ;

(3)

where τi represents the membrane time constant for neuron i, hi is its feedforward input, and ηi
is the process noise (capturing both intrinsic and extrinsic forms of neural variability). W è il
matrix of recurrent connections, and hence Wij represents the strength of the synapse connect-
ing neuron j to neuron i. As previously mentioned, the network is nonlinear, with firing rates

ri tð Þ ¼ k ui tð Þ

cm:

(4)

Here k and m represent the scale and exponent of the firing rate nonlinearity (Ahmadian et al.,
2013). Given the rotational symmetry of the problem, W itself was parametrized to be rota-
tionally symmetric. Neurons in the model are arranged in a ring of pairs of E and I cells accord-
ing to their preferred orientations (Figure 1C), where Wij was a smoothly decaying function of
the tuning difference between neurons i and j (see Supplementary Figure 1A, top and second
row). IL (stimulus-independent) process noise covariance was analogously parametrized (Vedere
Supplementary Figure 1A, third row). Following canonical models of V1 simple cells (Dayan &
Abbott, 2001), feedforward inputs to the network were computed by applying a linear filter Wff

Network Neuroscience

208

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Bridging physiological and perceptual views of autism

to the stimulus (the image patch) followed by a nonlinearity (see Supplementary Figure 1A,
bottom row).

The perturbation here employed to induce an inhibitory deficit has a single free parameter δI
ASD
(see Supplementary Figure 1A–B). In
that scales the inhibitory columns of W, W
IO
order to maintain the baseline level of activity, a second modification is introduced (simulating
homeostatic adaption of the excitatory connections), scaling the excitatory columns of W by a
NT
factor δE: W
E . This second factor was found by grid-search minimization of the
homeostatic cost

ASD
E = (1 − δE)W

NT
= (1 − δI)W
IO

(cid:4)
(cid:4)
Ch ¼ μNT

s − μASD

(cid:4)
(cid:4);

(5)

capturing the change in mean spontaneous activity levels (μs) between the original NT- E
perturbed ASD-network. This adaptation procedure returns a single δE value for each δI value
(Supplementary Figure 1C). We note that excitatory changes via this procedure always resulted
smaller than inhibitory ones (cf. to identity line in Supplementary Figure 1C, bottom plot).
Network results presented throughout this paper correspond to δI = 0.1, for which δE =
0.076. Numerical experiments were repeated for δI = 0.05 and δI = 0.15 without qualitative
differences (not shown).

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Numerical Simulations

Stationary moments of neural responses to a fixed input (Figure 3E) were computed from
20,000 independent samples (200 ms apart) generated by letting neural activity in the network
evolve over time via Equation 3 (excluding transients). Power spectra in Figure 4A were
obtained from simulated local field potentials (LFPs), computed as the average (across-cell)
membrane potential. Gamma peak frequencies in Figure 4B (left) were obtained as the local
maximum in the spectrum within the gamma range (20–80 Hz), while total gamma power in
Figure 4B (right) was computed as the integral of the spectrum over that same range.

Transient responses displayed in Figure 4C were computed as the mean (across E-cells and
trials) firing rates (n = 100), which are then further averaged over a 10-ms sliding window. UN
random delay time (sampled from a truncated Gaussian, with a mean of 45 ms and a standard
deviation of 5 ms) was employed for the feedforward input to each pair of E–I cells. These
procedures had been put in place to allow for a comparison to experimental data, and are
here kept in order to compare the ASD-netowork to replotted results from the original (here
NT-) rete. Maximal firing rates in Figure 4D were obtained as the peak rates from transient
firing rate responses.

ACKNOWLEDGMENTS

This work was supported by Argentina’s National Scientific and Technical Research Council
(CONICET), which covered all researchers’ salaries. We are grateful to Y. Nagai for pointing
out this potential avenue of research after discussing previous work.

CODE AVAILABILITY

IL (Python) code to create the ASD-network is provided in bitbucket.org/ RSE_1987
/inhibitory_dysfunction (Echeveste, 2021). The code for the numerical experiments can be
found at bitbucket.org/RSE_1987/ssn_inference_numerical_experiments (Echeveste, 2020).

Network Neuroscience

209

Bridging physiological and perceptual views of autism

SUPPORTING INFORMATION

Supporting information for this article is available at https://doi.org/10.1162/netn_a_00219.

AUTHOR CONTRIBUTIONS

Rodrigo Echeveste: Conceptualization; Formal analysis; Investigation; Visualization; Writing –
original draft. Enzo Ferrante: Conceptualization; Writing – original draft. Diego H. Milone:
Conceptualization; Supervision; Writing – original draft. Inés Samengo: Conceptualization;
Supervision; Writing – original draft.

FUNDING INFORMATION

Rodrigo Echeveste, Santa Fe Agency for Science, Tecnologia, and Innovation, Award ID:
IO-138-19.

REFERENCES

Ahmadian, Y., Rubin, D., & Mugnaio, K. (2013). Analysis of the stabilized
supralinear network. Calcolo neurale, 25(8), 1994–2037.
https://doi.org/10.1162/NECO_a_00472, PubMed: 23663149
Aitchison, L., & Lengyel, M. (2016). The Hamiltonian brain: Efficient
probabilistic inference with excitatory-inhibitory neural circuit
dynamics. PLoS Computational Biology, 12(12), e1005186.
https://doi.org/10.1371/journal.pcbi.1005186, PubMed:
28027294

Aitchison, L., & Lengyel, M. (2017). With or without you: Predictive
coding and Bayesian inference in the brain. Current Opinion in
Neurobiology, 46, 219–227. https://doi.org/10.1016/j.conb.2017
.08.010, PubMed: 28942084

American Psychiatric Association. (2013). Diagnostic and statistical
manual of mental disorders (DSM-5). American Psychiatric Associa-
tion Publishing. https://doi.org/10.1176/appi.books.9780890425596
Aronov, D., Nevers, R., & Tank, D. W. (2017). Mapping of a
non-spatial dimension by the hippocampal-entorhinal circuit.
Nature, 543(7647), 719–722. https://doi.org/10.1038
/nature21692, PubMed: 28358077

Bastos, A., Usrey, W., Adams, R., Mangun, G., Fries, P., & Friston,
K. (2012). Canonical microcircuits for predictive coding. Neuron,
76(4), 695–711. https://doi.org/10.1016/j.neuron.2012.10.038,
PubMed: 23177956

Berkes, P., Orbán, G., Lengyel, M., & Fiser, J. (2011). Spontaneous
cortical activity reveals hallmarks of an optimal internal model of
the environment. Scienza, 331(6013), 83–87. https://doi.org/10
.1126/science.1195870, PubMed: 21212356

Bolton, P. F., Carcani-Rathwell, I., Hutton, J., Goode, S., Howlin, P.,
& Rutter, M. (2011). Epilepsy in autism: Features and correlates.
British Journal of Psychiatry, 198(4), 289–294. https://doi.org/10
.1192/bjp.bp.109.076877, PubMed: 21972278

Carandini, M., & Heeger, D. (2012). Normalization as a canonical
neural computation. Nature Reviews Neuroscience, 13(1), 51.
https://doi.org/10.1038/nrn3136, PubMed: 22108672

Cellot, G., & Cherubini, E. (2014). Gabaergic signaling as therapeu-
tic target for autism spectrum disorders. Frontiers in Pediatrics, 2,
70. https://doi.org/10.3389/fped.2014.00070, PubMed:
25072038

Churchland, M. M., Yu, B. M., Cunningham, J. P., Sugrue, l. P.,
Cohen, M. R., Corrado, G. S., … Shenoy, K. V. (2010). Stimulus
onset quenches neural variability: A widespread cortical phe-
nomenon. Nature Neuroscience, 13(3), 369. https://doi.org/10
.1038/nn.2501, PubMed: 20173745

Coen-Cagli, R., Kohn, A., & Schwartz, O. (2015). Flexible gating of
contextual influences in natural vision. Nature Neuroscience, 18,
1648–1655. https://doi.org/10.1038/nn.4128, PubMed: 26436902
Dayan, P., & Abbott, l. (2001). Theoretical neuroscience (Vol. 806).

Cambridge, MA: CON Premere.

Dickinson, A., Bruyns-Haylett, M., Smith, R., Jones, M., & Milne, E.
(2016). Superior orientation discrimination and increased peak
gamma frequency in autism spectrum conditions. Journal of
Abnormal Psychology, 125(3), 412. https://doi.org/10.1037
/abn0000148, PubMed: 27043918

Dickinson, A., Jones, M., & Milne, E. (2014). Oblique orientation
discrimination thresholds are superior in those with a high level
of autistic traits. Journal of Autism and Developmental Disorders,
44(11), 2844–2850. https://doi.org/10.1007/s10803-014-2147-1,
PubMed: 24870402

Echeveste, R. (2020). SSN inference numerical experiments, Bitbucket,
https:// bitbucket.org/ RSE_1987/ssn_inference_numerical
_experiments

Echeveste, R. (2021). ASD sampling, Bitbucket, https://bitbucket

.org/RSE_1987/inhibitory_dysfunction

Echeveste, R., Aitchison, L., Hennequin, G., & Lengyel, M. (2020).
Cortical-like dynamics in recurrent circuits optimized for
sampling-based probabilistic inference. Nature Neuroscience,
23(9), 1138–1149. https://doi.org/10.1038/s41593-020-0671-1,
PubMed: 32778794

Fiser, J., Berkes, P., Orbán, G., & Lengyel, M. (2010). Statistically
optimal perception and learning: From behavior to neural repre-
sentations. Trends in Cognitive Sciences, 14(3), 119–130. https://
doi.org/10.1016/j.tics.2010.01.003, PubMed: 20153683

Georgopoulos, UN. P., Taira, M., & Lukashin, UN. (1993). Cognitive
neurophysiology of
the motor cortex. Scienza, 260(5104),
47–52. https://doi.org/10.1126/science.8465199, PubMed:
8465199

Network Neuroscience

210

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Bridging physiological and perceptual views of autism

H a e f n e r, R . , B e r k e s , P. , & F i s e r,

( 2 0 1 6 ) . P e r c e p t u a l
decision-making as probabilistic inference by neural sampling.
Neuron, 90(3), 649–660. https://doi.org/10.1016/j.neuron.2016
.03.020, PubMed: 27146267

J .

Haider, B., Häusser, M., & Carandini, M. (2013). Inhibition domi-
nates sensory responses in the awake cortex. Nature, 493(7430),
97–100. https://doi.org/10.1038/nature11665, PubMed:
23172139

Haigh, S. M., Heeger, D. J., Dinstein, I., Minshew, N., & Behrmann,
M. (2015). Cortical variability in the sensory-evoked response in
autism. Journal of Autism and Developmental Disorders, 45(5),
1176–1190. https://doi.org/10.1007/s10803-014-2276-6,
PubMed: 25326820

Happé, F., & Frith, U. (2006). The weak coherence account:
Detail-focused cognitive style in autism spectrum disorders. Jour-
nal of Autism and Developmental Disorders, 36(1), 5–25. https://
doi.org/10.1007/s10803-005-0039-0, PubMed: 16450045

Hénaff, O. J., Boundy-Singer, Z. M., Meding, K., Ziemba, C. M., &
Goris, R. l. (2020). Representation of visual uncertainty through
neural gain variability. Nature Communications, 11(1), 1–12.
https://doi.org/10.1038/s41467-020-15533-0, PubMed: 32427825
Hennequin, G., Ahmadian, Y., Rubin, D., Lengyel, M., & Mugnaio, K.
(2018). The dynamical regime of sensory cortex: Stable dynamics
around a single stimulus-tuned attractor account for patterns of
noise variability. Neuron, 98(4), 846–860. https://doi.org/10
.1016/j.neuron.2018.04.017, PubMed: 29772203

Hennequin, G., & Lengyel, M. (2016). Characterizing variability in

nonlinear recurrent neuronal networks. arXiv:1610.03110.

Hennequin, G., Vogels, T., & Gerstner, W. (2014). Optimal control
of transient dynamics in balanced networks supports generation
of complex movements. Neuron, 82(6), 1394–1406. https://doi
.org/10.1016/j.neuron.2014.04.045, PubMed: 24945778

Horder, J., Andersson, M., Mendez, M. A., Singh, N., Tangen, Ä.,
Lundberg, J., … Borg, J. (2018). GABAA receptor availability is not
altered in adults with autism spectrum disorder or in mouse
models. Science Translational Medicine, 10(461). https://doi.org
/10.1126/scitranslmed.aam8434, PubMed: 30282698

Kern, J. K., Trivedi, M. H., Garver, C. R., Grannemann, B. D.,
Andrews, UN. A., Savla, J. S., … Schroeder, J. l. (2006). The pattern
of sensory processing abnormalities in autism. Autism, 10(5),
480–494. https://doi.org/10.1177/1362361306066564,
PubMed: 16940314

Knill, D., & Richards, W. (1996). Perception as Bayesian inference.
C a m b r i d g e U n i v e r si t y P r e s s . h t t p s : / / d o i . o rg / 1 0 . 1 0 1 7
/CBO9780511984037

Kropff, E., Carmichael, J. E., Moser, M.-B., & Moser, E. IO. (2015).
Speed cells in the medial entorhinal cortex. Nature, 523(7561),
419–424. https://doi.org/10.1038/nature14622, PubMed:
26176924

Lanillos, P., Oliva, D., Philippsen, A., Yamashita, Y., Nagai, Y., &
Cheng, G. (2020). A review on neural network models of schizo-
phrenia and autism spectrum disorder. Neural Networks, 122,
338–363. https://doi.org/10.1016/j.neunet.2019.10.014,
PubMed: 31760370

Mamma, W., Beck, J., Latham, P., & Pouget, UN. (2006). Bayesian infer-
ence with probabilistic population codes. Nature Neuroscience,
9(11), 1432–1438. https://doi.org/10.1038/nn1790, PubMed:
17057707

McNaughton, B. L., Battaglia, F. P., Jensen, O., Moser, E. I., &
Moser, M.-B. (2006). Path integration and the neural basis of
the “cognitive map.” Nature Reviews Neuroscience, 7(8),
663–678. https://doi.org/10.1038/nrn1932, PubMed: 16858394
Milne, E. (2011). Increased intra-participant variability in children
with autistic spectrum disorders: Evidence from single-trial anal-
ysis of evoked EEG. Frontiers in Psychology, 2, 51. https://doi.org
/10.3389/fpsyg.2011.00051, PubMed: 21716921

Mottron, L., Dawson, M., Soulieres, I., Hubert, B., & Burack, J.
(2006). Enhanced perceptual functioning in autism: An update,
and eight principles of autistic perception. Journal of Autism
and Developmental Disorders, 36(1), 27–43. https://doi.org/10
.1007/s10803-005-0040-7, PubMed: 16453071

Nelson, S. B., & Valakh, V. (2015). Excitatory/inhibitory balance
and circuit homeostasis in autism spectrum disorders. Neuron,
87(4), 684–698. https://doi.org/10.1016/j.neuron.2015.07.033,
PubMed: 26291155

Noel, J.-P., Shivkumar, S., Dokka, K., Haefner, R., & Angelaki, D.
(2021). Aberrant causal inference and presence of a compensa-
tory mechanism in autism spectrum disorder. PsyArXiv. https://
doi.org/10.31234/osf.io/7gjhs

Orbán, G., Berkes, P., Fiser, J., & Lengyel, M. (2016). Neural vari-
ability and sampling-based probabilistic representations in the
visual cortex. Neuron, 92(2), 530–543. https://doi.org/10.1016/j
.neuron.2016.09.038, PubMed: 27764674

Orhan, A., & Mamma, W. (2017). Efficient probabilistic inference in
generic neural networks trained with non-probabilistic feedback.
Nature Communications, 8(1), 138. https://doi.org/10.1038
/s41467-017-00181-8, PubMed: 28743932

Ozonoff, S., Heung, K., Byrd, R., Hansen, R., & Hertz-Picciotto, IO.
(2008). The onset of autism: Patterns of symptom emergence in
the first years of life. Autism Research, 1(6), 320–328. https://doi
.org/10.1002/aur.53, PubMed: 19360687

Palmer, C. J., Lawson, R. P., & Hohwy, J. (2017). Bayesian
approaches to autism: Towards volatility, action, and behavior.
Psychological Bulletin, 143(5), 521. https://doi.org/10.1037
/bul0000097, PubMed: 28333493

Pellicano, E., & Burr, D. (2012). When the world becomes “too
real”: A Bayesian explanation of autistic perception. Trends in
Cognitive Sciences, 16(10), 504–510. https://doi.org/10.1016/j
.tics.2012.08.009, PubMed: 22959875

Rao, R., & Ballard, D. (1999). Predictive coding in the visual cortex:
A functional interpretation of some extra-classical receptive-field
effects. Nature Neuroscience, 2(1), 79. https://doi.org/10.1038
/4580, PubMed: 10195184

Ray, S., & Maunsell, J. H. (2010). Differences in gamma frequencies
across visual cortex restrict their possible use in computation.
Neuron, 67(5), 885–896. https://doi.org/10.1016/j.neuron.2010
.08.004, PubMed: 20826318

Remington, E., Narain, D., Hosseini, E., & Jazayeri, M. (2018). Flex-
ible sensorimotor computations through rapid reconfiguration of
cortical dynamics. Neuron, 98(5), 1005–1019. https://doi.org/10
.1016/j.neuron.2018.05.020, PubMed: 29879384

Roberts, M., Lowet, E., Brunet, N., Ter Wal, M., Tiesinga, P., Fries,
P., & De Weerd, P. (2013). Robust gamma coherence between
macaque V1 and V2 by dynamic frequency matching. Neuron,
78(3), 523–536. https://doi.org/10.1016/j.neuron.2013.03.003,
PubMed: 23664617

Network Neuroscience

211

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
N
e
N
UN
R
T
io
C
e
–
P
D

F
/

6
1
1
9
6
2
0
0
2
4
1
4
N
e
N
_
UN
_
0
0
2
1
9
P
D

B
sì
G
tu
e
S
T

o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3

Bridging physiological and perceptual views of autism

Robertson, C. E., Ratai, E.-M., & Kanwisher, N. (2016). Reduced
gabaergic action in the autistic brain. Current Biology, 26(1),
80–85. https://doi.org/10.1016/j.cub.2015.11.019, PubMed:
26711497

Rosenberg, A., Patterson, J. S., & Angelaki, D. E. (2015). A compu-
tational perspective on autism. Proceedings of the National Acad-
emy of Sciences, 112(30), 9158–9165. https://doi.org/10.1073
/pnas.1510583112, PubMed: 26170299

Rubenstein, J., & Merzenich, M. M. (2003). Model of autism:
Increased ratio of excitation/inhibition in key neural systems.
Genes, Brain and Behavior, 2(5), 255–267. https://doi.org/10
.1034/j.1601-183X.2003.00037.x, PubMed: 14606691

Schwartz, O., Sejnowski, T., & Dayan, P. (2009). Perceptual orga-
nization in the tilt illusion. Journal of Vision, 9(4), 19.1–1920.
https://doi.org/10.1167/9.4.19, PubMed: 19757928

Seung, S. (1998). Cognitive neurophysiology of the motor cortex.
Neural Networks, 11(7–8), 1253–1258. https://doi.org/10.1016
/S0893-6080(98)00064-1

Simone, D. M., & Wallace, M. T. (2016). Dysfunction of sensory
oscillations in autism spectrum disorder. Neuroscience & Bio-
behavioral Reviews, 68, 848–861. https://doi.org/10.1016/j
.neubiorev.2016.07.016, PubMed: 27451342

Sinha, P., Kjelgaard, M. M., Gandhi, T. K., Tsourides, K., Cardinaux,
UN. L., Pantazis, D., … Held, R. M. (2014). Autism as a disorder of
prediction. Proceedings of the National Academy of Sciences,
111 ( 4 2 ) , 1 5 2 2 0 – 1 5 2 2 5 . h t t p s : / / d o i . o rg / 1 0 . 1 0 7 3 / p n a s
.1416797111, PubMed: 25288765

Skaggs, W. E., Knierim, J. J., Kudrimoti, H. S., & McNaughton, B. l.
(1995). A model of the neural basis of the rat’s sense of direction.
Advances in Neural Information Processing Systems, 7, 173–180.
PubMed: 11539168

Song, H., Yang, G., & Wang, X. (2016). Training excitatory-
inhibitory recurrent neural networks for cognitive tasks: A simple
and flexible framework. PLoS Computational Biology, 12(2),
e1004792. https://doi.org/10.1371/journal.pcbi.1004792,
PubMed: 26928718

Tsao, A., Sugar, J., Lu, L., Wang, C., Knierim, J. J., Moser, M. B., &
Moser, E. IO. (2018). Integrating time from experience in the lateral
entorhinal cortex. Nature, 561(7721), 57–62. https://doi.org/10
.1038/s41586-018-0459-6, PubMed: 30158699

Tsodyks, M. V., Skaggs, W. E., Sejnowski, T. J., & McNaughton, B. l.
(1997). Paradoxical effects of external modulation of inhibitory
interneurons. Journal of Neuroscience, 17(11), 4382–4388.
https://doi.org/10.1523/ JNEUROSCI.17-11-04382.1997,
PubMed: 9151754

Van Boxtel, J. J., & Lu, H. (2013). A predictive coding perspective
on autism spectrum disorders. Frontiers in Psychology, 4, 19.
https://doi.org/10.3389/fpsyg.2013.00019, PubMed: 23372559
Van de Cruys, S., Evers, K., Van der Hallen, R., Van Eylen, L.,
Boets, B., de Wit, L., & Wagemans, J. (2014). Precise minds
in uncertain worlds: Predictive coding in autism. Psicologico
Review, 121(4), 649. https://doi.org/10.1037/a0037665, PubMed:
25347312

van Diessen, E., Senders, J., Jansen, F. E., Boersma, M., & Bruining,
H. (2015). Increased power of resting-state gamma oscillations
in autism spectrum disorder detected by routine electroencepha-
lography. European Archives of Psychiatry and Clinical Neuro-
science, 265(6), 537–540. https://doi.org/10.1007/s00406-014
-0527-3, PubMed: 25182536

Wainwright, M., & Simoncelli, E. (2000). Scale mixtures of Gauss-
ians and the statistics of natural images. In Advances in Neural
Information Processing Systems (pag. 855–861).

Westheimer, G., & Beard, B. l. (1998). Orientation dependency for
foveal line stimuli: Detection and intensity discrimination, reso-
lution, orientation discrimination and vernier acuity. Vision
Research, 38(8), 1097–1103. https://doi.org/10.1016/S0042
-6989(97)00248-4, PubMed: 9666969

Yamins, D., Hong, H., Cadieu, C., Solomon, E., Seibert, D., &
DiCarlo, J. (2014). Performance-optimized hierarchical models
predict neural responses in higher visual cortex. Proceedings of
the National Academy of Sciences, 111(23), 8619–8624. https://
doi.org/10.1073/pnas.1403112111, PubMed: 24812127