An ACT-R Representation of Information Processing in Autism

Michael Matessa (mmatessa@alionscience.com)
Alion Science and Technology
1789 S. Braddock Ave, Suite 400
Pittsburgh, PA 15218 USA


Abstract
The low-level cognitive processes involved in autism are not
well understood and are a target of ongoing research. This
paper proposes that autistic behavior can be modeled as an
adaptive response to underconnectivity in certain areas of the
brain. In the ACT-R architecture, this is represented by a
reduction in source activations between the declarative
module and other modules, corresponding to
underconnectivity between the dorsolateral prefrontal cortex
and other regions (anterior cingulate cortex, motor, parietal,
and fusiform). Resulting errors in contextual memory
retrieval can reward strategies using visual problem
representation in the parietal area.
Keywords: Autism; modeling; underconnectivity
Introduction
Autism is a condition involving a qualitative impairment in
social interaction and communication, as well as repetitive
behaviors or narrow, obsessive interests. The low-level
cognitive processes involved in autism are not well
understood and are a target of ongoing research. The
following behavioral pattern of impaired/intact capabilities
is emerging. No impairment has been found for simple
visual skills in tasks such as guided saccade (Luna et al.,
2002); simple motor skills in tasks such as finger tapping
(Minshew, Goldstein, & Siegel, 1997); inhibition in tasks
such as Stroop (Ozonoff & Jensen, 1999) and “go-no-go”
(Schmitz et al., 2006); and simple memory in tasks such as
recognition (Bennetto, Pennington, & Rogers, 1996), letter
sequence (Minshew & Goldstein, 2001), unrelated free
recall (Smith, Gardiner, & Bowler, 2007), and syntactic
priming (Preissler, personal communication, November 11
2007) Impairment has been found for contextual memory in
tasks such as semantically related free recall (Smith,
Gardiner, & Bowler, 2007), word and sentence span
(Minshew & Goldstein, 2001), and sentence comprehension
(Müller et al., 1998).
Self-report of people with autism indicate a preference for
visual representation. Examples of this visual mode of
thinking include Temple Grandin’s design of livestock
facilities (Grandin, 1995) and Daniel Tammet’s
multiplication of large numbers (Tammet, 2006). Using
fMRI studies, this visual representation has been shown to
occur in the parietal area of the brain. For example, Kana et
al. (2006) found that an autism group activated parietal
brain regions associated with imagery for comprehending
both low and high imagery sentences, suggesting that they
were using mental imagery in both conditions. In contrast,
imaging studies have found lower activation for autistic
groups compared to control groups in the dorsolateral
prefrontal cortex (DLPFC) area of the brain during working
memory (Koshino et al., 2005; Luna et al., 2002) and
sentence comprehension (Just et al., 2004; Müller et al.,
1998) tasks. This does not indicate a general impairment in
this area, as Müller et al. (1998) found a higher activation of
the DLPFC in an autistic group compared to a control group
when repeating sentences.
Brain imaging can also be used to measure connectivity
of brain regions by calculating the temporal correlation of
activation. The pattern of results in the current limited
functional connectivity MRI (fcMRI) literature of autism
suggests that functional connectivity between subcortical
nuclei and cerebral cortex tends to be increased whereas
cortico-cortical functional connectivity tends to be reduced.
In particular, Turner et al. (2006) found a pattern of greater
connectivity in an autism group from the subcortical caudate
area of the brain to areas across frontal, parietal, and
occipital lobes. Between cortical areas, underconnectivity
has been found from DLPFC to parietal and fusiform gyrus
regions of the brain (Just et al. 2004; Koshino, 2007).
In summary, autism researchers have found no
impairments for simple skills (including simple memory),
but have found impairments in contextual memory and
corresponding lower activation in the DLPFC. A preference
for visual representation has been found with corresponding
activation in the parietal area of the brain. These results can
be contrasted with those from individuals with Attention
Deficit Hyperactivity Disorder (ADHD), which show
impairment for inhibition in tasks such as Stroop and
corresponding lower activation in the anterior cingulate
cortex (ACC) area of the brain (Bush et al., 1999)
Many theories have been proposed for the information
processing profile found in people with autism. Most of
these theories are expressed with high level constructs that
leave gaps in the details of processing. Computational
models offer more detail, but the few computational models
of autism that have been developed have been limited to a
small number of tasks (Cohen, 1998; Kriete & Noelle, 2005;
McClelland, 2000; O’Laughlin & Thagard, 2000). This
paper proposes the first step of an account of autism
findings with a representation of information processing in
autism based on the ACT-R cognitive architecture
(Anderson et al., 2004). Cognitive architectures provide a
detailed explanation of processing from perception to
cognition to motor activity, and psychological research in
one area is captured in the architecture for other projects to
use.
2168

Figure 1: An illustration of the locations of the regions of interest. The Goal (7) and Procedural (8) areas
are deeper in the brain, with Procedural being sub-cortical.

High Level Constructs
Some of the high level constructs used to explain autistic
behavior include central coherence (Frith, 1989),
executive function (Pennington & Ozonoff, 1996),
working memory (Russell, Jarrold, & Henry, 1996),
complexity (Minshew & Goldstein, 1998), and
underconnectivity (Just et al., 2004). These constructs
have been useful in trying to understand autism, but
relying on them without looking at low level processing
can lead to conflicting results and gaps in understanding
of autistic behavior. For example, Edgin and Pennington
(2005) noted that executive function deficits may be less
pervasive in autism than was originally thought. Recent
studies have shown intact performance on executive
function measures in groups with autism when
extraneous task demands have been minimized.
McMahon-Griffith (doctoral dissertation, 2002) has
found that lessening the amount of experimenter
feedback on tasks such as the Wisconsin Card Sorting
Task (WCST) can help improve performance in those
with autism to a level equivalent to typical performance.
Also, other studies (Ozonoff, 1995; Pascualvaca et al.,
1998) have found that the deficits on executive function
measures were greater when administered by humans
than when administered by computer. In their study of
spatial cognition, Edgin and Pennington found no
evidence for impairments in executive function or in
processing global/local information, and they thought this
contradicted theories that rely on these constructs.
Koshino et al. (2007) noted that although several
behavioral studies have found deficits in working
memory in autism (e.g., Bennetto et al. 1996; Minshew et
al. 1997; Luna et al. 2002;), some others have not (e.g.,
Russell et al. 1996; Griffith et al. 1999; Ozonoff and
Strayer 2001). The detailed processing account of
computational models can make task requirements and
processing explicit and help to shed light on these results.
ACT-R
ACT-R (Anderson et al., 2004) is a computational theory of
human cognition incorporating both declarative knowledge
(e.g., addition facts) and procedural knowledge (e.g., rules
for solving multi-column addition) into a production system
where procedural rules act on declarative chunks. In ACT-R
declarative knowledge is represented in structures called
chunks and held in the Declarative module, whereas
procedural knowledge is represented as rules called
productions and held in the Procedural module. Rules also
have access to other modules including the Visual module
for perception, the Manual module for action, the Imaginal
module for holding visual problem representation, and the
Goal module for keeping track of current intentions. These
modules are proposed to occur in specific areas of the brain:
Manual in the motor cortex (BA 3/4), Imaginal in the
parietal cortex (BA 39/40), Declarative in the DLPFC (BA
45/46), Goal in the ACC (BA 24/32), Visual in the fusiform
gyrus (BA 37), and Procedural in the caudate of the basal
ganglia (Figure 1). The ACT-R theory also includes Aural
and Vocal modules but these will not be discussed for ease
of exposition.
The locations of these modules have been supported by a
number of brain imaging studies with tasks such as Tower
of Hanoi (Anderson, Albert, & Fincham, 2005), fan memory
(Sohn et al., 2005), associative memory (Anderson, Qin,
Jung, & Carter, 2007), anticipation of conflict monitoring
(Sohn et al., 2007), algebra (Danker & Anderson, 2007;
Stocco & Anderson, in press), and mental calculation
(Anderson & Qin, in press).
In addition to the symbolic level of facts and rules, ACT-
R includes a subsymbolic level of representation where facts
have an activation attribute which influences their
probability of retrieval and the time it takes to retrieve
them. Rules have a utility attribute which influences their
probability of being used. The activation Ai of a chunk i
is computed from two components – the base-level and a
2169