NIH Public Access

Author Manuscript
Neuropsychologia. Author manuscript; available in PMC 2015 September 01.
Published in final edited form as:
NIH-PA Author Manuscript

Neuropsychologia. 2014 September ; 0: 375–389. doi:10.1016/j.neuropsychologia.2014.04.014.

A Neural Network Model of Individual Differences in Task

Switching Abilities
Seth A. Herd1, Randall C. O’Reilly1, Tom E. Hazy1, Christopher H. Chatham1,2, Angela M.
Brant1,3, and Naomi P. Friedman1,4,*
1Department of Psychology and Neuroscience, University of Colorado Boulder, 345 UCB,
Boulder, CO 80309, USA
4Institute for Behavioral Genetics, University of Colorado Boulder, 447 UCB, Boulder, CO 80309,
USA

Abstract
NIH-PA Author Manuscript

We use a biologically grounded neural network model to investigate the brain mechanisms
underlying individual differences specific to the selection and instantiation of representations that
exert cognitive control in task switching. Existing computational models of task switching do not
focus on individual differences and so cannot explain why task switching abilities are separable
from other executive function (EF) abilities (such as response inhibition). We explore hypotheses
regarding neural mechanisms underlying the “Shifting-Specific” and “Common EF” components
of EF proposed in the Unity/Diversity model (Miyake & Friedman, 2012) and similar components
in related theoretical frameworks. We do so by adapting a well-developed neural network model
of working memory (Prefrontal cortex, Basal ganglia Working Memory or PBWM; Hazy, Frank,
& O’Reilly, 2007) to task switching and the Stroop task, and comparing its behavior on those
tasks under a variety of individual difference manipulations. Results are consistent with the
hypotheses that variation specific to task switching (i.e., Shifting-Specific) may be related to
uncontrolled, automatic persistence of goal representations, whereas variation general to multiple
EFs (i.e., Common EF) may be related to the strength of PFC representations and their effect on
processing in the remainder of the cognitive system. Moreover, increasing signal to noise ratio in
NIH-PA Author Manuscript

PFC, theoretically tied to levels of tonic dopamine and a genetic polymorphism in the COMT
gene, reduced Stroop interference but increased switch costs. This stability-flexibility tradeoff
provides an explanation for why these two EF components sometimes show opposing correlations
with other variables such as attention problems and self-restraint.

© 2014 Elsevier Ltd. All rights reserved.

*
Corresponding Author. Postal Address: Institute for Behavioral Genetics, University of Colorado Boulder, 447 UCB, Boulder, CO
80309, USA. [email protected]. Phone: 1-303-735-4457.
2Present address: Department of Cognitive, Linguistic, & Psychological Sciences, Brown University, Providence, RI 02906, USA
3Present address: Department of Psychology, Pennsylvania State University, State College, PA 16802, USA
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our
customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of
the resulting proof before it is published in its final citable form. Please note that during the production process errors may be
discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
 Herd et al. Page 2

Keywords
executive control; set shifting; computational model; genetics
NIH-PA Author Manuscript

1. Introduction
Understanding how people switch tasks (e.g., how the brain switches attention between a
conversation and oncoming traffic) has obvious relevance in itself. In addition, detailed
exploration of switch costs (the extra time it takes to perform a task when switching from a
different task) has provided numerous insights into the mechanisms of human executive
function (EF) (see reviews by Kiesel at al., 2010; Monsell, 2003; and Vandierendonck,
Liefooghe, & Verbruggen, 2010). A number of computational models have been proposed
as explanations of the computational and neural mechanisms of task switching. Although
these models have elucidated the sources of many task switching findings, they have tended
to focus on mean effects that occur across subjects, without considering patterns of
individual differences that might shed light on the mechanisms involved. In this paper, we
present a biologically based neural network model of task switching that can explain
patterns of individual differences in terms of specific neural mechanisms.
NIH-PA Author Manuscript

Importantly, this model is integrated with other models of executive tasks and so can
elucidate what factors distinguish vs. unify switching abilities from or with other executive
abilities. We focus on the largely unanswered question of why task switching abilities are
separable from other EF abilities (such as response inhibition and working memory
updating/capacity) in terms of individual differences (e.g., Miyake et al., 2000). Though an
individual’s performance on switch tasks correlates with other measures of EF such as
performance on the antisaccade or Stroop tasks, task-switching scores also capture unique
variance: Different switch tasks correlate more closely with each other than they do with
other EF tasks. This unique variance not only appears to be influenced by separate genes
(Friedman et al., 2008), but it also appears to show some trade offs with general executive
control; those who are better at switching, controlling for other EF abilities, seem to show
more behavioral problems such as more attention and externalizing problems and lower self-
restraint (Miyake & Friedman, 2012). Our model suggests variations in several neural
mechanisms that may explain this separation of and tradeoff between different aspects of
EF.
NIH-PA Author Manuscript

A novel aspect of this study is that we vary many biologically based model parameters over
a broad range to simulate individual differences within the normal range, and in some cases
extending into ranges that would be considered pathological. This approach goes beyond
typical manipulations in the computational literature that involve, for example, lesioning
parts of the model to ascertain their necessity for performing a task or to simulate specific
neurological insult. By focusing on more graded manipulations of key parameters, we
simulate individual differences in model performance that can explain observed patterns of
correlations and anti-correlations in the literature. We focus on simulating qualitative rather
than quantitative patterns. Moreover, we focus on explaining results found with normal
population samples, in which individuals may fall on a spectrum of ability (and disorder),
with the extremes capturing individuals who might meet criteria for a diagnosis. Our model

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 3

suggests variations in several neural mechanisms that may explain this separation of and
tradeoff between different aspects of EF. First, however, we situate our model in relation to
previous models and theories of task switching.
NIH-PA Author Manuscript

1.1. Previous Models of Task Switching

Theoretical and computational models of task switching have focused on two possible
sources of switch costs (Kiesel et al., 2010; Vandierendonck et al., 2010). While these
possibilities were considered competing explanations in early work, more recent theoretical
work holds that both are likely true (Vandierendonck et al., 2010). The first possibility is
that switch costs reflect time needed to resolve interference from prior conflicting task sets
(rules for mapping stimuli to responses on a given task, often thought to be instantiated
neurally as representations of continuously firing neurons in prefrontal cortex or PFC). This
task-set interference explanation includes, in some formulations, activation triggered by
previous stimulus-task associations (i.e., “task-set inertia”; Allport, Styles, & Hsieh, 1994).
The second possibility is that switch costs reflect a time cost for active reconfiguration of
task sets (e.g., Rogers & Monsell, 1995) arising, in some accounts, from retrieving the task
set from long-term memory (Altmann & Gray, 2008, Logan & Schneider, 2010). Task sets
are representations guiding performance of a particular stimulus-response mapping when
NIH-PA Author Manuscript

stimuli have been linked to many possible responses. They likely include rules about
categorization of particular stimuli, response mappings, attention orientation, and response
threshold (Vandierendonck et al., 2010). Mechanistically detailed models of task switching
and executive control usually hold that these task set representations consist of persistent
neural firing in PFC (e.g., Cohen, Dunbar, & McClelland 1990; Herd, Banich, & O’Reilly,
2006; Reynolds, Braver, Brown, & Van der Stigchel, 2006). This mechanism is based on
abundant empirical evidence from monkey electrophysiology in working memory tasks, and
human neuroimaging during task switching and other EF tasks. Most abstract mathematical
models are also consistent with this hypothesis (Logan & Gordon, 2001; Meiran, Kessler, &
Adi-Japha, 2008; Sohn & Anderson, 2001).

Although both explanations of switch costs are likely correct in part, Vandierendonck et al.
(2010) point out that an integrated view has not been implemented in a computational model
(although Brown, Reynolds, & Braver, 2007, have come close with a model that
incorporates a task-set representation system and interference control). Here we build on
NIH-PA Author Manuscript

prior work by modeling task switching with a biologically based model of working memory.
This model produces switch costs through both task set reconfiguration (i.e., updating the
contents of working memory with new task set information based on a cue) and interference
resolution (through top-down biasing and competitive inhibition). In this account, the two
sources of costs are inextricably intertwined (at the level of task set representations;
interference in stimulus-response mappings is separable, as we discuss later): Reconfiguring
a task set takes time in part because of interference with the previous task set.

The current model is consistent with previously proposed proximal neural mechanisms of
cognitive control, as explored in some detail in neural network models (Cohen, et al., 1990;
Herd et al., 2006). However, we further explore their origins and interactions with other
brain mechanisms crucial to making cognitive control flexible and appropriate to each

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 4

situation. These earlier models show how a task set representation, in the form of persistent
neural firing in prefrontal cortex (PFC), influences the neural processing that happens in
other brain regions. A variety of common neural learning mechanisms (both error-driven
NIH-PA Author Manuscript

and associative in nature) can create strong connections between neurons in the prefrontal
task set and those in motor and parietal areas, thereby causing them to drive those neurons
that carry out the correct mappings necessary for that task whenever the task set is active.

Crucially, those models consider task sets to be a kind of working memory. By using a well-
developed model of working memory and training it to perform task switching, we follow
and test this hypothesis. However, most of those previous neural network task switching
models simply assumed that a task set representation is appropriately maintained during the
performance of a task – that is, a specific set of prefrontal neurons remains active while a
given task is performed, and that set switches when the task to be performed switches. Only
a few models have included realistic mechanisms that learn task sets, (e.g., Collins & Frank,
2013; Rougier, Noelle, Braver, Cohen, & O’Reilly, 2005), and these studies do not
investigate resulting switch costs in detail. Here, we explore a set of brain mechanisms that
could appropriately perform that type of switching, including learning task set
representations, and how to switch them appropriately. The inclusion of these mechanisms
NIH-PA Author Manuscript

gives our model enough depth to explain not only mean effects across the population, but
individual differences in task performance.

1.2. Individual Differences in Task Switching are Separable From Other Cognitive Control
Abilities
Previous work has repeatedly demonstrated that EFs can be characterized as a family of
related but separable abilities, both at behavioral and neural levels (e.g., Collette et al., 2005;
Friedman et al., 2006; Hedden & Yoon, 2006; Lehto, Juuja rvi, Kooistra, & Pulkkinen,
2003; Miyake et al., 2000). That is, these abilities show unity and diversity (Teuber, 1972;
Miyake et al., 2000). Here we explore some possible mechanistic underpinnings of two of
these factors, one common to many or most EF tasks, and one specific to switching tasks.
These two factors appear to be somewhat anti-correlated, so that those better at EF in
general show a greater RT difference between switch and repeat trials in two-task switching
paradigms than those who are weaker on general EF (although they still tend to be faster in
both conditions in absolute terms). As described below, the evidence for this hypothesis is
NIH-PA Author Manuscript

indirect but strong.

Friedman et al. (2008) reported a large twin study of three commonly examined EFs,
measured with latent variables based on nine tasks: response inhibition (Inhibiting:
antisaccade, Stroop, and stop-signal tasks); working memory updating (Updating: n-back,
letter memory, and keep-track tasks); and task switching (Shifting: color-shape, number-
letter, and category switch tasks). Consistent with prior literature, they found that these three
factors were correlated but separable. Importantly, results of latent variable twin modeling
indicated that this unity and diversity was primarily genetic in origin. Specifically, these
three abilities were all influenced by a highly heritable common factor, but the Updating and
Shifting components were also influenced by separate genetic factors. Friedman et al.
reconfigured their model to capture the variance common to all the tasks (Common EF) and

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 5

specific to updating and shifting tasks, Updating-Specific and Shifting-Specific,

respectively. There was no specific variance for the Inhibiting component, suggesting that at
the level of individual differences, the common factor captured all the variance in inhibiting
NIH-PA Author Manuscript

abilities, a finding that has since been replicated in other independent datasets (see Miyake
& Friedman, 2012). Their resulting Unity/Diversity model is shown in Figure 1, in which
performance on shifting tasks are influenced by both the Common EF factor and the
Shifting-Specific factor.

Reparsing the correlated components of Inhibiting, Updating, and Shifting into these
orthogonal components of Common EF, Updating-Specific, and Shifting-Specific has
resulted in the important insight that that some variables tend to show opposite correlations
with the Common EF factor and Shifting-Specific factors. Table 1 provides a summary of
these results for the Longitudinal Twin Sample, taken from published studies or re-analyses
of published studies. The table also includes genetic correlations when available, which are
the correlations between just the genetic variance in each variable (environmental
correlations were generally not significant).

As this table shows, both cognitive and non-cognitive variables show a trade-off in their
NIH-PA Author Manuscript

correlations with Common EF and Shifting-Specific abilities; however, this pattern only
emerges when shifting ability is separated into the Common EF and Shifting-Specific
components. For example, Friedman, Miyake, Robinson, and Hewitt (2011) reported that
full-scale Wechsler Adult Intelligence Test (WAIS IQ; Wechsler, 1997) was positively
correlated with Inhibiting, Updating, and Shifting abilities, but the correlation with Shifting
was quite small (r = .16). However, when the same data were analyzed with the Unity/
Diversity model, IQ positively correlated with Common EF and Updating-Specific, but
significantly negatively correlated with Shifting-Specific. These results indicate the low
correlation with Shifting was due to the fact that IQ is positively related to the Common EF
part of Shifting, but negatively related to Shifting-Specific. Hence, the total correlation with
Shifting, which is a combination of these two factors, was low. This same trade-off pattern
can be seen in several non-cognitive phenotypes related to externalizing disorders and
behavior problems, measured at multiple earlier times in the longitudinal sample, shown in
Table 1. These measures include teacher-rated attention problems (Friedman et al., 2007),
Behavioral Disinhibition at two ages (a latent variable that captures externalizing behaviors
including ADHD, substance use, conduct disorder, and the personality dimension of novelty
NIH-PA Author Manuscript

seeking; Young et al, 2009), and even very early self-restraint, measured at ages 14 to 26
months (Friedman et al., 2011).1

Taken together, these data suggest that there are opposing mechanisms that tend to make
some variables, particularly behavioral problems, show negative correlations with Common
EF but positive correlations with Shifting-Specific. Although these two components are by

1It is important to note that we do not always find this strong pattern of trade-offs. For example, Friedman et al. (2009) found that the
slope of growth in sleep problems from ages 4 to 16 years was negatively correlated with Common EF (r = −.29), but was not
significantly correlated with Shifting-Specific (r = .12). Similarly, in unpublished data, we have found the number of past week
depression symptoms, as measured by the Center for Epidemiological Scales – Depression (CES-D; Radloff, 1977), showed a
significant negative correlation (r = −.28 phenotypic and r = −.38 genetic) with the Common EF factor but did not significantly
correlate with Shifting-Specific (r = .07 phenotypic and r = .07 genetic).

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 6

necessity uncorrelated in the Unity/Diversity model (because Shifting-Specific is essentially

a residual of Common EF), their trade-off becomes apparent when a third variable is added
to the model. These patterns have led us to postulate some mechanisms that might underlie
NIH-PA Author Manuscript

these factors and show such a trade-off. In this paper, we show that manipulations of these
mechanisms within the models do indeed result in opposite patterns of results for an
inhibition task (a four-response, manual variant of the common Stroop (1935) color-naming
vs. word-reading task) vs. a switching task.

Specifically, we propose that individual differences in the Common EF factor in large part
reflect the ability to actively maintain goals and goal-related information, often in the face of
interference (Friedman et al., 2008), and to use these goals to bias ongoing processing (i.e.,
top-down attention and task control). This basic ability, which has been frequently posited as
a key element of executive control and frontal lobe functioning (e.g., Engle, Kane, &
Tuholski, 1999; Miller & Cohen, 2001), is necessary for all types of EF tasks. It may also be
particularly influential in response inhibition tasks and other tasks that involve conflict
(which many of our EF tasks do), a view consistent with views of inhibitory control as a by-
product of goal maintenance (Chatham et al., 2012; Herd et al 2006; Kane & Engle, 2003;
Morton & Munakata, 2002; Munakata et al., 2011). This strong influence of goal
NIH-PA Author Manuscript

maintenance on response inhibition tasks may explain why Friedman et al. (2008) found no
inhibiting-specific variance in addition to Common EF explaining the correlations between
the response inhibition tasks. Within the computational model, we show how individual
differences in this ability are affected by two example manipulations that affect the influence
of maintained PFC task representations on posterior processing: the gain parameter of the
units in the PFC layer, and the strength of the connection between the PFC layer and
posterior hidden layers.

In contrast, we hypothesize that the Shifting-Specific component reflects flexibility—ease of

transitioning to new task set representations. To put it differently, poor performance on the
Shifting-Specific component may reflect “stickiness” (Altamirano, Miyake, & Whitmer,
2010; Blackwell, Chatham, Wiseheart & Munakata, this issue), a tendency for goal
representations to stick around after they are no longer relevant (i.e., in the absence of their
continued active maintenance). In our model, this ability primarily involves the replacement
of no-longer-active goals. We have explored two mechanisms that influence individual
differences in this trait: recurrent connection strength in PFC, and the degree of clearing of
NIH-PA Author Manuscript

old representations when gating occurs (clearing on gating).

1.3. PBWM Framework and Model Predictions

Our starting point for the modeling work was a well-developed neural network model of
working memory: PBWM (Prefrontal cortex, Basal ganglia Working Memory (Frank,
Loughry, & O’Reilly, 2001; Frank, Seeberger, & O’Reilly, 2004; Hazy, Frank, & O’Reilly,
2006, 2007; O’Reilly, 2006; O’Reilly & Frank, 2006). In part, our choice of a working
memory model served as a test of the theory that cognitive control is simply one function of
working memory mechanisms (Deco & Rolls, 2005; Engle et al., 1999; O’Reilly, Braver, &
Cohen, 1999). In the view we present here, cognitive control and working memory use
isomorphic brain mechanisms: Cognitive control arises from working memory

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 7

representations of goals. These goal representations exert useful biasing effects that give rise
to attention (Desimone & Duncan, 1995) and flexible stimulus-response mappings (Cohen et
al., 1990, Herd et al., 2006).
NIH-PA Author Manuscript

The PBWM series of models is based on a wealth of physiological data, and incorporates a
good deal of physiological detail. We used this detailed model of working memory to create
specific hypotheses about the brain mechanisms underlying different aspects of EF, and to
explore the degree to which a model intended to explain working memory could explain EF
more generally. We propose links from these physiological hypotheses to the level of
psychological constructs (Figure 2). Specifically, we examined the different mechanisms
called upon to rapidly switch between tasks versus those needed to maintain strong
representations that support all EF tasks. We discuss these distinctions primarily in the
context of the Unity/Diversity model of EF (Friedman et al., 2008; Miyake & Friedman,
2012); however, this idea of a flexibility/stability tradeoff in goal-directed behavior has been
discussed by a number of researchers (e.g., Bilder, 2012; Cools, 2012; Goschke, 2000;
Tunbridge, Harrison, & Weinberger, 2006).

The Common EF versus Shifting-Specific constructs map relatively well to particular

NIH-PA Author Manuscript

components of the PBWM model (although that mapping is not perfectly one-to-one, as
addressed in the discussion). PBWM focuses on the need to selectively and appropriately
“gate” information into working memory. Persistent neural firing in PFC over delay periods
has been shown to reflect relevant items, whereas distracting items cause brief but not
persistent activity (Miller, 2000). Some mechanism must appropriately select which
information is maintained. The PBWM framework proposes that such selection is performed
by a complex interplay between the PFC and the basal ganglia (BG), directly analogous to
the well-studied interaction between sensorimotor cortex and BG in selecting motor actions.

While the BG provides the “store or don’t” signal, the PFC is responsible for actually
maintaining the selected information. The BG input dramatically strengthens maintenance,
but the PFC by itself also has an intrinsic ability (and therefore tendency) to maintain any
and all representations it forms for a short period. While this automatic maintenance or
“stickiness” of representations in PFC is helpful in many cases, these intrinsic cortical
mechanisms also slow down the switch from one representation to the next. In the case of
task switching, this slowed transition can slow task performance. Thus, the several neural
NIH-PA Author Manuscript

factors that contribute to a tendency to maintain information in PFC (regardless of the gating
decision from BG) should contribute to worse switching performance. Neural factors that
are controlled by BG gating should not, on the other hand, slow down switching, as such
gating happens rapidly. Such factors should contribute to the Common EF component. One
exception is factors that are triggered by gating, but serve to actively clear out old
representations; these should also contribute to the Shifting-Specific component (see Figure
2).

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 8

2. Methods
2.1. General Model Architecture
NIH-PA Author Manuscript

Our model, shown in Figure 3, was constructed based on the Leabra neural network
modeling framework (O’Reilly, 1996; O’Reilly, Hazy & Herd, in press; O’Reilly &
Munakata, 2000). This framework fulfills some of the criteria laid out by Newell (1990) for
a cumulative model that explains many phenomena, and so constitutes a more constrained
theory. The Leabra framework has been used to model many different psychological and
neurophysiological findings with the same basic set of assumptions and most of the same
parameter values (see O’Reilly et al., in press, for a recent review). This model also gained
additional constraints in using parameters and network architecture to model not only one
task, as in most neural network work, but all nine of the tasks used by Friedman et al (2008)
in a closely related modeling project (Friedman et al., 2014; see also Chatham et al., 2011).2

The Leabra algorithm is described in greater detail in print (O’Reilly & Munakata, 2000)
and actively updated online textbooks (O’Reilly et al., 2012), and its purpose and tenets as a
cumulative theory are elaborated in O’Reilly et al. (in press); its more relevant general
characteristics are described here. All of these attributes are common choices in artificial
NIH-PA Author Manuscript

neural network models, except as noted. Point neurons are used, with no explicit modeling
of their spatial extent. Input from other units is summed and translated to an average firing
rate (a “rate coded” approach). The effect of a unit on each of its afferents is determined by
the product of its firing rate and the synaptic strength connecting the two. All such weights
in the current model were “learned” in the course of training, by adapting their strength
through a combination of error-driven and associative learning, a learning rule unique to
Leabra. The error-driven component is based on phasic differences between an early “guess”
or “minus” phase, and a late “correct answer” or “plus” phase; information from the correct
answer affects the network through feedback synaptic top-down projections (which are
identical in function to the excitatory feedforward projections). The model performs learning
with a similar overall effect to backpropagation learning, but calculates weight changes by
comparing differences in local synaptic activity, rather than the non-local and therefore
biologically implausible operation employed in backpropagation learning. The error-driven
component of the learning essentially drives the network to create states from inputs that are
similar to those that occur in the presence of the appropriate output. The source of the
NIH-PA Author Manuscript

correct answers needed for this type of learning can, in many cases, be conceptualized as
arriving from later sensory input or from other brain areas; an actual external teacher is thus
beneficial but not necessary for this type of learning. In the current training, we assume that
reliable feedback on desired answers is available in some form.

The PBWM model in particular is based on extensive lesion and electrophysiological

evidence from those brain areas involved in reward-motivated instrumental and working
memory tasks in animals, and it is constrained by human behavioral and neuroimaging

2That project focuses on using the same model to capture individual differences in all nine of the tasks used by Friedman et al. (2008).
It does not include discussions of benchmark findings but rather focuses on understanding the mechanisms that may lead multiple
tasks within a construct to intercorrelate. Although there is some discussion of the parameter manipulations used here, it is in the
broader context of understanding patterns seen across all nine tasks, including updating tasks.

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 9

studies on more complex working memory tasks in humans. The extensive work on the role
of the BG and cortex in motor learning is leveraged to explain the brain mechanisms
underlying working memory, because the anatomical relationships between these systems
NIH-PA Author Manuscript

are highly similar in the areas known to be responsible for motor learning and working
memory. In essence, the PFC creates representations of candidate actions, task sets, or
memory items, and the BG decides which of these representations are “gated” to be
performed or remembered. The PFC learns these representations through error-driven
mechanisms, while the BG learns the final gating decisions based on predictions of reward.
Those reward predictions are made by an array of subcortical systems working in concert
(Hazy, Frank, & O’Reilly, 2010; O’Reilly, Frank, Hazy, & Watz, 2007). These systems
produce dopamine release when new predictors of reward appear in the environment, and
dips when expected rewards do not occur. These phasic changes in dopamine concentration
in turn drive learning in the striatum (labeled Matrix in model illustrations for the relevant
subset of striatal neurons) to perform actions and maintain memory items that predict
reward. The details and further explanation of the computations of the PBWM model can be
found in the relevant publications (Frank, et al., 2001, 2004; Hazy, et al., 2006, 2007;
O’Reilly 2006; O’Reilly & Frank, 2006).
NIH-PA Author Manuscript

The model is highly similar but not identical to previously published versions, because the
framework is still being actively developed. This model was developed from an off-the-shelf
current version of the model. Parameters were changed only to put the model in a more
sensitive parameter regime to allow investigation of individual differences, since the
development version was optimized for the fastest possible learning on a small set of
benchmark tasks. In setting the default levels of various parameters, we selected
intermediate values so that performance would not be optimized, but could increase or
decrease, thus improving our chances for obtaining individual differences with
manipulations. In particular, we used default values for the parameters we manipulated that
would produce performance comparable to average human performance, and not at ceiling
or floor (e.g., gain parameters for the neural response functions in PFC and hidden layers
were reduced from 600 to 100; the clear-on-gating parameter was reduced from 1 to 0.5).

The other important difference between this and other PBWM models is the inclusion of two
hidden layers, and the persistence of neural activity in those layers between trials. These
layers produce an attractor state that gives the model an increased tendency to make the
NIH-PA Author Manuscript

same sensory-motor mappings that it did on the previous trials. This change captures the
persistence of representations in premotor cortex, higher sensory areas, and parietal cortex
that perform stimulus-response mappings (e.g., Curtis & Lee, 2011).

2.2. Task Details

The switch and Stroop tasks are implemented in the same model, so their architectures and
parameters are identical. We determined a training schedule for each task that would lead to
performance approximating average accuracy obtained in a large normal population sample
(Friedman et al., 2008; 2011) at default parameter settings. The same training schedules
were used for testing the benchmark findings (all tested with default parameter values) and
for testing individual difference manipulations (see Section 3.2.1).

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 10

2.2.1. Switch Paradigm—The results we report here were primarily based on a color-
naming/shape-naming task switching paradigm in which two classes of stimuli and manual
responses are used for each task (Miyake, Emerson, Padilla, & Ahn, 2004). In the first
NIH-PA Author Manuscript

model settling trial (cue naming), the task cue was presented, and the model was trained to
echo the task name in a verbal output layer (“color” or “shape”). In the immediately
following task performance trial, the task cue was not presented so that the network would
rely on the working memory capacity of PFC to maintain that information. Because activity
in the hidden layers (roughly corresponding to parietal cortex and other areas within the
stimulus-response arc) was decayed by reducing activations by half between trials, adequate
information could be recovered from the representations in the hidden layers to perform the
task. However, the model consistently learned to use the gating mechanisms of PFC to
maintain this information as well, and the significant effects of the manipulations of
parameters in PFC provide reassurance that it plays an important role in task switching in
the model, as it is known to do in people. The model was trained to produce the response
appropriate to the cued task on the manual output layer (assuming a manual response on a
button-box), and to produce the name of the cued task once again on the verbal layer (to
scaffold the creation of a representation of the previously cued task). Importantly, the
manual output responses are shared across tasks (e.g., the same output unit is used to
NIH-PA Author Manuscript

indicate “red” and “circle,” depending on the task, making these manual responses bivalent).
Each behavioral trial consisted of two model trials.

The use of pre-cuing vs. presenting both task cue and stimuli at once was crucial; unlike
humans, neural network models tend to learn such tasks as a flat, conjunctive mapping of
task cue plus stimulus to output when both cue and stimulus are presented simultaneously.
This type of learning produces fundamentally different kinds of switch costs, since tasks are
not explicitly represented. Including a small working memory requirement produces a better
model of human behavior. Training the network to name the task as well as to respond on
the performance trial did not appear to be critical, but did allow the model to learn faster and
more reliably. This procedure matches human behavior under the hypothesis that cue
processing and task set preparation are separate operations from actual task performance.

In this way, our model is unlike typical behavioral experiments, in which only a single
response time (RT) is available per trial; in the model, time preparing to perform the task is
tracked separately from time actually performing the task. Arrington, Logan, and Schneider
NIH-PA Author Manuscript

(2007) used a very similar design to separate cue and task switch costs in a behavioral
experiment, operating under the same assumption that cue and target processing are largely
serial (cf., Logan & Bundesen, 2003). They found switch effects in both the cue and target
phases, and reported that responding to the cue did not disrupt normal processing. Hence,
the version of the task used for this model would seem to generalize to more typical versions
in which only one response is required.

The model was trained using shaping — a technique widely employed in the training of
primates (e.g., Skinner, 1938; see also Krueger & Dayan, 2009, for application to
computational models). It was first trained for twenty epochs on switching at random: Each
task appeared equally often, with a 50% chance of switching tasks on each trial. This
frequent switching training was needed to establish basic performance. Next the model was

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 11

trained for 100 epochs, still performing each task equally often, but with only a 20% chance
of switching tasks on each trial, and without the provision of a task cue on repeat trials. The
cueing phase was eliminated on repeat trials, although the model still had to name the
NIH-PA Author Manuscript

appropriate task for every trial, using memory traces, and of course needed to maintain that
information to provide the correct manual response on conflict trials. The model was cued to
the new task with every switch. This style of training emulated the more standard experience
people have with task switching: Task sets tend to remain relevant for some time until a
signal in the environment indicates that a different task set is needed. This training built an
expectation for repeat trials. (While this style of training was not necessary for producing
switch costs, it did make the costs larger, and may be more ecologically valid.) Finally, the
model was trained for three epochs in the original 50% switch condition, but with the task
cue presented during the response portion of the trial as well as the cue period (emulating
practice trials at the laboratory task). The model was then tested for ten epochs without
learning (no change in any synaptic weights). All reported results are from those test epochs.

RTs were modeled as the number of cycles it took for an output unit to reach a threshold of
0.5 activation. Only RTs for trials that ended with correct responses were analyzed
(additionally, we eliminated trials following errors, as the correct task set may not have been
NIH-PA Author Manuscript

achieved on the prior trial. This elimination of trials following errors is common in
analyzing switch task data; e.g., Friedman et al., 2008).

2.2.2. Stroop Paradigm—We also modeled the Stroop task as an example of the
response inhibition tasks found to load only on the Common EF component (Miyake &
Friedman, 2012). The Stroop paradigm was nearly identical to that for the switch task (with
two tasks: word reading and color naming), but training involved a higher probability of the
word-reading task vs. the color-naming task to build a prepotent response to read the words.
Four responses and stimulus variations per dimension were used. The Shape layer was used
for word-reading inputs and the Color layer for color-naming inputs; words were locally
coded (a single unit for each) in the input layer in the same manner as color and shape, to
provide a simple mock-up of the input from sensory processing areas. For the Stroop task,
each model was trained for 20 epochs of basic training, with equal numbers of word-reading
and color-naming trials. In this phase, all trials were “neutral” (inputs from the irrelevant
modality were neither conflicting nor congruent, as they had no trained response mapping).
Next, the models were trained for another 40 epochs with asymmetric training, such that
NIH-PA Author Manuscript

33% of trials were color naming. In this phase, there were 40% neutral trials, 40% congruent
trials, and 20% incongruent trials. The model was then trained for 87 epochs (a number
chosen arbitrarily) on 80% word-reading trials, with no cue given for repeated trials. This
training forced the model to encode word reading in the hidden layers, with PFC task sets
engaged primarily for the rare color naming trials. This phase also used 40% neutral, 40%
congruent, and 20% incongruent trials for both tasks. This training was followed by three
epochs of training on color naming alone (simulating practice on the laboratory color-
naming Stroop task) with equal numbers of neutral, congruent, and incongruent trials. The
model was then tested on color naming in this same condition, without further learning.

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 12

3. Results
3. 1. General Task Switching Results (Benchmark Findings)
NIH-PA Author Manuscript

In this section we show that our model reproduces several key benchmark findings
(Vandierendonck et al., 2010), based on 100 runs. Although the model did not permit
investigation of every task switching effect in the literature (e.g., the “backward inhibition”
effect observed in tasks that require switching between three subtasks; Mayr & Keele, 2000;
or cue switch vs. task switch costs observed when multiple task cues are used), it does
permit investigation of most findings. Specifically, we examined whether the model showed
switch costs, interference costs, mixing costs, and, because our model can perform both
Stroop and switch tasks, asymmetric switch costs. These results were obtained using default
values of all parameters, in contrast to the individual difference manipulations discussed in
Section 3.2.

3.1.1. Switch Costs—First and most importantly, our model shows a switch cost.
Accuracy was near ceiling, so there was no significant switch cost in accuracy for the cue
naming trials (switch 98.6% vs. repeat 99.5%, t(99) = −1.44, p = .152) or task performance
trials (switch 98.7% vs. repeat 99.6%, t(99) = −1.33, p = .186), although the effects were in
NIH-PA Author Manuscript

the right direction.3 However, there were robust costs in the RTs. The total switch cost (sum
of the cue naming and task performance RTs) was 2.28 cycles (SE = .065; t(99) = 34.99, p
< .001). Cost was significant for both the time to name to the cue in switch vs. repeat trials
(M = 2.11 cycles, SE = .056; t(99) = 37.41, p < .001), and the time to categorize the stimulus
after responding to the cue, (M = 0.17 cycles, SE = .040; t(99) = 4.21, p < .001). The split of
the trials into cue naming and task performance can be considered analogous to
administering trials with long cue-to-stimulus intervals that allow the subject to prepare: In
this situation, the switch cost for the task performance trial can be considered analogous to
the “residual” switch cost that is observed even when subjects have ample time to process
the task cue. The fact that the switch cost for task performance trials was smaller than the
total cost is consistent with the typical finding of a preparation benefit; the fact that the task
performance trials still showed a significant switch cost despite preparation is consistent
with the phenomenon of residual switch cost (Kiesel et al., 2010, Monsell, 2003;
Vandierendonck et al., 2010).

Our residual switch cost is quite small in proportion to total RTs for the model (<10% of the
NIH-PA Author Manuscript

total switch cost). Typically, residual switch costs are larger (e.g., in our own research we
have found residual switch costs approximately 25% to 40% the magnitude of regular switch
costs; Friedman & Miyake, 2004). The small size of the residual switch cost in our model
might be interpreted as providing some support for each of the two competing theories about
residual switch costs. On the one hand, the small residual switch costs might be considered
consistent with theories in which residual switch costs largely reflect a “failure to engage”
on a small proportion of the trials (De Jong, 2000). The scope of our model does not extend
to simulating such occasional failures to prepare (c.f. Reynolds, et al., 2006); the model

3Switch costs in accuracy are often present in behavioral data. In our own data (unpublished), we have found that small but significant
mean switch costs in accuracy are not reliable in terms of individual differences, most likely because of restricted range due to near-
ceiling performance. Because most of the results that we aim to model are based on RT data, we focus on those effects.

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 13

cannot fail to prepare by focusing on task-irrelevant stimuli or thoughts, as humans do.

Thus, we are capturing less than half the full magnitude of typical residual costs, possibly
because our model never fails to engage. On the other hand, the fact that we still find some
NIH-PA Author Manuscript

significant residual switch cost (though very small), despite no failures to engage, suggests
that interference effects may also contribute to this residual cost. Thus, it seems plausible
that in human performance, residual switch costs could be due to both interference and
failure to engage, consistent with the proposed theoretical integration of Vandierendonck et
al. (2010).

3.1.2. Mixing Costs—In many switch tasks, although repeat trials are faster than switch
trials in mixed blocks (those that include both tasks), such repeat trials are still slower than
trials in pure-task blocks. This “mixing cost” has been attributed to a number of factors,
including increased interference of tasks in mixed lists and the requirement to maintain
multiple task sets and response mappings (Kiesel et al., 2010). Rubin and Meiran (2005)
found evidence supporting the hypothesis that mixing costs reflect resolution of task
ambiguity induced by bottom-up activation of competing task sets. This ambiguity may be
influenced by the subject’s knowledge about the presence of mixing vs. not, which is usually
given in the task procedure. That is, in single-task blocks, the subject knows the other task
NIH-PA Author Manuscript

will not be relevant, so it may be easier to maintain a stronger task set of the relevant task
than when subject knows the task will be constantly be switching. The opportunity to
practice a consistent set of stimulus-response mappings is likely also a contributing factor
for faster performance when blocks of trials are not mixed.

If practice is a contributing factor, we might expect to see large mixing costs in this
particular paradigm only when the model has experienced learning with probabilities (so the
model could adjust its weights for task probability). Thus, we examined whether the model
would show mixing costs in two different ways to shed light on the mechanisms in the
model. First, we trained the model with three epochs of learning on switch trials (as in the
standard learning paradigm), then we ran a non-learning test on both switch blocks and
single-task blocks. Thus, the model’s weights should be set as if it were always expecting
switch trials. Nevertheless, we did see a small mixing cost when comparing repeat trials in
mixed blocks (RT = 35.61 cycles, SE = 0.149) to trials in single-task blocks (averaged
across color and shape blocks, RT = 35.38 cycles, SE = 0.151), t(99) = 5.49, p < .001.
However, this slightly longer RT (0.23 cycles, SE = 0.041) on repeat trials in mixed blocks
NIH-PA Author Manuscript

was somewhat small compared to typical mix costs (which can be about half the size of
switch costs; e.g., Kray & Lindenberger, 2000), and likely reflects increased residual
activation from the other task set that persisted across multiple trials (i.e., from task sets that
were active two or more trials before).

To examine whether learning task probabilities play a role in the mixing costs, we used the
same procedure to test single task blocks: 3 epochs of learning with the task, then 10 epochs
of non-learning test. That is, the model received 3 epochs of learning with the switch blocks,
then a no-learning test; 3 epochs of learning with the color task before a no-learning test;
and 3 epochs of learning with the shape task before the no-learning test. In this case, we saw
a much larger mixing cost when comparing repeat trials in mixed blocks (RT = 35.61 cycles,
SE = 0.149) to trials in single-task blocks (RT = 33.91 cycles, SE = 0.109), t(99) = 20.79, p

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 14

< .001. This longer RT (1.70 cycles, SE = 0.082) on repeat trials in mixed blocks was more
comparable to typical mix costs. This mixing cost was also comparable to that obtained
when we allowed learning during test trials: Repeat trials in mixed blocks (RT = 34.99
NIH-PA Author Manuscript

cycles, SE = 0.127) were slower than trials in single-task blocks (RT = 33.66 cycles, SE =
0.108; mixing cost = 1.33, SE = 0.058), t(99) = 20.79, p < .001. Thus, it appears that the
mixing cost in this model is due to some small extent on residual activation from competing
task sets, but to a larger extent to consistent learning about stimulus-response mappings.

3.1.3. Response Interference Effects—A robust finding in switch tasks is that

responses to stimuli that afford different responses for each task are slower than those that
afford the same response — an incongruency cost. Moreover, this effect interacts with
switching, such that the incongruency cost is greater for switch than repeat trials (Kiesel et
al., 2010; Vandierendonck et al., 2010). We reproduce this pattern of results with the RTs
for the task performance trials. Specifically, incongruent trials (RT = 21.18 cycles, SE = .
139) were slower than congruent trials (RT = 16.70 cycles, SE = 0.092), F(1, 99) = 1494.80,
p < .001. This effect interacted with switch condition, F(1, 99) = 22.64, p < .001, such that
the incongruency cost was larger for switch trials (M = 4.68 cycles, SE = .125) than for
repeat trials (M = 4.31 cycles, SE = .118). The fact that we find this effect even with a prior
NIH-PA Author Manuscript

cue naming (analogous to a long preparation time) is consistent with the finding that task-
rule congruency is not affected by preparation time (Monsell, 2003, but see also Kiesel et
al., 2010).

Another effect that occurs in switch tasks is a response repetition effect in which there is a
benefit for response repetitions on repeat trials, but there is no response repetition benefit, or
in some studies even a response repetition cost, on switch trials (e.g., Rogers & Monsell,
1995). Our model does not reproduce this finding. Specifically, though the response
repetition benefit on switch trials (M = .02 cycles, SE = .037) was not significant, t(99) =
0.53, p = .595, it was only marginally significant on repeat trials (M = .06 cycles, SE = .
030), t(99) = 1.88, p = .063. The interaction between switch condition and response
repetition was not significant, F(1, 99) = 0.48, p = .489. Hence, although the general pattern
was present, the effect was not statistically significant. It is possible that the presence of the
cue naming trials (similar to long preparation times) may have contributed to this absence of
an effect. Because the task implementation included cue naming trials between each pair of
task performance trials, the carry-over activation from one categorization response trial to
NIH-PA Author Manuscript

the next decayed more than would be the case had the categorization response trials
immediately followed one another. Extending the model to include adaptive feedback
mechanisms in response to conflict or errors (see discussion) might also produce this effect.
Finally, adding the known stronger short-term effects of learning (the well-established Short
Term Potentiation (STP) phenomenon), should also increase this effect by increasing the
degree to which a particular response activates the task set with which it last occurred.

3.1.4. Asymmetric Switch Costs—When individuals have to shift between tasks that
have different strengths or prepotency, they show asymmetric switch costs. For example,
Allport et al. (1994) found that in a Stroop task involving switching between word reading
and color naming, switch costs were, somewhat counter intuitively, larger when switching to

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 15

the dominant word reading task. To reproduce this finding, we tested a model that had been
trained on the Stroop task (to produce a task asymmetry) with 10 epochs of cued switch
trials (50% word-reading, and 50% color-naming; as with other models, it was a test with no
NIH-PA Author Manuscript

learning that was preceded by three epochs of training with the exact task before learning
was turned off). Consistent with prior results, we found that color-naming task performance
trials (M = 15.05 cycles, SE = .058) were slower than word-naming task performance trials
(M = 14.63 cycles, SE = .049), F(1, 99) = 49.58, p < .001, and combining cue naming and
task performance trials, switch trials (M = 30.34 cycles, SE = .105) were slower than repeat
trials (M = 28.52 cycles, SE = .098), F(1, 99) = 938.56, p < .001; However, there was a
significant task by switch condition interaction, F(1, 99) = 17.39, p < .001, such that switch
costs were larger when switching to word-reading (M = 2.06 cycles, SE = .094) than to
color-naming (M = 1.58 cycles SE = .069). Consistent with the arguments of Allport et al.
(1994), this effect seems to be caused by the necessity of a stronger task set for the less
prepotent task; that task set is less easily replaced by the weaker task set needed to support
the more commonly practiced task (the task set representation for the more-practiced task is
weaker because those well-practiced stimulus-response mappings require less support from
the PFC, and the error-driven learning that creates those task sets makes them as strong as
they need to be to produce correct answers). The model’s relationship to other proposed
NIH-PA Author Manuscript

explanations for the effect are considered in Section 4.2.

3.2. Individual Difference Manipulations

The reproduction of the most commonly reported effects of task switching — including
task-switch and residual switch costs, mixing costs, congruency effects, and asymmetric
switch costs — suggests that our model is rich and accurate enough to perform its main
function: addressing the mechanisms underlying the different components of EF abilities. In
this section we report the results of our parameter manipulations designed to simulate
individual differences in the switch task, and also in the comparison Stroop task. We focus
on comparing the effects of manipulating parameters that correspond to individual
differences in EFs, which appear to be largely genetic (Friedman et al., 2008). To explore
the mechanisms underlying the Common EF and Shifting-Specific components, we tested
two candidate parameters for each. We demonstrate dissociations between effects on the
model’s performance in the Stroop task (one of the response inhibition tasks most strongly
dependent on the Common EF component in Miyake & Friedman, 2012), and effects on the
NIH-PA Author Manuscript

switch task we focus on here. First, though, we briefly outline the procedure we used to test
these parameter manipulations.

3.2.1. Procedures to Test Individual Difference Manipulations—We used the

same training procedures described in Section 2.2 for all models, and manipulated
parameters one at a time (i.e., keeping all other parameters at their defaults) from the
beginning of training, since these parameters model a variety of genetic differences
(Friedman et al., 2008). Performance was always assessed during the non-learning test
period of 10 epochs that followed this standardized training. Accuracy remained high in all
models; we were primarily interested in RTs for correct trials, as these are the typical
variables used to assess individual differences in these tasks.

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 16

We ran 25 models per parameter value tested, and each run within a task used a different
random seed for setting initial weights (though the same seeds were used across tasks and
parameter settings to reduce noise in results). In selecting the parameter values to use, we
NIH-PA Author Manuscript

used a range around the default and equated the range across manipulations as much as
possible. The following parameter values were used: PFC to hidden connection strength (.
1, .25, .5, .75, 1; note that a PFC to hidden connection strength of 0 would be equivalent to
no communication between PFC and posterior cortex, so we used a low value of .1 instead
for the lower bound); PFC recurrent connection strength (0, .25, .5, .75, 1); clearing on
gating (0, .25, .5, .75, 1); and PFC gain (35, 50, 75, 100, 150, 200, 300). The gain
manipulation is the gamma of the sigmoid rate-code activation function (O’Reilly et al.,
2012) and has most of its effect on high activation values, but signal-to-noise ratio also
involves inhibition of low values. Thus, when manipulating gain, we also varied the “nvar”
parameter, which is the variance of the Gaussian noise kernel that is convolved with the
activation function; this parameter determines the curvature of the activation function near
the threshold, with larger values allowing more graded responses at the threshold. The
values of nvar were set to 1.2/gain, with a minimum of .005 to keep this value from being
too close to zero at high gains.
NIH-PA Author Manuscript

3.2.2. Results For Parameters Hypothesized to Relate to Common EF Variance

3.2.2.1. Connection Strength From PFC to Posterior (Hidden) Layer: We hypothesized
that if top-down biasing from PFC is crucial for all EF tasks, the connection strength
between PFC and the hidden layers should affect both the switch and Stroop tasks. As
expected, both the switch cost, F(4,120) = 4.66, p = .002, R2 = .13 (linear F(1,120) = 11.51,
p = .001, R2 = .08; quadratic F(1,120) = 6.43, p = .013, R2 = .05), and Stroop interference,
F(4,120) = 3.02, p = .020, R2 = .09 (linear F(1,120) = 7.27, p = .008, R2 = .06; cubic
F(1,120) = 4.20, p = .043, R2 = .03), were significantly reduced by greater connection
strength from PFC to the hidden layers that perform stimulus-response mappings (see Figure
4A). The actual slopes are relatively shallow due to the fact that these scores are difference
scores based on conditions that all showed significant linear decreases in RTs with
increasing connection strength, all F(1,120) > 190.73, p < .001, in addition to nonlinear
effects, as shown in Figure 4B and 4C. This pattern might be expected given that both trial
types (e.g., switch and repeat) were mixed in blocks and so required some task set
maintenance and top-down control even on easier conditions (e.g., repeat trials or congruent
NIH-PA Author Manuscript

trials). This overall trend is consistent with the fact that, in our studies, RTs in all conditions
are faster for individuals high on the Common EF factor (unpublished results). Nevertheless,
the presence of significant effects on the difference scores indicates that the slope of the
effect was significantly steeper for the more difficult conditions (switch trials in Color-
Shape and incongruent trials in Stroop).

3.2.2.2. Signal to Noise in PFC: The second Common EF manipulation was unit gain in
PFC. This parameter determines signal-to-noise ratio in the layer, which is hypothesized to
be affected by tonic extracellular dopamine levels in PFC (Servan-Schreiber, Printz, &
Cohen, 1990). In our models gain was yoked to the closely related parameter nvar, as
described earlier. This manipulation is of particular interest given that the frequently studied
COMT val108/158met polymorphism is hypothesized to influence extracellular dopamine

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 17

levels in frontal cortex, thus affecting signal-to-noise ratio, with the val allele leading to
lower levels (e.g., Tunbridge et al., 2006). The effect of a higher signal-to-noise ratio in PFC
is deeper attractor states, which leads to crisper representations of goals, but also stronger
NIH-PA Author Manuscript

activations that then must be overcome when goals switch; thus, we expected this
manipulation to have opposite effects on Stroop and switch performance. Signal-to-noise
(i.e., gain) significantly influenced switch costs, F(6,168) = 12.44, p < .001, R2 = .31 (linear
F(1,168) = 23.76, p < .001, R2 = .10; quadratic F(1,168) = 23.38, p < .001, R2 = .09; cubic
F(1,168) = 18.90, p < .001, R2 = .08; 5th order F(1,168) = 5.69, p = .018, R2 = .02), and
Stroop interference, F(6,168) = 14.68, p < .001, R2 = .34 (linear F(1,168) = 74.53, p < .001,
R2 = .29; quadratic F(1,168) = 10.04, p = .002, R2 = .04). However, as shown in Figure 5A,
the two costs showed opposite patterns of results, with higher gain resulting in higher switch
costs but lower Stroop interference costs. This opposite pattern for the switch task was not
due to the manipulation having a qualitatively different effect on the individual trials; as
shown in Figure 5B and 5C, the RTs in both tasks both showed significant linear decreases
in response to gain, all F(1,168) > 137.76, p < .001, along with nonlinear effects. However,
the significantly higher switch costs with higher gain is due to the lower effect on switch
trials than repeat trials. In contrast, for the Stroop task, incongruent trials benefitted more
from higher gain, resulting in lower interference costs.
NIH-PA Author Manuscript

3.2.3. Parameters Hypothesized to Relate to Shifting-Specific Variance

3.2.3.1. Recurrent Connectivity in PFC: Interference occurs when no-longer-relevant task
goals stick around to interfere with current goals. Therefore, any factors contributing to
longevity of PFC representations that cannot be switched off by the BG should contribute to
switch-specific effects. We hypothesized that one mechanism that might lead to such
uncontrollable persistence or “stickiness” might be recurrent connectivity in the PFC,
because such connectivity serves to enable representations to self reinforce and thus remain
active in the absence of external input. Such stickiness may be particularly detrimental for
switch tasks because the goals are rapidly changing. Higher recurrent connectivity leads to
continued activation of previous goals/task sets, and that would cause interference on switch
trials.

This hypothesis was confirmed by our manipulation of recurrent connection strength. As

shown in Figure 6A, Color-Shape switch costs showed a significant increase as recurrent
NIH-PA Author Manuscript

connection strength increased, F(4,120) = 6.89, p < .001, R2 = .19 (linear F(1,120) = 21.82,
p < .001, R2 = .15), whereas Stroop interference costs showed a small decrease, F(4,120) =
4.32, p = .003, R2 = .13 (linear F(1,120) = 4.36, p = .039, R2 = .03; quadratic F(1,120) =
10.08, p = .002, R2 = .07). An examination of the RTs in the individual conditions (Figure
6B and 6C) indicated that, though there were significant nonlinear effects on both conditions
of the Color-Shape task, switch RTs showed a significant linear increase, F(1,120) = 18.48,
p < .001, R2 = .12, whereas repeat RTs did not, F(1,120) = 0.27, p = .606, R2 = .00. This
pattern resulted in larger switch costs at high levels of recurrent connectivity. In contrast,
Stroop RTs did not show significant linear decreases in any condition, all F(1,120) < 2.71, p
> .102, although there were significant quadratic effects on both the incongruent and neutral
conditions.

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 18

3.2.3.2. Clearing on Gating: The second manipulation of interference from irrelevant goals
was in a parameter affecting the degree to which representations “decay” once active
maintenance is toggled off. This parameter, called clearing on gating, is the fraction of unit
NIH-PA Author Manuscript

activity that remains after a “go” action by the BG (a decision, in essence, that it is time to
either replace or use the given working memory item). Thus, a value of .25 for clearing on
gating means that each unit in PFC is reduced by 25% of its previous activity on the trial
after the “go” event, corresponding to relatively weak inhibitory activity, while a value of
1.0 means that the activity, and so the representations, are eliminated entirely. We used this
parameter as a blanket manipulation of the varied neural mechanisms, both passive and
active, that might contribute to clearing of working memory representations after they are no
longer needed. The strength of GABAergic inhibition in cortex triggered by a wave of
excitatory activity when a thalamocortical circuit is activated is one such variable (see, e.g.,
Rigas & Castro-Alamancos, 2007), but others mechanisms may contribute to this effect as
well.

Our hypothesis was that higher levels of clearing on gating would lead to lower switch costs,
but should not affect tasks in which switching is not important. This prediction was
confirmed: As shown in Figure 7A, Color-Shape switch costs showed a significant increase
NIH-PA Author Manuscript

as clearing on gating decreased, F(4,120) = 15.78, p < .001, R2 = .34 (linear F(1,120) =
57.24, p < .001, R2 = .31), whereas Stroop interference costs were not significantly affected
by clearing on gating, F(4,120) = 1.64, p = .169, R2 = .05. This effect on switch costs was
driven by the switch cost for cue naming RTs, F(4,120) = 26.46, p < .001, R2 = .47; the
effect on switch costs for task performance RTs was not significant, F(4,120) = 0.99, p = .
417, R2 = .03. An examination of the RTs in the individual conditions (see Figure 7B)
indicated that Color-Shape switch RTs showed a significant linear decrease in RT with
lower clearing on gating, F(1,120) = 16.68, p < .001, R2 = .12, as did the repeat RTs,
F(1,120) = 57.91, p < .001, R2 = .32 (quadratic also significant), but repeat RTs decreased
more than switch RTs, consistent with the idea that less clearing on gating resulted in more
persistent task sets that helped speed re-establishment on repeat trials.

The fact that lower clearing also, on average, speeded responses on switch trials is
counterintuitive. Because much of the time to establish the new task set is taken by
developing activations, switching from an old to new task set representation is actually
faster than developing a new representation from scratch. The old representation in effect
NIH-PA Author Manuscript

provides a small “jump-start” for the new representation. Stroop RTs did not show
significant decreases with lower clearing on gating, all linear F(1,120) < 2.20, p > .072,
though there was a significant quadratic effect on neutral RTs (see Figure 7C). Because the
Stroop RTs did not include cue naming times, they did not benefit from a “jump start” from
previous trials. However, were such naming times to be included, we would expect a similar
jump start for all trial types (since all involve the same goal and no switching), leading to no
effect on the difference scores.

4. Discussion
We presented a biologically based neural network model of task switching that showed
individual differences in two separable components of switch costs — Common EF and

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 19

Shifting-Specific (summarized in Figure 2) — in addition to reproducing many benchmark

findings from the task-switching literature. Specifically, to simulate individual differences
we manipulated four parameters to affect gated active maintenance of task goals and top-
NIH-PA Author Manuscript

down biasing (Common EF), and residual activation of no-longer-relevant goals or goal
“stickiness” (Shifting-Specific). Whereas the top-down biasing manipulation (PFC to
posterior (hidden) layers connection strength) had beneficial influences on both the
inhibition and switch tasks, the strength of goal maintenance (unit gain in PFC, determining
signal-to-noise ratio) had opposite effects on the two tasks. These results are consistent with
a trade-off between stability and flexibility (Goschke, 2000), influenced by individual
differences in cortical signal-to-noise ratio that is modulated by (among other things)
dopamine function (Colzato, Waszak, Nieuwenhuis, Posthuma, & Hommel, 2010). The
manipulations of non-gated maintenance (stickiness) in PFC had significant overall effects
on switch costs, while having either small opposite (recurrent connection strength) or null
(clearing on gating) effects on Stroop interference, as predicted.

4.1. Mapping EF Components to Biological Mechanisms

These results demonstrate that detailed neural network models based closely on known
biology can be useful in distinguishing the neural mechanisms underlying different
NIH-PA Author Manuscript

components of individual differences in EF. Mapping from biological differences to their

effects on composite measures of comparative performance across many tasks is a complex
undertaking. This type of theory development may thus particularly stand to benefit from the
application of detailed computational models that track the interacting effects of many
theoretical commitments, and can reveal unexpected consequences of manipulations. For
instance, we did not fully predict the effects of either PFC gain or recurrent connectivity on
both switch and Stroop tasks, so the model directly informed our perspective on the likely
contributions of such factors to particular EF component abilities.

More specifically, our model suggests some mechanistic underpinnings of the individual
differences in different components of EF abilities. Parameter manipulations in the model
(and therefore, we predict, genetic and learned differences in brains) that contribute to the
uncontrolled (un-gated), automatic persistence of representations are good candidates for the
Shifting-Specific component of EF identified by Friedman et al. (2008). Such neural factors
include the density and strength of recurrent connections within PFC, intrinsic tendency of
NIH-PA Author Manuscript

neurons to fire persistently (hysteresis), and signal-to-noise ratio (gain) as influenced by

dopamine and the COMT genetic variation, among others. This component should also be
controlled by neural factors that affect active clearing of PFC representations with gating (as
captured by the clearing on gating parameter). These factors would be complex, and are less
well understood, since there are no known inhibitory neurons whose function is unique to
gating interactions. Any factors causing cortical inhibitory neurons to fire strongly
specifically when a wave of activity is first gated through the thalamus should also affect the
Shifting-Specific components; however, we are not aware of any such properties that are
currently known.

The Common EF component should be determined by variance in the quality/strength of

gated PFC representations and their effect on processing in the rest of the cognitive system.

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 20

Such mechanisms include the influence of PFC on other brain areas, including density and
strength of connections. They also include the variety of mechanisms determining how
strongly the BG affect cortical processing, including strength of BG connections with
NIH-PA Author Manuscript

thalamus, and those from thalamus to PFC, etc. Importantly, the general tendency for the
BG to gate should also affect the Common EF component, in the sense that incorrectly
gating might lead to goal neglect or performing the wrong task. Our model could not capture
this effect, as its limited training regime consisting of only two tasks trained gating so
effectively that gating was effectively perfect. This is in sharp contrast to everyday human
experience, in which it is important to not attend to and maintain each new experience.
Variances in this tendency can be captured in the PBWM framework, and have been used as
the basis of a theory that, among other variables, decreased tonic dopamine levels in the BG
lead to gating too rarely, causing task neglect and, in extreme cases, attentional deficit
disorder (Frank et al, 2007).

Importantly, these simulations provide a mechanistic account for the surprising tradeoffs
observed in many studies between switching and other components of EF. In particular, the
PFC gain parameter and, to a smaller extent, the recurrent connection parameter, decreased
Stroop interference but increased switch costs, mimicking the pattern of correlations seen
NIH-PA Author Manuscript

with certain variables (Miyake & Friedman, 2012; see also Blackwell et al., this issue).
These manipulations suggest that one source of these tradeoffs is in the strength of goal
maintenance (tied to dopamine function in cortex, among other biological factors), versus
stickiness of representations (affected by the clearing on gating parameter and hypothesized
to be related to GABA), and as such provide testable predictions. These results are
compatible with a recent finding that having more val alleles in the COMT val108/158met
polymorphism, which results in lower tonic DA in PFC, is associated with lower switch
costs (Colzato et al., 2010), despite being also associated with worse executive performance
in tasks such as n-back and the Wisconsin Card Sorting Test (e.g., Egan et al., 2001; Caldú
et al., 2007). In our model the parameter most related to the COMT polymorphism would be
gain in PFC units (hypothesized to affect the Common EF component), and it shows the
same trade-off in terms of Stroop and switch.

In discussing some potential mechanistic underpinnings of dissociable EF abilities, it is

worth clarifying that the mapping is complex and not one-to-one. Each biological
mechanism is likely to contribute to performance in many EF tasks, but to different degrees;
NIH-PA Author Manuscript

as such, the loading of each mechanism on each factor will depend upon the analysis used to
separate components (e.g., whether components are constrained to be orthogonal, and
whether a single common component is included). In addition to the complexity of single
biological factors contributing different amounts of variance to multiple EF components (as
in the effects of gain and recurrent connection strength), it is important to recognize that our
consideration of the Common EF and Shifting-Specific components in no way suggests that
there are precisely two underlying biological factors, or even two closely functionally
related classes of such factors. Instead, each component can be affected by biological
variations with many different computational functions. For example, in this model, the
Shifting-Specific factor seems to be affected not only by uncontrollable persistence of PFC
representations, as captured by our recurrent connectivity manipulation, but also by the

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 21

strength of mechanisms which specifically clear or disrupt those representations when gating
signals the need for them to change, as captured by our clearing-on-gating mechanism.
While these are only two classes of functional mechanisms, it is likely that there are other
NIH-PA Author Manuscript

classes we have not identified.

In a similar vein, we have noted in previous work that while response inhibition, working
memory updating, and task shifting are some of the most commonly discussed EFs, there are
other possible separable EFs (e.g., Friedman & Miyake, 2004; Friedman et al., 2008;
Miyake et al., 2000). Thus, it is important to note here that we do not assume there are
precisely three components of EF, but merely that different EFs tasks tap both common and
distinct mechanisms (the Unity/Diversity model; Miyake & Friedman, 2012). While existing
studies are insufficient to identify all components of executive control across all possible
types of tasks, they are sufficient to demonstrate that there is a switching factor that is
distinct from a common factor, and it is that distinction we address.

4.2. Extensions and Future Directions

Although the current model is rich enough to produce hypotheses regarding several
important mechanisms underlying different components of EF ability, it could be
NIH-PA Author Manuscript

productively expanded in future work to provide a more complete model of task switching,
which would in turn make more specific predictions about that mapping. The current model
begins to address meta-control by capturing some of the mechanisms controlling switching
between different cognitive control representations, but two expansions on different aspects
of meta-control suggest themselves in particular.

The first such expansion on mechanisms of meta-control is an adaptive feedback

mechanism, as in work addressing anterior cingulate cortex (ACC) contributions to EF.
Although our current understanding of the relevant systems is incomplete (for instance, the
role of brainstem neuromodulatory systems in performance-based adaptation of cognitive
control, Krichmar, 2008), adding even a simple mechanism that increases the strength of
maintained goals in response to poor task performance has been shown to improve model fit
to inter-trial dependencies in task performance (Botvinick, Braver, Barch, Carter, & Cohen,
2001; Brown et al., 2007). Although such data has not, to our knowledge, been analyzed in
terms of individual differences, much less a multi-factor model, inclusion of such
mechanisms would improve the fidelity of the model, and so likely provide suggestions of
NIH-PA Author Manuscript

other mechanisms that might contribute to the EF components we address here. For instance,
mechanistic (biological) differences contributing to stronger error-driven goal activation
could be hypothesized to contribute to the Common EF component, since that mechanism
would contribute to success in all tasks. However, these same mechanisms might also
contribute negatively to the Shifting-Specific component, if that feedback enhances
activation of the current goal after an error (or conflict), and thereby makes that goal more
persistent and so more difficult to replace with the next goal on switch trials. Future
modeling work would help establish how specific individual differences in mechanisms
underlying this trial-to-trial task set regulation map to specific EF components.

Another extension to the model that would address a second proposed explanation for
asymmetrical switch costs is persistent inhibition of previous task sets: If the dominant task

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 22

must be inhibited more strongly, persistent inhibition would make it harder to re-activate
(Koch, Gade, Schuch & Philipp, 2010). Support for this explanation comes from studies
showing that switch costs are higher when switching to a task that has been recently
NIH-PA Author Manuscript

switched away from (the “backward inhibition” effect; e.g., Mayr & Keele, 2000). We
regard this explanation as incomplete, since it requires some further mechanism whereby a
neural representation becomes harder to re-activate after it has been inhibited, as the firing
of inhibitory neurons is not persistent across the relevant time scales of several seconds. The
competitive learning effects hypothesized and supported by Norman and colleagues offers
one such explanation; they propose that neurons that are briefly activated, then are
competitively inhibited, weaken their connections (Norman, Newman, & Detre, 2007). This
learning rule is not employed in this particular model, but is fully consistent with it. Another
possible explanation is a learned connection from an incorrect task set to the subthalamic
nucleus (STN), which appears to delay or halt actions (e.g., Frank, 2006). This connection
could be strengthened whenever a task set is not properly reconfigured, and so produces a
wrong answer. This persistent learning would cause a slower RT, but the time course of
synaptic potentiation does not, by itself, explain the strong difference in recently used versus
more distant task sets.
NIH-PA Author Manuscript

Another significant area for enhancement of the model is in including an “outer loop” for
meta-task goal maintenance. The current model switches only between the two trained tasks,
and so has goal representations for only those tasks. As such, it cannot display goal neglect
(e.g., Duncan, Emslie, Williams, Johnson, & Freer, 1996) by switching to a goal irrelevant
to the current task. In contrast, humans have trained on a huge variety of “tasks” over the
lifespan instead of just two, and have goals outside of the current task. Our model does not
show much residual switch costs with long preparation times (the switch costs for actual
performance trials are small or nonexistent depending on parameters), which we take to be
consistent with the hypothesis that infrequent failure to prepare the current task set is one
prominent source of residual switch costs (De Jong, 2000). Such “failure to engage” could in
part reflect goal neglect at the meta-control level — i.e., failing to perform the more abstract
task of attending to and using the task cues to switch goal representations.

Humans may use such “outer loop” goals, like performing task switching, by hierarchical
active maintenance of sub-goals and superordinate-goals simultaneously in PFC (e.g., Badre
& D’Esposito, 2007; Koechlin, Ody & Kouneieher, 2003). These meta-task goals would be
NIH-PA Author Manuscript

represented in the same type of PFC/BG circuit, but in different areas of PFC. In the current
model, this “task set” is represented only implicitly, in the switching rules learned by PFC
and BG in response to a task cue; there is no “switching” goal held in active maintenance. If
such meta-task goals were actively maintained, they would be consistent across sub-tasks.
They would not need frequent switching, and so would not rely on switching-specific
mechanisms; rather they would rely more upon the parameters controlling strength of
maintenance, which we have here identified with the Common EF component. As such, this
aspect of the model might be more accurate for very well practiced task switches, for which
learning can transfer knowledge of switch rules from active maintenance or episodic
memory to cortical synaptic memory as it has in the model. Evidence for this explicit
structure of maintained goals in novel tasks is mixed (Crittenden & Duncan, 2014;
Reynolds, O’Reilly, Cohen & Braver, 2012), perhaps because many factors may influence

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 23

whether goals (whether sub- or superordinate) are maintained, versus retrieved from
episodic memory (e.g., Braver, Gray & Burgess, 2007; Chatham, Frank & Munakata, 2009;
D’Ardenne et al., 2012). For example, task switching might be more likely to rely on the
NIH-PA Author Manuscript

maintenance of both sub- and superordinate control representations, depending on the time
pressure and frequency of switching in the paradigm used. By contrast, retrieval from
episodic memory might be a more prominent feature of tasks requiring less speeded
responses, or less frequent access to meta-control representations. Further empirical and
modeling work will be needed to identify how and when these mechanisms contribute to
meta-control of cognition.

Adding “meta-task-set” and performance-based adaptive mechanisms to the model are only
two examples of how such models might be made more complete and accurate, and so
produce better hypotheses on the brain mechanisms underlying human EF. Given the wealth
of data pertinent to the issue, further work using explicit, implemented models that can span
levels of analysis could be particularly useful in this domain.

4.3. Summary and Conclusion

The current model extends the literature on computational models of task switching to
NIH-PA Author Manuscript

include a model that can account for individual differences in switch costs. Because the
model built on the existing PBWM framework (Frank et al., 2007), it was integrated with
other EF tasks, and we used this integration to compare how parameter manipulations
influenced tasks tapping different components of EF (i.e., Common EF vs. Shifting-Specific;
Miyake & Friedman, 2012).

We confirmed two predictions about the underlying mechanisms for these factors within the
model. First, parameters that increased goal “stickiness” increased switch costs but not
Stroop interference, consistent with our prediction that the Shifting-Specific component of
the Unity/Diversity model may reflect the extent to which no-longer-relevant goals persist in
PFC. Second, parameters that affected the model’s ability to actively maintain goals in the
PFC layer and use those goals to bias processing in other brain areas influenced both the
switch and Stroop tasks, consistent with our hypothesis that this ability is key to Common
EF variation. Importantly, one of the manipulations (PFC unit gain, affecting signal-to-noise
ratio in PFC representations) showed opposite effects on Stroop interference and switch
costs, mirroring stability-flexibility trade-offs evident in individual difference (see Miyake
NIH-PA Author Manuscript

& Friedman, 2012) and molecular genetic literature (e.g., Colzato et al., 2010). This model
thus provides a foundation for future investigation into the biological and genetic
mechanisms underlying variation in executive control.

Acknowledgments
This research was supported by grants MH063207 and MH079485 from the National Institute of Health. We thank
Akira Miyake and Trent Kriete for helpful comments and suggestions on this project, as well as all the members of
the Interdisciplinary Behavioral Science Center on the Determinants of Executive Function and Dysfunction
(NIMH P50 MH079485) for their discussions of and contributions to the ideas presented here. We also thank Iring
Koch for a particularly thorough and helpful signed review that improved the manuscript in many ways.

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 24

References
Allport, DA.; Styles, EA.; Hsieh, S. Shifting intentional set: Exploring the dynamic control of tasks.
NIH-PA Author Manuscript

In: Umilta, C.; Moscovitch, M., editors. Attention and Performance XV. Cambridge, MA: MIT
Press; 1994. p. 421-452.
Altamirano LJ, Miyake A, Whitmer AJ. When mental inflexibility facilitates executive control:
Beneficial side effects of ruminative tendencies on goal maintenance. Psychological Science. 2010;
21:1377–1382. [PubMed: 20798398]
Altmann EM, Gray WD. An integrated model of cognitive control in task switching. Psychological
Review. 2008; 115:602–639. [PubMed: 18729594]
Arrington CM, Logan GD, Schneider DW. Separating cue encoding from target processing in the
explicit task-cuing procedure: Are there “true” task switch effects? Journal of Experimental
Psychology: Learning, Memory, and Cognition. 2007; 33:484–502.
Badre D, D’Esposito M. Functional magnetic resonance imaging evidence for a hierarchical
organization of the prefrontal cortex. Journal of Cognitive Neuroscience. 2007; 19:2082–2099.
[PubMed: 17892391]
Bilder RM. Executive control: balancing stability and flexibility via the duality of evolutionary
neuroanatomical trends. Dialogues In Clinical Neuroscience. 2012; 14:39–47. [PubMed: 22577303]
Blackwell, KA.; Chatham, CH.; Wiseheart, M.; Munakata, Y. A developmental window into trade-offs
in executive function: The case of task switching versus response inhibition in 6-year-olds. (this
issue)Manuscript under review
NIH-PA Author Manuscript

Botvinick MM, Braver TS, Barch DM, Carter CS, Cohen JD. Conflict monitoring and cognitive
control. Psychological Review. 2001; 108:624–652. [PubMed: 11488380]
Braver, TS.; Gray, JR.; Burgess, GC. Explaining the many varieties of working memory variation:
Dual mechanisms of cognitive control. In: Conway, ARA.; Jarrold, C.; Kane, MJ.; Miyake, A.;
Towse, JN., editors. Variation in working memory. New York: Oxford University Press; 2007. p.
76-106.
Brown JW, Reynolds JR, Braver TS. A computational neural model of fractionated conflict-control
mechanisms in task-switching. Cognitive Psychology. 2007; 55:37–85. [PubMed: 17078941]
Caldú X, Vendrell P, Bartres-Faz D, Clemente I, Bargallo N, Jurado MA, Serra-Grabulosa JM, Junqué
C. Impact of the COMT Val108/158 Met and DAT genotypes on prefrontal function in healthy
subjects. Neuroimage. 2007; 37:1437–1444. [PubMed: 17689985]
Chatham CH, Claus ED, Banich MT, Curran T, Kim A, Munakata Y. Cognitive control reflects
context monitoring, not motoric stopping, in response inhibition. PLoS One. 2012; 7:e31546.
[PubMed: 22384038]
Chatham CH, Frank MJ, Munakata Y. Pupillometric and behavioral markers of a developmental shift
in the temporal dynamics of cognitive control. Proceedings of the National Academy of Sciences.
2009; 106:5529–5533.
Chatham CH, Herd SA, Brant AM, Hazy TE, Miyake A, O’Reilly R, Friedman NP. From an executive
network to executive control: a computational model of the n-back task. Journal of Cognitive
NIH-PA Author Manuscript

Neuroscience. 2011; 23:3598–3619. [PubMed: 21563882]

Cohen JD, Dunbar K, McClelland JL. On the control of automatic processes: A parallel distributed
processing model of the Stroop effect. Psychological Review. 1990; 97:332–361. [PubMed:
2200075]
Collette F, Van der Linden M, Laureys S, Delfiore G, Degueldre C, Luxen A, Salmon E. Exploring the
unity and diversity of the neural substrates of executive functioning. Human Brain Mapping. 2005;
25:409–423. [PubMed: 15852470]
Collins AGE, Frank MJ. Cognitive control over learning: Creating, clustering, and generalizing task-
set structure. Psychological Review. 2013; 120:190–229. [PubMed: 23356780]
Colzato LS, Waszak F, Nieuwenhuis S, Posthuma D, Hommel B. The flexible mind is associated with
the catechol-o-methyltransferase (COMT) Val158Met polymorphism: Evidence for a role of
dopamine in the control of task-switching. Neuropsychologia. 2010; 48:2764–2768. [PubMed:
20434465]

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 25

Cools, R. Chemical neuromodulation of goal-directed behavior. In: Todd, PM.; Hills, TT.; Robbins,
TW.; Lupp, J., editors. Cognitive search: Evolution, algorithms, and the brain. Vol. 9. Cambridge,
MA: MIT Press; 2012. p. 111-123.Strüngmann Forum Report
NIH-PA Author Manuscript

Crittenden BM, Duncan J. Task difficulty manipulation reveals multiple demand activity but no frontal
lobe hierarchy. Cerebral Cortex. 2014; 24:532–540. [PubMed: 23131804]
Curtis CE, Lee D. Beyond working memory: The role of persistent activity in decision making. Trends
in Cognitive Sciences. 2011; 14:216–222. [PubMed: 20381406]
D’Ardenne K, Eshel N, Luka J, Lenartowicz A, Nystrom LE, Cohen JD. Role of prefrontal cortex and
the midbrain dopamine system in working memory updating. Proceedings of the National
Academy of Sciences of the United States of America. 2012; 109:19900–19909. [PubMed:
23086162]
De Jong, R. An intention-activation account of residual switch costs. In: Monsell, S.; Driver, J.,
editors. Control of cognitive processes: Attention and performance XVIII. Cambridge, MA: MIT
Press; 2000. p. 357-376.
Deco G, Rolls ET. Synaptic and spiking dynamics underlying reward reversal in the orbitofrontal
cortex. Cerebral Cortex. 2005; 15:15–30. [PubMed: 15238449]
Desimone R, Duncan J. Neural mechanisms of selective visual attention. Annual Review of
Neuroscience. 1995; 18:193–222.
Duncan J, Emslie H, Williams P, Johnson R, Freer C. Intelligence and the frontal lobe: The
organization of goal-directed behavior. Cognitive psychology. 1996; 30:257–303. [PubMed:
8660786]
NIH-PA Author Manuscript

Egan MF, Goldberg TE, Kolachana BS, Callicott JH, Mazzanti CM, Straub RE, Goldman D,
Weinberger DR. Effect of COMT val108/158met genotype on frontal lobe function and risk for
schizophrenia. Proceedings of the National Academy of Sciences of the United States of America.
2001; 98:6917–6922. [PubMed: 11381111]
Engle, RW.; Kane, M.; Tuholski, S. Individual differences in working memory capacity and what they
tell us about controlled attention, general fluid intelligence and functions of the prefrontal cortex.
In: Miyake, A.; Shah, P., editors. Models of working memory: Mechanisms of active maintenance
and executive control. New York: Cambridge University Press; 1999. p. 102-134.
Frank MJ. Hold your horses: A dynamic computational role for the subthalamic nucleus in decision
making. Neural Networks. 2006; 19:1120–1136. [PubMed: 16945502]
Frank MJ, Loughry B, O’Reilly RC. Interactions between the frontal cortex and basal ganglia in
working memory: A computational model. Cognitive, Affective, and Behavioral Neuroscience.
2001; 1:137–160.
Frank MJ, Seeberger LC, O’Reilly RC. By carrot or by stick: Cognitive reinforcement learning in
Parkinsonism. Science. 2004; 306:1940–1943. [PubMed: 15528409]
Friedman, NP.; Herd, SA.; Brant, AM.; Hazy, TE.; Kriete, T.; Chatham, CH.; O’Reilly, RC. Neural
network models of individual differences in executive functions. 2014. Manuscript in preparation
Friedman N, Miyake A, Corley R, Young S, DeFries J, Hewitt J. Not all executive functions are
related to intelligence. Psychological Science. 2006; 17:172–179. [PubMed: 16466426]
NIH-PA Author Manuscript

Friedman NP, Haberstick BC, Willcutt EG, Miyake A, Young SE, Corley RP, Hewitt JK. Greater
attention problems during childhood predict poorer executive functioning in late adolescence.
Psychological science. 2007; 18:893–900. [PubMed: 17894607]
Friedman NP, Miyake A. The relations among inhibition and interference control functions: A latent-
variable analysis. Journal of Experimental Psychology: General. 2004; 133:101–135. [PubMed:
14979754]
Friedman NP, Miyake A, Robinson JL, Hewitt JK. Developmental trajectories in toddlers’ self-
restraint predict individual differences in executive functions 14 years later: A behavioral genetic
analysis. Developmental Psychology. 2011; 47:1410–1430. [PubMed: 21668099]
Friedman NP, Miyake A, Young SE, Defries JC, Corley RP, Hewitt JK. Individual differences in
executive functions are almost entirely genetic in origin. Journal of Experimental Psychology:
General. 2008; 137:201–225. [PubMed: 18473654]

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 26

Goschke, T. Intentional reconfiguration and involuntary persistence in task set switching. In: Monsell,
S.; Driver, J., editors. Control of cognitive processes: Attention and performance XVIII.
Cambridge, MA: MIT Press; 2000. p. 331-355.
NIH-PA Author Manuscript

Hazy TE, Frank MJ, O’Reilly RC. Banishing the homunculus: Making working memory work.
Neuroscience. 2006; 139:105–118. [PubMed: 16343792]
Hazy TE, Frank MJ, O’Reilly RC. Towards an executive without a homunculus: Computational
models of the prefrontal cortex/basal ganglia system. Philosophical Transactions of the Royal
Society of London, Series B, Biological Sciences. 2007; 362:105–118.
Hazy TE, Frank MJ, O’Reilly RC. Neural mechanisms of acquired phasic dopamine responses in
learning. Neuroscience and Biobehavioral Reviews. 2010; 34:701–720. [PubMed: 19944716]
Hedden T, Yoon C. Individual differences in executive processing predict susceptibility to interference
in verbal working memory. Neuropsychology. 2006; 20:511–528. [PubMed: 16938014]
Herd SA, Banich MT, O’Reilly RC. Neural mechanisms of cognitive control: An integrative model of
Stroop task performance and fMRI data. Journal of Cognitive Neuroscience. 2006; 18:22–32.
[PubMed: 16417680]
Kane MJ, Engle RW. Working-memory capacity and the control of attention: The contributions of
goal neglect, response competition, and task set to Stroop interference. Journal of Experimental
Psychology: General. 2003; 132:47–70. [PubMed: 12656297]
Kiesel A, Steinhauser M, Wendt M, Falkenstein M, Jost K, Philipp AM, Koch I. Control and
interference in task switching — A review. Psychological Bulletin. 2010; 136:849–874. [PubMed:
20804238]
NIH-PA Author Manuscript

Koch I, Gade M, Schuch S, Philipp AM. The role of inhibition in task switching: A review.
Psychonomic Bulletin & Review. 2010; 17:1–14. [PubMed: 20081154]
Koechlin E, Ody C, Kouneiher F. Neuroscience: The architecture of cognitive control in the human
prefrontal cortex. Science. 2003; 302:1181–1184. [PubMed: 14615530]
Krichmar JA. The Neuromodulatory system: A framework for survival and adaptive behavior in a
challenging world. Adaptive Behavior. 2008; 16:385–399.
Kray J, Lindenberger U. Adult age differences in task switching. Psychology and Aging. 2000;
15:126–147. [PubMed: 10755295]
Krueger KA, Dayan P. Flexible shaping: How learning in small steps helps. Cognition. 2009;
110:380–394. [PubMed: 19121518]
Lehto JE, Juujarvi P, Kooistra L, Pulkkinen L. Dimensions of executive functioning: Evidence from
children. British Journal of Developmental Psychology. 2003; 21:59–80.
Logan GD, Bundesen C. Clever homunculus: Is there an endogenous act of control in the explicit task-
cuing procedure? Journal of Experimental Psychology: Human Perception and Performance. 2003;
29:575–599. [PubMed: 12848327]
Logan GD, Gordon RD. Executive control of visual attention in dual-task situations. Psychological
Review. 2001; 108:393–434. [PubMed: 11381835]
Logan GD, Schneider DW. Distinguishing reconfiguration and compound-cue retrieval in task
NIH-PA Author Manuscript

switching. Psychologica Belgica. 2010; 50:413–433.

Mayr U, Keele SW. Changing internal constraints on action: the role of backward inhibition. Journal
of Experimental Psychology: General. 2000; 129:4–26. [PubMed: 10756484]
Meiran N, Kessler Y, Adi-Japha E. Control by action representation and input selection (CARIS): A
theoretical framework for task switching. Psychological Research. 2008; 72:473–500. [PubMed:
18350316]
Miller EK. The prefrontal cortex and cognitive control. Nature Reviews Neuroscience. 2000; 1:59–65.
Miller EK, Cohen JD. An integrative theory of prefrontal cortex function. Annual Review of
Neuroscience. 2001; 24:167–202.
Miyake A, Emerson MJ, Padilla F, Ahn J. Inner speech as a retrieval aid for task goals: The effects of
cue type and articulatory suppression in the random task cuing paradigm. Acta Psychologica.
2004; 115:123–142. [PubMed: 14962397]

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 27

Miyake A, Friedman NP. The nature and organization of individual differences in executive functions:
Four general conclusions. Current Directions in Psychological Science. 2012; 21:8–14. [PubMed:
22773897]
NIH-PA Author Manuscript

Miyake A, Friedman NP, Emerson MJ, Witzki AH, Howerter A, Wager TD. The unity and diversity of
executive functions and their contributions to complex “frontal lobe” tasks: A latent variable
analysis. Cognitive Psychology. 2000; 41:49–100. [PubMed: 10945922]
Monsell S. Task switching. Trends in Cognitive Sciences. 2003; 7:134–140. [PubMed: 12639695]
Morton JB, Munakata Y. Active versus latent representations: A neural network model of
perseveration, dissociation, and decalage. Developmental Psychobiology. 2002; 40:255–265.
[PubMed: 11891637]
Munakata Y, Herd SA, Chatham CH, Depue BE, Banich MT, O’Reilly RC. A unified framework for
inhibitory control. Trends in Cognitive Sciences. 2011; 15:453–459. [PubMed: 21889391]
Norman KA, Newman EL, Detre G. A neural network model of retrieval-induced forgetting.
Psychological Review. 2007; 114:887–953. [PubMed: 17907868]
Newell, A. Unified Theories of Cognition. Cambridge, MA: Harvard University Press; 1990.
O’Reilly R. Biologically based computational models of high-level cognition. Science. 2006; 314:91–
94. [PubMed: 17023651]
O’Reilly, RC.; Braver, TS.; Cohen, JD. A biologically based computational model of working
memory. In: Miyake, A.; Shah, P., editors. Models of working memory: Mechanisms of active
maintenance and executive control. New York: Cambridge University Press; 1999. p. 375-411.
O’Reilly RC, Frank MJ. Making working memory work: A computational model of learning in the
NIH-PA Author Manuscript

prefrontal cortex and basal ganglia. Neural Computation. 2006; 18:283–328. [PubMed: 16378516]
O’Reilly RC, Frank MJ, Hazy TE, Watz B. PVLV: The primary value and learned value Pavlovian
learning algorithm. Behavioral Neuroscience. 2007; 121:31–49. [PubMed: 17324049]
O’Reilly, RC.; Hazy, TE.; Herd, SA. The Leabra cognitive architecture: How to play 20 principles
with nature and win!. In: Chipman, S., editor. The Oxford handbook of cognitive science. New
York: Oxford University Press; (in press)
O’Reilly, RC.; Munakata, Y. Computational explorations in cognitive neuroscience: Understanding the
mind by simulating the brain. Cambridge, MA: MIT Press; 2000.
O’Reilly, RC.; Munakata, Y.; Frank, MJ.; Hazy, TE., et al. Computational cognitive neuroscience. 1.
Wiki Book; 2012. URL: http://ccnbook.colorado.edu
Radloff LS. The CES-D scale: A self-report depression scale for research in the general population.
Applied Psychological Measurement. 1977; 1:385–401.
Reynolds JR, Braver TS, Brown JW, Van der Stigchel S. Computational and neural mechanisms of
task switching. Neurocomputing. 2006; 69:1332–1336.
Reynolds JR, O’Reilly RC, Cohen JD, Braver TS. The function and organization of lateral prefrontal
cortex: a test of competing hypotheses. PloS One. 2012; 7:e30284. [PubMed: 22355309]
Rigas P, Castro-Alamancos MA. Thalamocortical up states: Differential effects of intrinsic and
extrinsic cortical inputs on persistent activity. Journal of Neuroscience. 2007; 27:4261–4272.
NIH-PA Author Manuscript

[PubMed: 17442810]
Rogers RD, Monsell S. Costs of a predictable switch between simple cognitive tasks. Journal of
Experimental Psychology: General. 1995; 124:207–231.
Rougier NP, Noelle D, Braver TS, Cohen JD, O’Reilly RC. Prefrontal cortex and the flexibility of
cognitive control: Rules without symbols. Proceedings of the National Academy of Sciences.
2005; 102:7338–7343.
Rubin O, Meiran N. On the origins of the task mixing cost in the cuing task-switching paradigm.
Journal of Experimental Psychology: Learning, Memory, and Cognition. 2005; 31:1477–1491.
Servan-Schreiber D, Printz H, Cohen JD. A network model of catecholamine effects: Gain, signal-to-
noise ratio, and behavior. Science. 1990; 249:892–895. [PubMed: 2392679]
Skinner, BF. The Behavior of Organisms. New York: Appleton-Century-Crofts; 1938.
Sohn MH, Anderson JR. Task preparation and task repetition: Two-component model of task
switching. Journal of Experimental Psychology, General. 2001; 130:764–778. [PubMed:
11757879]

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 28

Stroop JR. Studies of interference in serial verbal reactions. Journal of Experimental Psychology.
1935; 18:643–662.
Teuber HL. Unity and diversity of frontal lobe functions. Acta Neurobiologiae Experimentalis. 1972;
NIH-PA Author Manuscript

32:615–656. [PubMed: 4627626]

Tunbridge EM, Harrison PJ, Weinberger DR. Catechol-O-methyltransferase, cognition, and psychosis:
Val158Met and beyond. Biological Psychiatry. 2006; 60:141–151. [PubMed: 16476412]
Vandierendonck A, Liefooghe B, Verbruggen F. Task switching: Interplay of reconfiguration and
interference control. Psychological Bulletin. 2010; 136:601–626. [PubMed: 20565170]
Wechsler, D. WAIS-III: Wechsler adult intelligence scale. San Antonio, TX: Psychological
Corporation; 1997.
Young SE, Friedman NP, Miyake A, Willcutt EG, Corley RP, Haberstick BC, Hewitt JK. Behavioral
disinhibition: Liability for externalizing spectrum disorders and its genetic and environmental
relation to response inhibition across adolescence. Journal of Abnormal Psychology. 2009;
118:117–130. [PubMed: 19222319]
NIH-PA Author Manuscript
NIH-PA Author Manuscript

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 29

Highlights
• We present a neural network model of individual differences in task switching.
NIH-PA Author Manuscript

• We compared results for a Stroop model that captures general executive ability.

• Active goal maintenance and top-down biasing affected both switch and Stroop
cost.

• Persistence of no-longer-relevant goals increased switch cost but not Stroop

cost.

• These mechanisms may create some of the unity and diversity of executive
functions.
NIH-PA Author Manuscript
NIH-PA Author Manuscript

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 30
NIH-PA Author Manuscript

Figure 1.
Unity/Diversity model from Friedman et al. (2008). There is a Common EF latent variable
on which all nine EF tasks load, as well as two “nested” latent variables on which the
updating and shifting tasks, respectively, also load. The Common EF variance turned out to
be isomorphic with the Inhibiting latent variable (see Friedman et al., 2008); there was not
significant inhibiting-specific variance. Because the Common EF factor captures the
variance common to all three EFs, the Updating-specific and Shifting-specific factors
capture the variance that is unique to updating and shifting, respectively. Hence, they are
uncorrelated with the Common EF factor and with each other. Antisac = antisaccade, Stop =
stop-signal, Letter = letter memory, S2back = spatial 2-back, Number = number-letter, Color
= color-shape, Category = category-switch.
NIH-PA Author Manuscript
NIH-PA Author Manuscript

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 31
NIH-PA Author Manuscript

Figure 2.
Key model components and predictions for contributions to individual difference in EF.
Hypothesized mechanisms influencing Common EF and Shifting-Specific components are
depicted above the ovals (model parameter used to manipulate each mechanism in
NIH-PA Author Manuscript

parentheses), and inside the ovals are the predicted effects on Common EF (top) and
Shifting-Specific (bottom). Arrows are excitatory connections. For the gating connections,
the connection “opens” the gate to allow new representations into PFC for maintenance
under appropriate conditions, and the terminal circle is an indirect inhibitory effect of the
basal ganglia gating signal; this wave of inhibitory cortical activity helps “clear” old PFC
representations to allow new representations to come in. The key Common EF mechanisms
we manipulate are Strength of PFC outputs and representations. The strength of PFC outputs
(e.g, corticortical and cortico-thalamo-cortical connection density and average synaptic
strength) has a positive effect on Common EF since it allows makes all task sets to have a
stronger top-down biasing effect on posterior areas. It has no effect on switching speed
because it does not affect the persistence of PFC representations in any way. The strength of
PFC representations (i.e., signal-to-noise ratio in PFC) also has a positive effect on Common
EF because it enables the task set representations to be stronger and more precise. However,
it slows down switching, as stronger representations are more difficult to dislodge when no
longer relevant. The key Shifting-Specific mechanisms we manipulated were non-gated
persistence of PFC representations and clearing on gating. Non-gated PFC persistence
NIH-PA Author Manuscript

(manipulated with recurrent connection strength) slows down switching because it slows
down the process of replacing task sets. It generally has no effect (as in our parameter
manipulation and tasks) or a positive effect on many EF tasks in which persistence of task
sets is useful, or when persistence is linked to stronger PFC representations. Clearing on
gating speeds up switching while not affecting most EF tasks, as it was implemented to have
no effect outside of the moment of gating. Many physiological sources of clearing
differences (such as strength of inhibitory interneurons in PFC) would likely produce effects
on other EF components as well.

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 32
NIH-PA Author Manuscript

Figure 3.
Model Architecture. Each small rectangle is a single unit. Lines to the bottom right of a
NIH-PA Author Manuscript

layer are incoming connections, while those from the left are outgoing connections. The
general flow of information processing proceeds from left to right, from Color, Shape, and
Task inputs, through the two Processing (hidden) layers, to the Verbal and Manual outputs.
The PFC layer holds information in active states from previous trials and that information
modulates processing in the two hidden layers. The PFC layer is composed of stripes, each
of which is “gated” to maintain information based on the result of learning in the Matrix
(striatum). This learning is in turn based on a predicted reward, calculated in the PVLV
module (lower left), which is shown without units and connections to make other
connections visible. The PBWM framework and PVLV system are described in detail
elsewhere; see text for references.
NIH-PA Author Manuscript

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 33
NIH-PA Author Manuscript
NIH-PA Author Manuscript

Figure 4.
NIH-PA Author Manuscript

Interference and Switch costs (A), and RTs in the (B) Color Shape task and (C) Stroop Task
as a function of PFC to posterior (hidden) layer connection strength. Bars indicate ± 2
standard errors, approximating 95% confidence intervals.

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 34
NIH-PA Author Manuscript
NIH-PA Author Manuscript

Figure 5.
Interference and Switch costs (A), and RTs in the (B) Color Shape task and (C) Stroop Task
as a function of PFC gain. Bars indicate ± 2 standard errors, approximating 95% confidence
intervals.
NIH-PA Author Manuscript

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 35
NIH-PA Author Manuscript
NIH-PA Author Manuscript

Figure 6.
Interference and Switch costs (A), and RTs in the (B) Color Shape task and (C) Stroop Task
as a function of PFC recurrent connection strength. Bars indicate ± 2 standard errors,
approximating 95% confidence intervals.
NIH-PA Author Manuscript

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 36
NIH-PA Author Manuscript
NIH-PA Author Manuscript

Figure 7.
Interference and Switch costs (A), and RTs in the (B) Color Shape task and (C) Stroop Task
as a function of the clearing on gating parameter. Bars indicate ± 2 standard errors,
approximating 95% confidence intervals.
NIH-PA Author Manuscript

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.

 Herd et al. Page 37

Table 1

Correlations of Problems in a Longitudinal Twin Sample with Unity/Diversity EF Latent Variables at Age 17.
NIH-PA Author Manuscript

Correlated Variable Common EF Updating-Specific Shifting-Specific

WAIS-IQa .51* .49* −.24*

Attention problems ages 7 to 14b

Intercept of growth model −.52* .04 .20~

Slope of growth modelc .03 −.20 −.34*

Behavioral Disinhibition Age 12d

Age 12 phenotypic −.43* .05 .14~
Age 12 genetic −.58* .07 .34*

Behavioral Disinhibition Age 17d

Age 17 phenotypic −.39* .11 .11

Age 17 genetic −.63* .16 .30*

Self Restraint Age 14–36 monthse

Latent class membership phenotypicf .29* .02 −.21*

NIH-PA Author Manuscript

Latent class membership genetic .49* .01 −.34

*
p < .05.
~
p < .10.
a
From Friedman et al. (2011).
b
Reanalysis of Friedman et al. (2007).
c
Higher values indicate greater decreases in problems across time.
d
Reanalysis of Young et al. (2009).
e
From Friedman et al. (2011).
f
Cohen’s D to r conversions
NIH-PA Author Manuscript

Neuropsychologia. Author manuscript; available in PMC 2015 September 01.