A computational model of fractionated conflict-control mechanisms in
a, c, , baJoshua W. Brown, Jeremy R. Reynolds and Todd S. Braver aDepartment of Psychology, Washington University, St. Louis, MO 63130, USA bDepartment of Psychology, University of Colorado Boulder, Boulder, CO 80309, USA cDepartment of Psychological and Brain Sciences, Indiana University, Bloomington, IN 47405, USA
Accepted 14 September 2006. Available online 31 October 2006. Abstract
A feature of human cognition is the ability to monitor and adjust one’s own behavior under changing circumstances. A dynamic balance between controlled and rapid responding is needed to adapt to a fluctuating environment. We suggest that cognitive control may include, among other things, two distinct processes. Incongruent stimuli may drive top-down facilitation of task-relevant responses to bias performance toward exploitation vs. exploration. Task or response switches may generally slow responses to bias toward accuracy vs. speed and exploration vs. exploitation. Behavioral results from a task switching study demonstrate these two distinct processes as revealed by higher-order sequential effects. A computational model implements the two conflict-control mechanisms, which allow it to capture many complex and novel sequential effects. Lesion studies with the model demonstrate that the model is unable to capture these effects without the conflict-control loops and show how each monitoring component modulates cognitive control. The results suggest numerous testable predictions regarding the neural substrates of cognitive control.
Keywords: Cognitive control; Task switching; Sequential effects; Performance monitoring; Neural network; Computational modeling
1.1. Cognitive control in task switching 1.2. Higher-order sequential effects 2. Behavioral study: Sequential effects in task-switching
3.2. Tasks and stimuli
3.4. Data analysis
4. Results and discussion
4.1. Current trial effects
4.2. Effects of preceding trials 4.3. Interactions between current and previous trials
4.4. Error rate
4.5. Speed–accuracy effects
5. Experiment discussion
6. Modelling study: Simulating task-switching sequential effects
7. Model mechanisms
Model methods 8.
Priming mechanisms 8.3.
8.4. Performance monitoring modules 9. Model analysis methods
10. Model results
10.1. Higher-order sequences
10.2. Performance monitor activity signatures 10.3. Nested model analysis
10.4. Lesions of cognitive control mechanisms 10.5. Switch costs
11. Model discussion
12. General discussion
12.1. Switch cost
12.2. Cognitive control modulates performance tradeoffs
12.3. Cognitive neuroscience
Appendix A. Appendix
A.1. Parameter optimization
Appendix A. Supplementary data References
Environments often change unpredictably over time. In some cases, adaptation to such changing environments requires the learning of new responses. However, in other cases, the same response must be generated with slightly changed parameters. This necessitates an ongoing adjustment of behavioral control. A classic example is the tradeoff between speed
[Osman et al., 2000], [Plamondon and Alimi, 1997] and and accuracy (
[Strayer and Kramer, 1994]). If the cost or likelihood of errors is low
and speed essential, then one will do well to execute a given action as quickly as possible with less regard for accuracy. Conversely, if the cost of errors is high and speed less important, then one will do well to increase behavioral control, slow down, and be more careful. Subjects may adopt a strategy ranging between rapid (emphasis on speed) and controlled (emphasis on accuracy) responding. In this case, shifting the bias toward accuracy rather than speed might be considered an example of a simple form of cognitive control. More complex forms of control might involve adjusting attentional allocation between focused exploitation of known aspects of an environment versus exploration of unknown components [Ishii et al., 2002], [Kaelbling et al., 1996], [Sutton and Barto, 1998] (
and [Usher et al., 1999]) or between fast, pre-potent versus controlled, non-prepotent responding ([Pardo et al., 1990] and [Weissman et al.,
2005]). Notably, the cognitive control strategy may change independently of changes in the tasks being performed. The domain of cognitive or executive control may thus include many different kinds of control effects and underlying mechanisms (Norman & Shallice, 1986). Recent efforts have
been made to dissect various components on empirical and meta-analytic grounds into such categories as shifting, monitoring or updating, inhibition, and selective attention ([Miyake et al., 2000] and [Wager and
Smith, 2003]). Our goal in this paper is to develop a fractionation of cognitive control that derives from an integration of theoretical, computational, and behavioral analyses. To do this, we focus on task switching as a well-studied representative paradigm ([Allport et al.,
1994], [De Jong et al., 1999], [Dreisbach et al., 2002], [Hodgson et al.,
2002], [Jersild, 1927], [Meiran, 1996], [Meiran, 2000a], [Meiran, 2000b],
[Meiran et al., 2000], [Meiran and Gotler, 2001], [Meiran and Marciano,
2002], [Nieuwenhuis and Monsell, 2002], [Rogers and Monsell, 1995], [Sohn
and Anderson, 2001] and [Wylie and Allport, 2000b]). Studies of switching
between two cognitive tasks afford significant insight into control of both cognitive and motor processes by dissociating changes in task set (cognitive) from changes in the required response (motor). 1.1. Cognitive control in task switching
In this paper, we have focused the scope of the analysis and simulations toward the specific aims of elucidating cognitive control mechanisms.
There has been significant recent controversy regarding the extent to which executive control mechanisms are necessary to drive a task switch ([Altmann, 2003], [Logan and Bundesen, 2003], [Monsell, 2003] and [Rogers
and Monsell, 1995]). We do not take an absolute position on this issue. Nonetheless, the controversies suggest that further specification is needed of different types of control functions and how they might influence performance during task-switching. Rather than investigating whether control functions are necessary to drive a task switch, we focus instead on delineating mechanisms of performance monitoring and control that modulate response parameters within a particular task set (Altmann
& Gray, 2002). In so doing, we develop a computational model of cognitive control in standard task switching experimental paradigms. Implementing theoretical hypotheses as computational models provides a critical means to clarify the relationship between complex effects and relatively simpler underlying mechanisms ([Botvinick et al., 2001], [Braver et al.,
1999], [Cho et al., 2002], [Gilbert and Shallice, 2002], [Jones et al.,
2002], [Melara and Algom, 2003], [Miller and Cohen, 2001] and [Phaf et
al., 1990]). However, our purpose in developing a computational framework is not to provide a comprehensive model of task switching but rather to use it as a means to an end. We use task switching as a representative domain in which multiple control mechanisms may interact within a single experimental paradigm. Our aim in this endeavor is twofold. First, we aim to explore competing hypotheses that behavioral effects in task-switching can be accounted for solely by bottom up mechanisms, versus the hypothesis that top-down effects such as control loops provide better accounts of a range of data, even if such control loops are not required for task-switching per se. Second, we aim to concretely delineate specific mechanisms of cognitive control that may generalize beyond task switching to a range of behavioral paradigms.
A central issue in any task-switching study is the importance of sequential relationships between trials. At a basic level, a task switch involves information from two consecutive trials, in that the task changes from trial n ? 1 to trial n. Studies of task-switching have, with some
exceptions ([Altmann and Gray, 2002] and [Mayr, 2002]) generally focused
on the current and immediately preceding trial, in that a task switch involves task A in the preceding trial and task B in the current trial. Perhaps the most prominent finding is that of switch costs, which refer to the slower, less accurate performance when the task switches relative to the preceding trial as compared to when it does not switch (Jersild,
1927). The switch cost effect persists residually even with long preparation times ([Allport et al., 1994], [de Jong, 2000], [Meiran et
al., 2000], [Nieuwenhuis and Monsell, 2002] and [Rogers and Monsell,
The origin of the switch cost is controversial. Some investigators (Rogers
& Monsell, 1995) interpret the residual switch cost as the time needed to reconfigure the system, which cannot be completed until the target stimuli appear. This interpretation implies a putative top-down executive
control mechanism, possibly distinct from task-specific
stimulus-response pathways, that implements the task switch and produces the residual switch cost. On the other hand, Allport and colleagues argue for a explanation of switch costs ([Allport et al., 1994] and bottom-up
[Wylie and Allport, 2000b]). According to their associative
task-set-interference (TSI) hypothesis, the switch cost originates from associative strengthening between task-related stimuli and an internal representation of the task-set with which they are paired on a previous trial. In the current trial, stimuli will tend to re-evoke the previous task-set, even if the previous task-set is different from the one relevant for the current trial (i.e., the task switched). This pairing may lead to interference that lengthens RTs (Waszak, Hommel, & Allport, 2003). In
the same vein, recent modeling work (Gilbert & Shallice, 2002) has shown
that switch costs can be accounted for in part by a combination of residual activity in task set representations and associative TSI. In addition to target-related priming effects, some portion of the apparent switch cost
[Arrington and Logan, may be due to a loss of cue repetition benefits (
2004a], [Arrington and Logan, 2004b], [Logan and Bundesen, 2003] and [Mayr
and Kliegl, 2003]) related to stimulus-specific cue priming effects. Also in favor of a bottom-up explanation of switch costs, several authors have argued that switch costs represent a failure to proactively reconfigure
task set on a subset of trials and maintain this configuration across a delay ([De Jong, 2000], [De Jong et al., 1999], [Meiran, 1996] and
[Nieuwenhuis and Monsell, 2002]). This is referred to as the
“failure-to-engage” (FTE) hypothesis. Thus, the switch cost would result from the time required to reactively instate the correct task set
representation at the time of target stimulus presentation, while also overriding interference from the prepotent but inappropriate task-set. Computational modelling work has shown that the FTE hypothesis can, in fact, provide a quantitative account of the relevant behavioral data (Reynolds, Braver, Brown, & Stigchel, 2006). Others have suggested that
switching away from a given task involves a possibly top-down inhibitory process, namely backward inhibition, that suppresses the previous task set representation, making it more difficult to return to it again ([Mayr,
2002] and [Mayr and Keele, 2000]). Finally, on the basis of previous work
([Cho et al., 2002] and [Jones et al., 2002]), we will suggest below that
top-down mechanisms might impose a protracted slowing effect that persists into subsequent trials. While all of these factors may contribute to the switch cost to some degree, further specification is needed regarding the relative contributions of each.
Aside from the switch cost, incongruency effects are also prominent in task switching studies. It is well-established that stimuli associated with incongruent responses lead to interference and thus poorer performance in a range of cognitive tasks ([Botvinick et al., 2001],
[Botvinick et al., 1999], [Egner and Hirsch, 2005], [Eriksen and Eriksen,
1974] and [MacLeod, 1991]). Moreover, incongruency effects have been
associated with a conflict adaptation effect (Mayr, Awh, & Laurey, 2003),
i.e. reduced incongruency effects on subsequent trials, possibly due to a corresponding increase in cognitive control ([Kerns et al., 2004] and
[MacDonald et al., 2000]). Within task-switching paradigms, incongruency effects occur when a feature of the target stimulus is associated with an incompatible response according to the currently irrelevant task. In the case of incongruent stimuli, cognitive control mechanisms may be needed to increase activity of the (possibly already active) task set representation or pathway, and directly or indirectly suppress irrelevant task set representations or pathways (Egner & Hirsch, 2005). Given
incongruent stimuli, a performance monitor may serve to increase
-related activity and subsequent attentional focus to persistent task-set
the current task; once increased, the effect may persist into subsequent trials. Several predictions follow from this hypothesis. First, despite the cost of incongruency on both response time and error rate in the current trial, subsequent trials of the same task would be expected to show an improvement in performance. In support of this account, recent work (Goschke, 2000) suggests that incongruency leads to greater switch costs in the subsequent trial. Specifically, prior incongruency increased response time on switch trials and reduced response time on no-switch trials. This may reflect facilitation of repeating the same task (though not necessarily the same response), as well as increased inhibition of the previously irrelevant task, both of which are expected if incongruency does lead to an enhancement of representations of the current task set (at the expense of representations of the currently irrelevant task set), which persists into subsequent trials. Furthermore, if the subsequent trial is incongruent as well as a switch, then this incongruency should interact with the persistent previous task set representation to increase switch costs further. Although some of the effect of incongruency in the previous trial may be due to stimulus repeat effects (Mayr et al., 2003),
in the case of task switching, the effect persists even when episodic stimulus repeats are eliminated in the sequence of interest (Goschke,
2000). In the experiment below, we begin by attempting to replicate these findings.
1.2. Higher-order sequential effects
As the above suggests, cognitive control mechanisms may exert protracted and persistent effects on performance, and therefore they may be evident most clearly in the effects of longer sequences of trial conditions. Higher order sequential effects refer to conditions involving sequences of more than two trials. In general, a full descriptive model of performance monitoring will need to make reference to higher order effects (Laming, 1968). To our knowledge, there have been no previous studies that have systematically examined the presence of higher-order sequential effects in task-switching. However, there are strong reasons to a priori
suggest that these effects will be present and significant. Specifically, within simpler task paradigms, such as two-alternative forced choice (2AFC), there has been a great deal of examination of the role of higher-order sequential effects in modulating behavioral performance. For example, there is a long tradition of research establishing that error commission ([Laming, 1968] and [Rabbitt, 1966]) or a change in the
required response (Bertelson, 1961) produces a relatively persistent
slowing in response time on subsequent task trials. Furthermore, there is evidence that such sequential effects may reflect specifically top-down control mechanisms, such as expectation of particular sequences (Soetens, Boer, & Hueting, 1985), in addition to bottom-up mechanisms.
Evidence for the co-existence of top-down and bottom-up mechanisms of sequential effects has also been found previously with the aid of computational modeling of 2AFC tasks (Cho et al., 2002).
In principle, different forms of conflict might usefully modulate distinct components of cognitive control. One form of conflict may be evoked by switching tasks from one trial to the next. Specifically, if the required task or response changes frequently, then it is difficult to predict where to focus attention for optimal performance. Thus, conflict between the expected and actual required responses due to changes in the required task or response may be effectively addressed by generally slowing responses in subsequent trials, to prevent an anticipated response from being prematurely (and erroneously) generated before external stimuli can be adequately processed. On the other hand, another form of conflict may be evoked by incongruent stimuli. If the task requirements change little but strong task-irrelevant, conflicting stimuli appear, then performance may be best served if conflict due to task-irrelevant stimuli drives increased attentional focus to the relevant stimuli. Conversely, if there were only a single form of conflict-control mechanism used to adjust performance, responses could not be effectively adapted to the constraints of specific task situations. Thus, a non-specific response slowing mechanism would not produce an appropriate shift in attentional focus towards task-relevant stimuli and away from task-irrelevant ones. Likewise, an attentional focusing
mechanism would be ineffective in responding to unexpected changes in task-requirements. Thus, we postulate that there are multiple conflict-control loop mechanisms in the brain that are each associated with regulating adjustment in specific forms of cognitive control. It is well-established that incongruency effects across a range of task paradigms are associated with a form of conflict, and that this conflict engages the anterior cingulate cortex (ACC) and related brain areas that appear to perform performance monitoring and control functions ([Botvinick et al., 1999], [Carter et al., 1998] and [Kerns et al., 2004]).
Likewise, the results of 2AFC and related paradigms (e.g., go-nogo) indicate that response switching also appears to engage similar conflict and performance monitoring processes in the ACC and elsewhere ([Jones et
al., 2002] and [Stuphorn et al., 2000]). In a task-switching environment,
switches in both response and task also occur frequently from one trial to the next, and as such may engage a conflict-detection mechanism that non-specifically slows responses on subsequent trials. There is some evidence that task-switching engages conflict-detection mechanisms in the ACC ([Dove et al., 2000] and [Luks et al., 2002]).
Below, we present a systematic analysis of sequential effects from an experimental study of task-switching that demonstrates the presence of potentially complex and confusing higher-order sequential relationships in behavioral performance. We then show that a computational model with two performance monitoring and control mechanisms can account for the observed complex behavioral effects, while the model without the two conflict-control loops cannot account for the observed effects. The model thus supports the hypothesis of multiple generalized cognitive control mechanisms and produces quantitative predictions that can be tested with functional imaging and lesion studies.
2. Behavioral study: Sequential effects in task-switching Our aim was to test for the existence of effects consistent with cognitive control mechanisms as discussed above. To this end, we conducted a task switching study and looked at factors of: (1) task switch (S) vs. no-switch (N); (2) response alternation (A) vs. repetition (R), and (3) stimulus incongruency (I) vs. congruency (C). Trials were coded this way in accordance with previous task-switching studies that have observed significant effects or interactions of these factors on behavioral performance ([Allport et al., 1994] and [Rogers and Monsell, 1995]). To
avoid confusion, we use the terms repetition and alternation to refer to
changes in the required response from trial to trial, and switch or
no-switch to denote changes in the required task set. In order to analyze
the interaction of these factors in higher-order sequential effects, we crossed each of the eight possible current-trial conditions with the eight previous trial conditions, for a total of 64 task conditions (6 factors). Following earlier conventions ([Botvinick et al., 1999] and [Kerns et al.,
2004]), we use uppercase letters to denote current-trial conditions and lowercase letters to denote previous trial conditions. Importantly, this coding method implicitly includes information from three consecutive trials (though not all information from the earliest trial was analyzed—in particular, the incongruency factor was omitted). For example, one possible trial condition is nrcSAI, i.e., the previous trial
conditions were no-task-switch, response repetition, congruent stimulus followed by current trial conditions of task switch, response alternation, and incongruent stimulus.
The complexity of this analysis method was justified in that it allowed us to investigate several issues. First, we attempted to replicate previous results (Goschke, 2000), to show increased switch cost as a
function of prior trial incongruency. Second, we attempted to demonstrate effects consistent with a non-specific, task-switch-induced slowing that persisted to subsequent trials, regardless of subsequent trial type. This would be found as a slowing in response times due to task switching in the previous trial, regardless of the current trial type. Furthermore, if such slowing due to task switches in the previous trial was found even
when no switch occurred in the current trial, then we would conclude that
the slowing differed from backward inhibition effects, since backward inhibition entails a switch in the current trial to a previously abandoned task set. A positive result would suggest that the cognitive control mechanisms of the Jones et al. (Jones et al., 2002) model, and associated
with the ACC, might also govern the tradeoff between control and prepotency in cognitive as well as motor mechanisms.
Sixteen participants, age 19–22 (9 female) underwent behavioral testing.
These participants came from the Washington University area and were compensated for their participation by being paid $10/hour. 3.2. Tasks and stimuli
Subjects performed a variant of the Rogers and Monsell (1995) letter-digit
paradigm which involved two different tasks performed on visually presented stimuli: consonant/vowel classification of letters and
odd/even classification of digits (see Fig. 1). On each trial, a single
uppercase letter and digit were centrally presented side-by-side in 24-point Times New Roman font, white on a black background. The location of each stimulus type (letter or digit) was random and varied across trials. Letters and digits were selected randomly and with uniform probability from the following two sets: letters: (A, E, I, U, X, P, L, Z); digits: (2–9, inclusive). Classification judgments were indicated via a manual button press with the index finger of each hand. Only two buttons were provided for classification, which produced response overlap across tasks. The mapping of response (odd/even, consonant/vowel) to hand was counterbalanced across participants, but was fixed for a participant across all trials. On any given trial, only a single task was to be performed, and this was indicated by a task-cue presented prior to the onset of the target stimuli. The task cue was the word “LETTER” or “NUMBER” presented visually at central fixation in 24-point Times New
Roman font. Each task cue occurred randomly with 50% probability, leading to an equal proportion of trials in which the current task switched or repeated from the one performed just previously.
Display Full Size version of this image (15K)
Fig. 1. The task-switching task, adapted from (Rogers & Monsell, 1995).
Each trial begins with the cue word “LETTER” or “NUMBER” appearing briefly in the middle of the screen. After a delay during which the screen is blank, a letter and a number appear side-by-side. If the cue word was “LETTER,” the subject must respond differentially for vowels and
consonants. If “NUMBER,” the subject must respond differentially for odd and even numbers.
The timing of each trial was as follows: (1) task-cue presented for 300 ms; (2) constant preparatory interval of 1500 ms starting at cue offset, during which the display went blank; (3) presentation of target stimuli (e.g., “A 2” or “9 P”) immediately following the preparatory interval, with stimulus duration lasting until a response was made or 5000 ms elapsed; (4) constant response-cue-interval (RCI) of 200 ms occurring prior to onset of the next trial.
Testing was performed on a Macintosh G3 computer running PsyScope software. Button press responses were made on the PsyScope button box. After