Elsevier

Neurocomputing

Volume 69, Issues 10–12, June 2006, Pages 1332-1336
Neurocomputing

Computational and neural mechanisms of task switching

https://doi.org/10.1016/j.neucom.2005.12.102Get rights and content

Abstract

Switching between tasks that overlap in perceptual and response characteristics is assumed to rely upon the maintenance of task representations in prefrontal cortex (PFC). However, task-switching studies demonstrate “switch costs,” even when there is sufficient time to prepare for a new task. These costs suggest that task-switching performance reflects a complex interplay between priming and the updating and maintenance of task representations. We describe a computational model in which this interaction is made explicit and linked to the dynamics of PFC. Simulation results account for a variety of empirical phenomena and predict a double dissociation in lateral PFC that was subsequently identified.

Introduction

PFC is hypothesized to subserve the active maintenance of task representations in situations that require rapidly switching between multiple demands [8]. However, one potential problem for this account is that these active maintenance processes appear to be nonoptimal: trials in which the task switches demonstrate increased response times and error rates relative to trials in which the task remains the same (i.e., “switch costs”) [1]. These switch costs persist even when sufficient preparation time is given for a new task [1], [5], [7]. One potential explanation for this finding is that the active maintenance process subserved by PFC is optimal, but is only engaged on a subset of trials in which the task switches. Distributional analyses of response times support this hypothesis. When there is sufficient time to prepare for a task switch, response times appear to come from two stochastic distributions: a prepared distribution of fast trials in which there are no switch costs, and an unprepared distribution of slow response times in which there are severe switch costs [5]. These data suggest that performance on fast trials is governed by a process that effectively suppresses processing of the irrelevant stimulus dimension (i.e., the one associated with the other task), and that the absence of this process on slow trials reveals the lasting effects of previous stimuli (e.g., priming), which are likely subserved by associative learning. To investigate this claim, we developed a connectionist model of task switching that linked the interactions between priming effects and controlled processing to the dynamics of the PFC.

Section snippets

Methods

The model used the connectionist architecture seen in Fig. 1. The simulations were conducted from within the LEABRA framework of the pdp++ software. Model units had a nearly sigmoidal point neuron activation function that maintains key aspects of the electrophysiological process of firing neurons, such as different ion channel types (for details, see [9]).

Activations were calculated for each unit by clamping the inputs at the beginning of each event and allowing activity to propagate through

Results and discussion

In order to evaluate the model's performance, data produced by the network (performance measures as well as PFC activity dynamics) were compared to data produced by 13 participants whose brain activity was measured with functional magnetic resonance imaging (fMRI) while performing a random-cueing task-switching paradigm [3].

The network captured several general phenomena in the task-switching literature. Task switch trials were slower and less accurate than task repeats (13.0 cycles; effect

Conclusions

This model demonstrates that a combination of active maintenance processes and associative learning mechanisms accounts for several relevant behavioral phenomena in the task-switching literature. The model provides a mechanistic account of how PFC (and DA) may contribute to task-switching performance through the active maintenance of task sets, while also capturing a behavioral phenomenon (i.e., “residual switch costs”) that, at first glance, appears to be at odds with this hypothesis. Further,

Acknowledgments

We are grateful to Andy Jones and Nicole Speer for helpful comments and suggestions. Supported by the Office of Naval Research (grant N00014-00-1- 0715) awarded to TSB and a National Defense Science and Engineering Graduate Fellowship awarded to JRR.

Jeremy R. Reynolds is a research associate in the Computational Cognitive Neuroscience Lab at the University of Colorado at Boulder. He received his Ph.D. in Psychology from Washington University in Saint Louis in 2005. His research uses converging methods to understand how the prefrontal cortex subserves controlled behavior.

References (9)

There are more references available in the full text version of this article.

Cited by (22)

  • A unified model of rule-set learning and selection

    2020, Neural Networks
    Citation Excerpt :

    Like many of the other models, the Ardid and Wang model is also subject to combinatorial explosion as the number of category cells required is equal to the product of the responses and the rule-sets. Another approach to rule-set maintenance is used in a model by Reynolds, Braver, Brown, and Van der Stigchel (2006) that simulated the residual switch-cost after priming. In the model, the rule-set is selected by the cue and stored in the PFC. Dopamine probabilistically maintains the rule-set until the stimulus is presented.

  • A neural network model of individual differences in task switching abilities

    2014, Neuropsychologia
    Citation Excerpt :

    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 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.

  • Why won't you do what I want? The informative failures of children and models

    2012, Cognitive Development
    Citation Excerpt :

    This updating function is thought to be subserved by striatal mechanisms; indeed, striatal mechanisms are becoming an important component of network models of higher-level cognition for precisely this reason (Chatham et al., 2011; Hazy, Frank, & O’Reilly, 2007; Kruger & Dayan, 2009; O’Reilly & Frank, 2006; Rougier, Noelle, Braver, Cohen, & O’Reilly, 2005). In addition to these biological and the aforementioned computational reasons for including an updating mechanism in our model, previous work indicates that updating mechanisms may be particularly important for capturing switch cost phenomena in adults (Reynolds, Braver, Brown, & van der Stigchel, 2006). This finding naturally raises the possibility that updating mechanisms are also important for capturing task-switching phenomena in children, such as the one we have observed.

  • Positive affect increases cognitive control in the antisaccade task

    2011, Brain and Cognition
    Citation Excerpt :

    Evidence for this theory originates from studies which have shown that DA agonists that enhance dopaminergic activity strengthen positive affect (Beatty, 1995). Furthermore, there is ample evidence that DA selectively modulates cognitive control processes (Braver, Barch, & Cohen, 1999; Reynolds, Braver, Brown, & Van der Stigchel, 2006). Interestingly, the antisaccade task has been shown to be modulated by dopamine levels in the brain.

View all citing articles on Scopus

Jeremy R. Reynolds is a research associate in the Computational Cognitive Neuroscience Lab at the University of Colorado at Boulder. He received his Ph.D. in Psychology from Washington University in Saint Louis in 2005. His research uses converging methods to understand how the prefrontal cortex subserves controlled behavior.

Todd S. Braver is an associate professor and director of the Cognitive Control and Psychopathology laboratory at the Washington University in Saint Louis. He received his Ph.D. in Cognitive Neuroscience from Carnegie Mellon University in 1997. His research uses multiple methods to understand how humans exert control over their thoughts and behavior.

Josh W. Brown is a research scientist in the Cognitive Control and Psychopathology laboratory at the Washington University in Saint Louis. He received his Ph.D. in Cognitive and Neural Systems from Boston University in 2000. His research focuses on the neural mechanisms of cognitive control in healthy and clinical populations, using integrated computational neural modeling, cognitive psychology, functional neuroimaging, and systems neurophysiology.

Stefan Van der Stigchel received his master's degree in Cognitive Artificial Intelligence in 2003 from Utrecht University, The Netherlands. He is currently pursuing a Ph.D. in Cognitive Psychology at the Vrije Universiteit, Amsterdam. His research interests include the relation between attention and eye movement behavior and computational models of cognitive control.

View full text