Computational and neural mechanisms of task switching

Jeremy R. Reynolds, Todd S. Braver, Joshua W. Brown, Stefan Van der StigchelNeurocomputing 2006 · DOI

TL;DR

This paper presents a computational model that explains how the brain switches between tasks, even when there's enough time to prepare. It proposes that task switching involves a dynamic interplay between actively maintaining task information in the prefrontal cortex (PFC) and the lingering effects of previous tasks (priming), mediated by dopamine. The model successfully predicts behavioral patterns and a specific pattern of brain activity in the PFC, offering a unified account of cognitive control.

What Problem Does This Paper Address?

The ability to switch between different tasks is a cornerstone of flexible, goal-directed behavior. Neuroscientific theories often attribute this capability to the prefrontal cortex (PFC), which is thought to actively maintain task representations. However, a persistent puzzle in the field is the existence of "switch costs" – slower reaction times and increased errors on trials where the task changes compared to trials where it stays the same. These costs remain even when participants have ample time to prepare for the upcoming task. This suggests that simply "holding" a task representation in mind isn't enough, and that other factors, like the influence of past experiences (priming), play a crucial role. The authors aim to reconcile these observations by proposing a model that explicitly links the dynamics of PFC activity, the influence of dopamine, and the impact of associative learning (priming) to explain task switching performance.

Methods

The researchers developed a connectionist model, implemented within the LEABRA framework of the pdp++ software. This model uses units with sigmoidal activation functions, mimicking the behavior of neurons with different ion channel types. Key to the model's operation is a simulated "gating signal" that represents a phasic dopamine (DA) response. This signal, occurring randomly on 50% of cue events, is hypothesized to enable active maintenance of task representations in the PFC during the delay period. If the gating signal occurs, task information is actively maintained; if it doesn't, the information decays, leaving the network unprepared. Associative learning, modeled using a Hebbian learning rule, strengthens connections between co-active units, contributing to priming effects. The model's weights are also subject to decay over trials.

Datasets

The paper does not explicitly mention training datasets. For evaluation, the model's performance was compared to data from 13 human participants who performed a random-cueing task-switching paradigm while their brain activity was measured using functional magnetic resonance imaging (fMRI).

Key Insights and Ideas

  • Dopamine as a Gatekeeper for Active Maintenance: The model proposes that phasic dopamine acts as a crucial "gate" for the PFC to actively maintain task representations. This explains why task switching can be "unprepared" even with sufficient cue time – the dopamine signal might simply not occur.
  • The Interplay of Priming and Control: Switch costs aren't solely due to a failure of active maintenance. Instead, they arise from a complex interaction where associative learning (priming) can bias processing, and active maintenance (when engaged) can override these biases.
  • Two Distributions of Performance: The model accounts for the observation that response times in task switching often come from two distributions: fast trials (where control is effective, suppressing priming) and slow trials (where control is less effective, and priming has a stronger influence).
  • Predicting a Double Dissociation in PFC Activity: A significant outcome of the model was its prediction of a specific pattern of brain activity in lateral PFC. This prediction, concerning differential activity during cue/delay versus target periods depending on trial speed, was later confirmed by empirical fMRI data.

Key Results

  • The model successfully replicated general task-switching phenomena, showing slower and less accurate performance on switch trials compared to repeat trials.
  • Switch costs in the model were comparable to those observed in human participants (e.g., 13.0 cycles vs. 74 msec).
  • The model demonstrated a switch × response repetition interaction, with greater switch costs when the correct response repeated.
  • Crucially, the model predicted and the empirical data confirmed a double dissociation in lateral PFC activity: "prepared" (faster) trials showed greater delay-period activity (reflecting active maintenance), while "unprepared" (slower) trials showed greater target-period activity (reflecting re-activation due to priming).
  • Even without active maintenance, the model performed above chance, highlighting the role of associative learning in task switching.

Limitations and Future Work

The authors acknowledge that the model's parameters were adjusted to fit empirical data, and future work could explore how these parameters might be learned. They also suggest that the model could be extended to investigate other aspects of cognitive control, such as the role of different PFC subregions or the impact of more complex task structures. The current model focuses on a simplified task switching paradigm, and its applicability to more complex, real-world scenarios remains an area for future exploration.

Summary updated

BibTeX
@article{reynolds2006computational,
  author = {Jeremy R. Reynolds and Todd S. Braver and Joshua W. Brown and Stefan Van der Stigchel},
  title = {{Computational and neural mechanisms of task switching}},
  journal = {Neurocomputing},
  year = {2006},
  doi = {10.1016/j.neucom.2005.12.102},
}
Figure