Dynamics of Cognitive Control

This capacity for cognitive control allows people not only to pursue goal-directed behaviors over extended periods of time, but also to rapidly and flexibly adapt behavior to novel circumstances in a fast-changing world. This flexibility  requires dynamic adjustments in the allocation of control that we seek to understand through the use of well-crafted psychophysics experiments, neuroimaging studies, and neural network modeling. Where possible, we exploit the mathematical tools of control theory to construct normative models – that is, to explain how control is used to optimize behavior –  and how control allocation is impacted by fundamental computational tradeoffs, such as  the stability versus flexibility of representations and the use of shared (general purpose) versus separated (task-dedicated) representations in neural networks, the explore/exploit tradeoff, and how learning interacts with and shapes the balance between these tradeoffs.  This work is cast within the framework of “bounded optimality,” or “resource rationality,” which assume that the human brain has evolved to optimize its computational functions subject to the constraints imposed by the tradeoffs above, and the constraints imposed by its structure and the world in which it must operate.

An active area of research in the lab involves how cognitive control manages learning. Should you noodle around the guitar today, or do some of those technique exercises? Noodling a bit and bumbling through some tabs will certainly be less effortful than grueling finger exercises. But, what will happen to your skill level in a year’s time of making this choice weekly? You face a tradeoff between immediate reward (playing some simple stuff) with relatively low constant effort, and potentially higher long term reward (playing some complicated stuff) with relatively high effort right now that may decrease over time as you get better. This isn’t an easy intertemporal choice – it requires predicting the relative costs and benefits of your effort over time.

Recent work shows that rats choose the effortful option for the benefits of faster learning and higher future reward, suggesting that they strategically manage their own learning through the use of cognitive control. This general strategy for more effective learning falls out naturally from an important theory of cognitive control allocation developed in the lab when one assumes that an agent is aware that their effort will have an effect on their learning trajectory. This line of research has important implications for general theories of cognition, such as bounded optimality, which proposes that a cognitive system should be evaluated according to its inherent limitations or bounds. An important assumption of these theories is that those bounds are fixed. If I'm not very good at the guitar, my skill level will act as a bound on my performance: no matter how hard I try, I will only be able to play so well. However, I can improve my skill level through learning. Thus, it’s important for general theories of cognition to account for the fact that modifying our cognitive bounds is possible and must be considered when deciding how to allocate mental effort. 

Another major focus in our lab is to develop cognitive models that formalize the role, implementation and dynamics of cognitive control. One part of this endeavor is to understand how control dynamics can account for the stability-flexibility trade-off in attention: the fact that stable focus on one task comes at a cost of slower switching of attention to another task. In our work (Musslick et al., 2018, 2019; Jongkees et al., 2023) we proposed that stronger allocation of control increases the weighting of relevant stimulus inputs in a decision making process (thus improving ‘stability’), but also increases the time necessary to reconfigure control when a shift to a different task is required (thus impairing ‘flexibility’). We formalized these ideas in a dynamical system, where the intensity of control allocation is determined by the gain of task-representing units (Figure 2A). Model simulations show high gain (corresponding to high intensity of control), as compared to low gain, improves stability (indexed by the maximum level of activation of the relevant task unit and minimum level of activation of the irrelevant task unit) but impairs flexible switching (indexed by the time required for task units to settle into the appropriate activation state during a task-switch; Figure 2B, C). Simulations further show the optimal intensity of control (in terms of maximizing response accuracy) depends on how much flexibility is required during a task: when task-switches are frequent the optimal intensity of control is lower, and model fitting of behavioral data indicates people adjust their control intensity in line with this model prediction. These findings provide a mechanistic account of the stability-flexibility trade-off and allow for rational predictions on optimal control intensity as a function of a task’s demand for cognitive flexibility.

Figure 2. Mechanistic modeling of the stability-flexibiltiy trade-off in attention. (A) Effect of gain modulation on the nonlinear activation function of task units, which relates the net input of a unit (e.g. task cues) to its activity state. (B-C) Phase portraits for activity of task units in response to task cues, under low gain (B) and high gain (C), showing trajectory from prior activation of one task (e.g., red dot for task 1) to the currently-relevant task (e.g., blue for task 2). Contour lines and arrows indicate the energy and shape of the attractor landscape after a task-switch from one task to the other.

The importance of the work above was recently underscored by Jongkees et al. (2023), who showed that such models of control dynamics are necessary to dissociate and discriminate between different mechanism of control that could account for changes in behavior. Specifically, we showed adaptations of a speed-accuracy trade-off in responding can be mistaken for adaptations of cognitive flexibility (i.e., control intensity) when focussing just on reaction time (RT) data, which is often a primary outcome measure in cognitive science studies. For example, when attempting to increase speed of responding in a flexibility-demanding situation, people could lower their intensity of control allocation to facilitate cognitive flexibility. Lower intensity of control would make control easier to reconfigure and thereby decrease ‘task-switch cost’, i.e. the RT difference on task-switches and task-repeats. However, a similar decrease in task-switch cost could be achieved not by adjusting control intensity but instead simply by lowering one’s response threshold. As a result, RT on both task-switch and task-repeat trials would decrease but because task-switch RT is generally slower it would decrease more than task-repeat RT. Thus the difference in RT, the task-switch cost, would decrease with lower threshold and create the appearance of increased cognitive flexibility that actually reflects a speed-accuracy trade-off in responding. Using our model that dissociates control dynamics from speed-accuracy trade-offs in responding, we showed that monetary reward for fast and accurate responses, an experimental manipulation typically assumed to motivate an increase in cognitive flexibility, only induced an emphasis on speed at the cost of accuracy. As such, our findings highlight the utility of precise cognitive modeling to determine the mechanisms through which people use control to adapt to task demands.