The Wiley Handbook of Cognitive Control

Tobias Egner

• Publisher: Wiley-Blackwell
• Year: 2017
• Print ISBN: 9781119121442, 9781118920541
• eBook ISBN: 9781118920473
• Edition: 1
• Topic: Biological Sciences
• Subtopic: Neuroscience
• Index: Biological Sciences

Covering basic theory, new research, and intersections with adjacent fields, this is the first comprehensive reference work on cognitive control – our ability to use internal goals to guide thought and behavior.

• Draws together expert perspectives from a range of disciplines, including cognitive psychology, neuropsychology, neuroscience, cognitive science, and neurology
• Covers behavioral phenomena of cognitive control, neuroanatomical and computational models of frontal lobe function, and the interface between cognitive control and other mental processes
• Explores the ways in which cognitive control research can inform and enhance our understanding of brain development and neurological and psychiatric conditions

Section II
Models of Cognitive Control: Computations, Mechanisms, and Neuroanatomy

8
Computational Models of Cognitive Control

Tom Verguts

Introduction

A tennis player is ready to serve his match point, when lightning flashes across the sky. What should he do? Serve the ball? Run for shelter? Or first pick up his kid from the playground? In a situation like this, cognitive control is required to act adaptively. Broadly, cognitive control refers to organising, selecting, or otherwise operating on more basic cognitive or motor processes (e.g., serving a tennis ball). This is particularly, but not exclusively, needed when something goes wrong (e.g., lightning strikes).

Over the last decades, cognitive control has witnessed a remarkable surge of interest in cognitive neuroscience. In parallel, several computational models of cognitive control have been proposed, which help in formulating precise cognitive mechanisms to both integrate existing data and provide theoretical guidance to steer future empirical work. Especially in a multidisciplinary field like cognitive control, characterised by a large and bewildering number of experimental methods, experimental paradigms, studied organisms, and theories, such guidance from computational modelling is indispensable.

I here review some of these models. To structure the review, I use two core ideas that have figured repeatedly in computational models of cognitive control. The first core idea is that cognitive control is a form of hierarchically higher decision making. The lower level of decision making chooses or optimises actions; the higher level of decision making (cognitive control) is concerned with choosing or optimising the brain or cognitive processes that do so. The second core idea is that the process of learning to make such decisions can be modelled using the same computational machinery as lower‐level decision making. One advantage is that the same theoretical frameworks that have been proposed for solving lower‐level problems can be used for cognitive control as well. In this respect, I discuss how three influential modelling frameworks (parallel distributed processing, reinforcement learning, and Bayesian inference) have been used to formulate computational models of cognitive control.

Cognitive Control: Solving Higher‐Level Problems

The first core idea is that cognitive control is about solving problems, but not all types of problems. As long as the sky is blue, the tennis player can keep on playing and leave control to primary controllers. Identifying an object as a house, reading a word, and serving a tennis ball are not the business of cognitive control, but are handled by visual perception, language, and motor controllers, respectively. A single primary controller corresponds to a relatively well‐circumscribed task that can be solved with a single stimulus–action mapping (possibly with hidden representations). An example from actual cognitive control laboratory tasks would be evaluating a number’s parity status or identifying which of two numbers is larger. Cognitive control operates to control such primary controllers when needed (Braver & Cohen, 2000). I will call such hierarchically higher controllers secondary controllers. A class of secondary controllers that are in charge of a similar task (e.g., inhibition) will be called a cognitive control function. It is worth stressing that this is not meant to imply that a function (class of secondary controllers) is localised in one brain area dedicated to this task. It simply identifies one computational task (function). Also, my choice of the cognitive control functions is ‘literature driven’; the selection was made on the basis of which functions are often discussed (and modelled) in the literature. Furthermore, a full hierarchy may involve many more levels than just those two. For example, there are surely controllers below the ‘primary’ level of serving a tennis ball (Haruno, Wolpert, & Kawato, 2001). With these caveats addressed, I find the notions of primary and secondary controllers useful for the current purpose, so I will adopt them.

Cognitive control is needed when a problem arises in primary controllers or the responses they generate (see also Chapter 17 by Ullsperger in this volume). In the example above, simply continuing to run the tennis‐serving controller may no longer be optimal when lightning strikes, and the tennis player might change his course of action and head for the playground. However, this is not always appropriate. When there is no child in the playground, he can opt to energise the currently operating primary controller (tennis serving) and inhibit the perception of dark clouds across the sky, in the hope of quickly winning the match. Hence, every challenge can be confronted in several ways, requiring different secondary controllers. Here I identify five classes of secondary controllers (i.e., cognitive control functions). The first three roughly correspond to the cognitive control functions as identified by Miyake and colleagues (Miyake et al., 2000); the other two involve energising or optimizing a given primary controller in some way (see below for details).

The first cognitive control function is choosing which of two (or more) primary controllers is allowed to steer behaviour (Rogers & Monsell, 1995; Chapter 2 by Monsell in this volume); the tennis player from the example is about to make such a choice. I will refer to this function as choosing between controllers. It has been studied extensively in the task‐switching literature, but that term is too narrow for the current purpose since choosing between controllers does not necessarily involve a task switch if the current (primary) controller remains appropriate. The second cognitive control function concerns inhibition of the current (primary) controller (Aron, Robbins, & Poldrack, 2004; Chapter 6 by Verbruggen & Logan in this volume) or of the (currently prepared) response. The third cognitive control function is to provide context in working memory for primary controllers (Miller & Cohen, 2001; O’Reilly & Frank, 2006).

The fourth cognitive control function is energising the current primary controller. Different implementations for this function have been discussed (Botvinick, Braver, Barch, Carter, & Cohen, 2001). A fifth and final cognitive control function is to choose a learning rate (of a primary controller; Behrens, Woolrich, Walton, & Rushworth, 2007). Changing the learning rate can involve either an up‐ or downregulation depending on the context.

When cognitive control is triggered by an actual problem (e.g., a lightning strike), it is called reactive. However, the problems to which cognitive control responds must not be actually experienced; predicting them may suffice. When cognitive control is triggered by a stimulus predicting that a problem will occur, cognitive control is said to be proactive (Braver, 2012; Chapter 9 by Chiew & Braver in this volume). The tennis player in the example may either comfort his child when it comes running toward him (reactive); or the lightning strike may lead him to foresee a scared child and fetch the child before it starts crying (proactive).

Solving Higher‐Level Problems: Learning, Value, and Inference

I discussed that cognitive control concerns solving higher‐order problems; the next issue is how to solve them. Here the basic modelling strategy is to seek inspiration from the literature on primary controllers (e.g., auditory perception, depth processing, motor control), and how these primary controllers know what to do. This question has resulted in a number of theories and algorithms, with varying emphasis on learning, optimisation, or inference. Over the last several decades, three modelling traditions in particular have been very influential.

Modellers in the parallel distributed processing (PDP) tradition emphasise that processing occurs in parallel between several (distributed) controllers. These controllers know what to do from a gradual approach toward an optimal point of a goal function (e.g., the minimum of an error function). This gradual process is called learning (Rumelhart, Williams, & Hinton, 1986; Seidenberg & McClelland, 1989). The parameters that change (i.e., are optimised) during learning are typically weights in a neural network (controller) in this tradition. The second modelling tradition is reinforcement learning (RL): It emphasises the optimisation of value functions to obtain a goal (e.g., finding the maximum of a value function; Montague, Dayan, & Sejnowski, 1996; Sutton & Barto, 1998). These value functions combine the expected reward (reinforcement) and cost attached to available stimuli or actions. The parameters that change during optimisati...

Table of contents

• Cover
• Title Page
• Table of Contents
• Notes on Contributors
• Acknowledgments
• Section I: The Basics of Cognitive Control
• Section II: Models of Cognitive Control
• Section III: Cognitive Control in Context
• Section IV: Cognitive Control in Practice
• Index
• End User License Agreement