1
Cognitive Control: Core Constructs and Current Considerations

Jonathan D. Cohen

The capacity for cognitive control is perhaps the most distinguishing characteristic of human behaviour. Broadly defined, cognitive control refers to the ability to pursue goal-directed behaviour, in the face of otherwise more habitual or immediately compelling behaviours. This ability is engaged by every faculty that distinguishes human abilities from those of other species, and in virtually every domain of human function from perception to action, decision making to planning, and problem solving to language processing. Understanding the mechanisms that underlie our capacity for cognitive control seems essential to unravelling the mystery of why, on the one hand, we are capable of intelligent, goal-directed behaviour, whereas on the other hand this ability seems so vulnerable to irrational influences and failure. Not surprisingly, the distinction between controlled and automatic processing is one of the most fundamental and long-standing principles of cognitive psychology. However, as fundamental as the construct of cognitive control is, it has been almost equally elusive. Most importantly, the construct on its own says little about the mechanisms involved.

Fortunately, in the half-century since the concept of control was first introduced into psychology (Miller, Galanter, & Pribram, 1960), and afforded a central role in cognitive psychology not long thereafter (Posner & Snyder, 1975; Shiffrin & Schneider, 1977), considerable progress has been made in characterising cognitive control in more precise and mechanistic terms, at both the psychological and neurobiological levels of analysis (e.g., Anderson, 1983; Botvinick & Cohen, 2014; Collins & Frank, 2013; Daw, Niv, & Dayan, 2005; Dayan, 2012; Duncan, 2010; Koechlin & Summerfield, 2007; Miller & Cohen, 2001; O'Reilly, 2006). Much of this progress is reflected in the chapters of this volume. Needless to say, however, considerable progress remains to be made, and formulating a way forward may benefit by revisiting, and carefully reconsidering some of the foundational ideas that originally motivated the construct of cognitive control, and how these have evolved over the past half-century.

In this introduction, I review the original formulations of the distinction between controlled and automatic processing, the issues that this distinction raised, how these have been addressed, and questions that remain. This review is intended to be useful in at least two ways: (a) as a guide to the central constructs and most pressing issues concerning cognitive control for those who are new to this area of research; and (b) as an inventory of challenges that a satisfying account of cognitive control must address for those who are familiar with the area. I have organised the issues into three broad categories: (a) core, defining features of cognitive control; (b) the relationship of cognitive control to other closely related constructs in psychology; and (c) ways in which the understanding of cognitive control may be informed by theoretical approaches that have proved valuable in other areas such as computer science.

Core Constructs

Definitional Attributes of Controlled Versus Automatic Processing

As a theoretical construct, cognitive control grew out the study of communications and control systems, including the discipline of cybernetics that flourished in the middle of the last century (e.g., Wiener, 1948). In particular, an influential book by Miller et al. (1960) explicitly drew the connection between control theory, the goal-directed nature of human cognition, and its apparently hierarchical structure-topics that have regained attention in modern research (as will be discussed below). However, three articles are generally credited with operationalising the construct of cognitive control, and placing it at the centre of experimental research in cognitive psychology: one by Posner & Snyder (1975), and a pair by Shiffrin and Schneider (Schneider & Shiffrin, 1977; Shiffrin & Schneider, 1977). These articles focused on three attributes that distinguish controlled from automatic processes: (a) controlled processes are slower to execute; (b) are subject to interference by competing automatic processes; (c) and rely on a central, limited-capacity processing mechanism.

The canonical example chosen by Posner & Snyder (1975) to illustrate this was a comparison of colour naming and word reading in the Stroop task (MacLeod, 1991; Stroop, 1935). Adults are almost universally faster to read a word out loud than to name the colour of a stimulus (criterion 1). Critically, when responding to incongruent stimuli (e.g., the word 'RED' displayed in green), the colour of the stimulus has almost no impact on the word reading response, whereas the word invariably interferes with naming the colour. Furthermore, attempts to name the colour while performing another unrelated task (such as mental arithmetic) are likely to be impaired. These properties generally do not apply to word reading. These findings were explained by proposing that colour naming is a controlled process, whereas word reading is automatic. This account of findings in the Stroop task quickly became-and in many areas still remains-a foundational paradigm for studying controlled and automatic processing (for example, the same principles are used to infer the influence of automatic processes using the Implicit Association Task-IAT; Greenwald & Banaji, 1995). However, almost as soon as the construct of controlled processing was introduced, it raised concerns.

Capacity Constraints

Central, limited-capacity mechanism. Perhaps the most important and controversial assertion was that cognitive control relies on a central, limited-capacity processing mechanism that imposes a serial constraint on the execution of controlled processes, as distinct from automatic processes that can be carried out in parallel.1 The importance of this assumption cannot be overestimated. The idea was paradigmatic in the literal sense. It provided the operational criterion that is used almost universally to identify a process as control demanding: dual-task interference. If performance of a task suffers when another task that is unrelated (i.e., does not involve the same stimuli or responses) must be performed at the same time, then the first task is deemed to be control demanding. However, as practically-and introspectively-appealing as this assertion is, it is equally problematic.

The capacity constraints on control are generally attributed to its reliance on a limited resource; however, neither the nature of the resource, nor the reason for its limitation has yet been identified. Some have argued that the resource may be metabolic (see discussion of effort below). However, there is no reliable evidentiary basis for this (Carter, Kofler, Forster, & McCullough, 2015), and it seems improbable given the importance of the function, the metabolic resources available to the brain, and the scale on which it is able to commit metabolic resources to other processes.

Another suggestion is that the limitation is structural. For example, most models of cognitive control propose that control-demanding processes rely on the activation and maintenance of control representations that are used to guide execution (see discussion below). These representations are considered the 'resource' upon which control relies, and the limited capacity of control is attributed to a limitation in the scope of such representations that can be actively maintained (e.g., a limitation in working memory capacity). However, this begs the question: Why is that capacity limited? One possibility is a physical limitation (akin to the limited number of memory registers in a CPU). However, like metabolic constraints, this seems highly improbable. There are 100 billion neurons in the human brain, of which about one third reside in areas thought to be responsible for cognitive control (e.g., the prefrontal and dorsal parietal cortex). With those resources available, evolution would have to be viewed as a poor engineer to be incapable of maintaining more than a single control representation at a time. Another possibility is that there are functional constraints on the system; for example, the number of representations that can be simultaneously maintained in an attractor system, or a tension between their number and resolution (Edin et al., 2009; Ma & Huang, 2009; Usher, Cohen, Haarmann, & Horn, 2001). Such efforts reflect important progress being made in developing quantitative, mechanistically explicit accounts of representation and processing in neural systems, and may well explain constraints within circumscribed domains of processing. However, once again, this begs the question: Why cannot a system as vast as the human neocortex proliferate attractor systems for a function as valuable as cognitive control?

Multiple resources hypothesis. An alternative to the idea that dual-task interference reflects a constraint in the control system itself is the idea that, instead, it reflects something about the processes being controlled. This idea has its origins in multiple resource theories of attention (Allport, 1980; Logan, 1985; Navon & Gopher, 1979; also see Allport, Antonis, & Reynolds, 1972; Wickens, 1984). They argued that interference between tasks may reflect cross-talk within local resources (e.g.,...