SEARCH

SEARCH BY CITATION

Keywords:

  • Cognitive control;
  • Representation;
  • Goal neglect;
  • Automaticity;
  • Task difficulty;
  • Representational acuity

Abstract

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Componential control representations
  5. 3. … versus emergent control processes
  6. 4. Oblivious to perfect control?
  7. References

In this commentary, I focus on the difference between processes and representations and how this distinction relates to the question of what is controlled. Despite some views that task switching is a prototypical control process, the analysis concludes that task switching depends on the task goal representation and that control processes are there to prevent goal representations from disintegrating. Over time, these processes become obsolete, leaving behind a representation that automatically controls task performance. The distinction between processes and representations relates to practice effects and automaticity and sheds light on what is meant by the phrase “automatic control.”


1. Introduction

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Componential control representations
  5. 3. … versus emergent control processes
  6. 4. Oblivious to perfect control?
  7. References

The target articles in TopiCS volume 2, issue 4 cover a wide range of approaches to cognitive control and seem to promote task-switching as a prototypical cognitive control process. Nevertheless, the articles are somewhat inconsistent in using the terms “process” and “representation.” Where Stout (2010) hints at “internal representations” being used in control, Cragg and Nation (2010) discuss processes that maintain the integrity of goal representations, Alexander and Brown (2010) describe their theory of how a control representation comes about, and Lenartowicz, Kalar, Congdon, and Poldrack (2010) assume that cognitive control is a cognitive construct that has a neural substrate. Given the varied views, it is critical to appreciate the differences between a control process and a control representation in order to understand the different computational and neural signatures of cognitive control.

2. Componential control representations

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Componential control representations
  5. 3. … versus emergent control processes
  6. 4. Oblivious to perfect control?
  7. References

In every cognitive task conducted in the laboratory and in real-world tasks such as tool-making, the individual has to create or is given a representation of the task. This representation includes the begin-state, the end-state, and can include the relevant operators. In cognitive architectures such as SOAR and ACT-R, these representations consist of a set of IF-THEN statements and end-point states. Connectionist models of such tasks may contain the same representations in the form of a matrix of connection weights, but one can imagine a more dynamic representation (i.e., same weight matrix, but different patterns of activation for different tasks). Instead of developing a theory of representation, I will argue that control representations in and of themselves have limited utility because they are task specific.

I will take the example of task-switching (Monsell, 2003), in which parity judgment alternates with magnitude judgment (smaller/larger than 5), with the position of the stimulus onscreen indicating the judgment required on any specific trial. A goal representation1 must consist of the following rules:

  • (a)
    IF above_line AND number ∈{2, 4, 6, 8} THEN press left-key
  • (b)
    IF above_line AND number ∈{1, 3, 7, 9} THEN press right-key
  • (c)
    IF below_line AND number ∈{1, 2, 3, 4} THEN press left-key
  • (d)
    IF below_line AND number ∈{6, 7, 8, 9} THEN press right-key

I make the following assumptions about the features of (goal) representations:

  • 1
    Goal representations are bindings of their subrepresentations. In the above paradigm, this would require subrepresentations of above_line, below_line, the numbers, and the motor programs for pressing left-key and right-key.
  • 2
    These bindings might be completely novel (as with arbitrary stimulus-response mappings) or already existent (retrieved from memory).
  • 3
    Novel intrarepresentational bindings are fragile. This assumption comes from the rationale that a robust control system should not be able to create new goal representations out of noise. To avoid the creation of inappropriate (or even improbable) goals, the formation of new goals should require some time or effort. This necessarily means that new bindings are fragile.
  • 4
    The entire collection of subrepresentational components is componential, but it can be bound to a single memory trace. This trace may be likened to a multidimensional pointer, which points to the locations of the subrepresentations. Activating this pointer will reinstate the entire collection and thus “upload” the task-relevant information. This pointer is therefore akin to an instance in Logan’s (1988) instance theory and to an event file in Hommel’s (2004) feature binding hypothesis. The ability to learn the multidimensional representation gives rise to the characteristic practice effects (Logan, 1988).
  • 5
    More speculatively, goal representations can be localized in the brain, although this is likely to be statistical with representation X involving a (distributed) neural group Y with mostly activation pattern Z, with changing group membership and activation pattern.

Goal representations can be extracted from the task instructions in a laboratory setting by translating the verbal instructions given by the researcher in a set of IF-THEN rules. A connectionist architecture of the instructions may resemble the connectionist models developed by Cohen and coworkers (Cohen, Dunbar, & McClelland, 1990; Rougier, Noelle, Braver, Cohen, & O’Reilly, 2005). These models (where one node represents an entire group of neurons) consist of a stimulus layer that receives external input and a response layer that activates motor programs for depressing response keys or producing verbal outputs. These two layers are connected via a hidden layer, which may connect all possible stimuli with all possible responses. In the task-switching paradigm, the hidden units and the associated connection weights will encode the aforementioned four IF-THEN rules. As the hidden layer will contain conflicting stimulus-response mapping, many models include a fourth layer that biases some mappings more than others. This is the biased competition model described by Alexander and Brown (2010) and forms a general view of control in models of this kind (Miller & Cohen, 2001). In such models, this fourth layer contains the control representations. The control layer is structured such as to activate or keep active those goal representations that lead to optimal (as defined by the experimenter) performance with minimal conflict. These control dynamics can operate from one trial to the next (e.g., Botvinick, Braver, Barch, Carter, & Cohen, 2001) or within the same trial (Davelaar, 2008). Both between- and within-trial control dynamics have their signatures in response time distributions (Davelaar, 2008; Pratte, Rouder, Morey, & Feng, 2010). It is important to realize that these control representations are goal representations themselves and inherit the features mentioned above for goal representations and thus are componential.

I assert that goal representations, especially novel ones, are fragile and that they are active at different levels. The stronger the activity, the more probable presentation of the predicates will lead to execution of the consequences. Duncan and colleagues (Duncan, Emslie, Williams, Johnson, & Freer, 1996; Duncan et al., 2008) provide evidence for the fragility of goal representations. In their task, participants monitor two streams of digits and letters presented at a fast rate on either side of the center of the computer screen. At the beginning of the trial, the participants receive two instructions. The first informs the participant which side to monitor. The second addresses the possible switching operation toward the end of the stimulus stream when a character appears in between the two streams. This character is either a cue to switch the attention from one to the other side or to stay with the original side. Goal neglect is defined as not switching when cued to do so while the participant has knowledge of the goal. Very old adults, frontal patients, and those with low fluid intelligence do not execute the switch when given the cue, even though they have not forgotten the rule. The phenomenon of goal neglect is easily explained in terms of the goal representation not becoming activated strongly enough to allow goal execution.

3. … versus emergent control processes

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Componential control representations
  5. 3. … versus emergent control processes
  6. 4. Oblivious to perfect control?
  7. References

Apart from representations having a graded activation level, when activation levels are too low, the entire goal representation may disintegrate into its subcomponents. In other words, the participant may forget one of the components of the task (e.g., which key to press). To prevent disintegration, a number of simple processes can be utilized: (a) active maintenance of goal subcomponents, (b) associating the subcomponents with each other, (c) and with a current memory context, and (d) reactivating the subcomponents if they were deactivated. The first, active maintenance, has been recognized as a separate working memory control process (Baddeley, 1996), whereas the second and third suggest a link between memory and cognitive control that underlies the progression from controlled to automatic responding. The final process, reactivation, is akin to the process of (covert verbal) rehearsal. As such, inner speech helps reactivating subcomponents of the goal representation and thus supports cognitive control (Cragg & Nation, 2010). All these processes help in keeping the integrity of the goal representation for as long as is needed or possible, as losing the goal representation will result in increased errors. These processes may be intermittent and most likely do not correspond to a dedicated cluster of neurons.

The formation of goals may be a sensitive period affecting representational integrity. Duncan et al. (2008) showed that when the relevant task instruction was combined with multiple irrelevant instructions, performance was compromised. This suggests that a limited pool of resources contribute to goal representations and that representational acuity is an important determinant of performance. It is noteworthy to highlight the potential of involving stable representations in the formation and (re)activation of goal representations. Cragg and Nation (2010) review the literature on the use of language in cognitive control. In the current context, this review highlights the role that language may fulfill in keeping the subcomponents together. In particular, verbal labels may help to bind the different subcomponents. The role of language in cognitive control (through verbal labeling) is to prevent disintegration of the goal representation and thus fulfills the supporting role of cognitive glue. Does this make language a control process?

The goal representation gets activated by an external stimulus (or an internal response) and has as outcome the activation of an internal (e.g., activating a subgoal) or an external response. In and of itself, this would be sufficient to perform a task adequately and thus no control is needed (unless automatic control is assumed). That is, the goal representation is not the locus of cognitive control. Then what is controlled? If it is assumed that cognitive control is emergent, then what is controlled is either the performance levels and/or the integrity of the goal representation. In other words, cognitive control is emergent when task-nonspecific processes are used to maintain or increase that which is controlled. This resonates with recent work using production systems (Altmann & Gray, 2008; Salvucci & Taatgen, 2008) in which general processes are recruited in cognitive control. Due to the nonspecificity of the processes recruited in control tasks, brain localization of a particular control process, such as “task switching,” will necessarily yield the brain areas that support the task-nonspecific processes.

4. Oblivious to perfect control?

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Componential control representations
  5. 3. … versus emergent control processes
  6. 4. Oblivious to perfect control?
  7. References

Mandik (2010) argues against a sensory theory of control consciousness and suggests that motor commands directly contribute to the subjective feeling of control consciousness. However, the sensory theories that Mandik rejects are external-sensory in nature, still allowing the possibility for internal-sensory theories of control consciousness. That is, control consciousness relates to the awareness of the amount of work that needs to be done in order to do the task. For example, when I verbally repeat the digits one, two, three at a rate of 3 digits/s, while simultaneously doing a manual Stroop task in which I alternate between color “naming” and word “reading,” I may find it very difficult in the first 15 min. After ten 1-h sessions of practice, I will find it much easier. Nothing (external-sensory) has changed to the task or task environment. Yet my sense of having to do work is largest in the first 15 min of the first session. This type of consciousness differs from external-sensory awareness and cannot be reduced to the sensory theories that Mandik (2010) argues against. This sense of effort can be quantified as internal conflict (Botvinick et al., 2001) and has been shown to provide a sufficient vocabulary to combine metacognitive judgments with the sense of difficulty (Davelaar, 2009). Importantly, the need for motor commands to precede cognitive conflict highlights the contribution of motor commands for control consciousness without denying the direct contributions of internal-sensory (or metacognitive) signals.

In this commentary, I have focused mainly on the distinction between processes and representations, as the literature frequently confuses these concepts. I identified as differences that control representations are task specific, fragile in formation, can be bound to a stable memory trace, and can be localized in the brain, and that control processes are task nonspecific, support the integrity of representations, and do not have a specific location in the brain. In the language of Mandik (2010), we are “control conscious” when stabilizing the control representation, but work on “autopilot” when those representations are stable.

Footnotes
  • 1

    . To clarify, I use the term “goal representation” as synonymous to “task representation” and “task set.” I make a distinction between “control representations” and “goal representations,” where “control representations” are a subset of “goal representations.”

References

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Componential control representations
  5. 3. … versus emergent control processes
  6. 4. Oblivious to perfect control?
  7. References