Proactive control and episodic binding in context processing effects

Abstract

Decision making often involves using prior contextual information to evaluate relevant events. A laboratory equivalent of context processing situations, the AX-Continuous Performance Task (AX-CPT), yields errors when a target probe appears in a nontarget context, or when a nontarget probe appears in a target context. According to goal-driven accounts, context information is used for top-down preparation towards processing of context-appropriate information. Performance costs in nontarget trials are attributed to inefficient cognitive control. In contrast, the episodic binding account predicts that prior experiences with combinations of a context, probe, and response are bound in episodic memory and thus bias future actions. Performance costs in this view are caused by the need to overrule prepotent associations. The current experiments tested the relative importance of proactive cognitive control versus episodic learning of appropriate stimulus–response relations for CPT performance. Support was obtained for both contributions.

Introduction

Maintaining the contents of working memory (WM) and adjusting it to the task at hand involves cognitive control. By cognitive control we refer to the properties of the cognitive system that emerge as we configure basic cognitive processes in accordance with instructions, intentions and (rapidly changing) environmental demands. For example, a car driver approaching a crossroad has to organize task goals (getting to a specific destination before a specific time), task representations (traffic rules and car operating principles), task context (recent and current road signs; and road map information constraining the possible routes from A to B), and incoming sensory information (current traffic and position relative to destination). Performing such a complex task involves top-down proactive control (anticipating and preparing for certain situations and events) as well as reactive control (flexible adaptive responding to unanticipated situations and events). In addition, such tasks involve the formation of episodic associations between temporally adjacent relevant events. The stronger the association between specific events, the more reliably they can be used for routine-like control over behavior, but also the more difficult it will be to overcome these associations when the current situation calls for a different behavior. In the current study, the importance of episodic binding vis-à-vis proactive control processes in activating the task context will be tested in a continuous performance task (CPT).

The importance of proactive control processes in the maintenance and updating of WM content is largely beyond controversy (e.g., Miller & Cohen, 2001). Failures to update the current task goal are reflected in perseverative behavior under conditions that actually require a switch from one set of task rules to another, as in set shifting experiments (Jersild, 1927) or in the Wisconsin Card Sorting Task (Grant and Berg, 1948, Milner, 1963). Failures to implement and maintain a goal-driven decision bias may lead to distractibility, for example in the Stroop color-word task (Stroop, 1935) or the Eriksen flanker task (Eriksen & Eriksen, 1974) if task-irrelevant information captures attention and elicits an inappropriate response.

Failures to maintain or update WM content are also illustrated by performance errors in the AX-CPT (e.g., Braver & Barch, 2002), a variant of the classic CPT (Rosvold, Mirsky, Sarason, Bransome, & Beck, 1956) and focus of the current study. Braver and colleagues (Braver and Barch, 2002, Braver et al., 2005) used the term context representation to refer to representations that influence planning, behavior and attentional processes. The context can contain task instructions, information from prior stimuli, or intended actions. It configures the cognitive system for the performance of challenging and nonroutine tasks.

According to the context processing model (Braver and Barch, 2002, Braver et al., 1999, Braver et al., 2001), context information is part of the representations in WM. For each new incoming contextual stimulus (referred to here as context cues) the model changes its current context representation, which helps in interpreting imperative stimuli (referred to here as probes). Top-down control can be exerted because the context cue (as currently represented in prefrontal cortex, PFC) biases or primes the activation of a response or goal, as previously associated with that particular cue, in succeeding trials. On the other hand, pre-existing associations between specific probe stimuli and specific responses can create a bottom-up bias in responding. Some probe stimuli are for example more strongly associated to one action, while others have become associated more strongly to another action. When context updating or maintenance is not intact, response activation may be driven more strongly by these probe–response associations than by context cues. The hypotheses generated by the context processing model have been tested using the AX-version of the CPT (AX-CPT, Braver and Barch, 2002, Braver et al., 1999, Braver et al., 2001).

AX-CPT paradigm: the AX-CPT paradigm is a modified version of the classic CPT (Rosvold et al., 1956). During each AX-CPT trial, participants are presented with a sequence of stimuli containing a context cue (A or B) and a probe (X or Y) on the computer screen. They have to respond to a target probe (X) with a manual response on the keyboard, the target response key, but only when the target probe is immediately preceded by a specific context cue (A). In every other case, for example in AY, BX or BY sequences, they have to respond to a probe with a nontarget response key. AX trials occur very frequently during the experiment to induce a strong tendency to make a target response to the X-probe.

Goal-driven bias and CPT: performance errors in the AX-CPT have been interpreted commonly in terms of impulsivity, attention, or inhibition (Halperin et al., 1994, Riccio et al., 2001). Target trials (AX) typically occur in the majority of the trials in the AX-CPT, and this probability induces a strong bias to issue a target response, even in trials other than AX (BX and AY).

According to the context processing model, context information must therefore be used in these trials to give the appropriate nontarget response and overcome the bias to activate the target response. Braver et al. (Braver and Barch, 2002, Braver et al., 1999, Braver et al., 2001) hypothesized that in intact context processing (i.e., representation, updating and maintenance of context) AX-CPT performance is better (faster and more accurate) in AX trials than in AY and BX trials. In AY trials, subjects incorrectly expect a target probe to appear after an A-context cue and are thus inclined to respond with an incorrect (target) response (false alarm). If performance would be entirely goal-driven or rule-based, performance would be hampered exclusively in AY trials, because in BX trials intact context maintenance would result in correct rejection. However, performance in BX trials is often hampered by the X-probe which is strongly associated with an A context and target response. Thus, performance costs in BX trials may be due to inefficient proactive preparation or the result of strong stimulus–response associations, or both. The present study sought to evaluate the contribution of goal-driven control and episodic item-specific bindings in an AX-CPT.

Instead of via proactive preparation, a decision can be biased by currently available stimulus information that reactivates previously associated information from episodic memory. The associative account has recently received some support; performance costs attributed to top-down control may be confounded with episodic effects (e.g., Mayr et al., 2003, Nieuwenhuis et al., 2006), thus it seems relevant to consider whether they may play a role in the AX-CPT as well.

The feature integration account advocated by Hommel, 2004, Hommel et al., 2004 emphasizes on a bottom-up (stimulus-driven) influence and the effect of stimulus–stimulus or stimulus–response associations on current performance, rather than on a proactive control bias. The binding account proposes that on every occurrence of a perceived event, task-relevant or salient information (i.e., certain stimulus and response features) is temporarily bound into an episodic memory trace, the so-called event file (Hommel, 2004). If a feature of the encoded event is encountered on a subsequent occasion, this feature reactivates the associated features from the previously created event file. Performance can be hampered in succeeding trials if only one feature of an event file is reactivated and the other is not. This is the case, for instance, when a stimulus feature is repeated, but the associated response (as encoded in the event file) is incompatible with the currently correct response. The elements of the event file then have to be unbound or rebound which takes time and is error-prone. As a result, performance is worse compared to a situation in which both stimulus and action features are repeated or changed (Hommel, 1998, Hommel, 2004).

Recently, both goal-driven accounts and stimulus-driven accounts have been shown to explain behavioral performance costs in several experimental paradigms (for a review see Egner, 2007, Verguts and Notebaert, 2008). Studies on negative priming (Huang et al., 2004, Tipper, 2001), inhibition (Verbruggen, Logan, Liefooghe, & Vandierendonck, 2008), task switching (Waszak, Hommel, & Allport, 2004) and spatial incompatibility (Hommel et al., 2004) indicated that the reactivation of competing information as a result of a retrieved episodic binding created in a previous trial, biases performance in the current trial.

Stimulus-specific episodic bindings in the AX-CPT may arise between context cue stimuli, probe stimuli, and responses, and can affect performance in subsequent trials in the task. Thus, the binding account may give rise to a different explanation for the costs associated with AY and BX pairs in the AX-CPT than an entirely proactive goal-based account would. The features A, X, and target response become strongly associated with each other because of their prevalent occurrence. When an AY pair is presented incidentally, the A-context cue activates the representations of the X-probe stimulus and the target response; unbinding and re-binding are then either slow or unsuccessful. Likewise, when a BX pair is presented incidentally, the X-probe will activate the associated representations of the A-cue and the target response; again, unbinding and re-binding may either fail or consume time.

The aim of the present study is to establish whether, in addition to proactive goal-driven biases, stimulus-specific mechanisms can explain part of the variance in performance in the AX-CPT. One possibility is that a goal-driven bias will influence AY and BX performance without any effect of bindings between cue, probe, and response on performance. Alternatively, the effect of event binding might fully explain the variance in AY and BX performance. A third scenario might be that goal information as well as the influence of episodic bindings each partially accounts for the AY and BX effects.

To investigate these differential accounts, the traditional use of characters (A, B, X and Y) was replaced by the use of words (cues A, B) and pictures of faces (probes X and Y). Word and face stimuli should allow participants to distinguish between occurrences of specific context cues and probes in individual trials and to create separate event files for separate combinations of features. Previous research by Colzato, Raffone, and Hommel (2006) showed that the bindings between stimulus features are most powerful when using real-life pictures as compared to arbitrary feature conjunctions.

Subjects were asked to respond with a specific response based on the features of cues (words; presented uppercase or lowercase) and probes (pictures of faces; male or female). For example, an uppercase word succeeded by a female face asked for a target response.

A purely proactive goal-driven bias would predict that performance in the AX-CPT relies on the correct representation and maintenance of rule information, predicting more errors in AY trials than in AX trials. Note that the context processing model does take into account that associations between probe and target response drive performance which might lead to additional errors in BX trials. However, with intact context processing, they are thought to occur less often than AY errors.

In terms of the current stimuli, a participant may be instructed to give a target response to the picture of a female face preceded by a word in UPPERCASE script (AX trial). Because uppercase words are frequently followed by female faces, the words induce bias for a target response, also if they are followed by a male face (AY trial). Because female faces often require a target response, these faces induce a bias for a target response even if it is preceded by a lowercase word (BX trial).

A stimulus-specific binding account predicts that every occurrence of a context cue and a concurring target feature is bound into an event file. When in subsequent trials one of these features must be bound to a competing feature code it takes time to build a new association. The costs of processing a probe that deviates from a dominant association are investigated by comparing two types of AY and two types of BX trials that differ in association strength between context cues and probes. That is, some stimuli are presented in consistent combinations, thereby leading to strong binding and associated bias, and others in random combinations, thus leading to no bias. Table 1 shows trial examples for the main contrasts of interest.

When a specific target probe is shown consistently in the context of a particular word [an AX trial, FIRE –female face 1], then the specific context cue and probe are represented as a unitary episodic memory trace. If on a later occasion this particular context cue is succeeded by a photo of a context-inappropriate Y-probe, [FIRE –male face], then this context cue has to be unbound from the association with the female face and target response. This is time-consuming and prone to failure, explaining the AY costs. Other context cues, however, from regular AY cue–probe pairs [TABLE –male face], are never succeeded by a context-appropriate target probe [female face] throughout the experiment, thus no unbinding of the context cue from a previously associated probe is necessary and performance in these trials should not be time-consuming or error-prone.

Thus, a stimulus-specific binding account predicts that AY costs occur only for A-context cues that are associated consistently with a particular X-probe [A_xY trials], but not for those context cues that are paired nonsystematically to varying context-inappropriate Y-probes [A_yY trials]. Recall that if decisions entirely depend on proactive goal-driven preparation, these types of context cues are indistinguishable and AY costs are equivalent in both cases.

BX costs can be explained similarly; the binding account predicts that BX costs occur only for X-probes that are associated consistently with particular A-context cues [BX_a trials], but not for those probes that are paired inconsistently to varying B-context cues [BX_b trials]. Consider again an AX pair [WATER –female face 2]. If on a later occasion female face 2 is presented in an inappropriate B-context train, then this probe has to be unbound from the association with the context WATER which is time-consuming and error-prone. Other probe pictures are never preceded by A-context cues, but always by varying B-context cues, for example [house –female face 3]. Throughout the experiment female face 3 was never presented in an A-context WATER, thus the X-probe was never bound together with an A-context cue or target response. Hence, no time-consuming or error-prone unbinding costs are predicted in these BX_b trials. Again, if performance is merely goal-driven B-cue information will override any type of X-probe and thus performance will be equivalent across all types of BX pairs.

The current study consists of two experiments. Experiment 1 investigated the contribution the goal-driven and stimulus-driven biases in an adapted AX-CPT. To test whether binding effects would be subject to decay, the cue–probe inter-stimulus interval was varied between subjects. Experiment 2 was employed to replicate and extend the results of Experiment 1 with a more difficult version of the AX-CPT, which was accomplished by introducing a distracter between a cue and a probe.