Sequential dependencies in human behavior offer insights into cognitive control

We present a perspective on cognitive control that is motivated by an examination of sequential dependencies in human behavior. A sequential dependency is an influence of one incidental experience on subsequent experience. Sequential dependencies arise in psychological experiments when individuals perform a task repeatedly or perform a series of tasks, and one task trial influences behavior on subsequent trials. For example, in a naming task, individuals are faster to name a word after having just named easy (e.g., orthographically regular) words than after having just named difficult words. And in a choice task, individuals are faster to press a response key if the same response was made on recent trials than if a different response had been made. We view sequential dependencies as reflecting the fine tuning of cognitive control to the structure of the environment. We discuss the two sequential phenomena just mentioned, and present accounts of the phenomena in terms of the adaptation of cognitive control. For each phenomenon, we characterize cognitive control in terms of constructing a predictive model of the environment and using this model to optimize future performance. This same perspective offers insight not only into adaptation of control, but how task instructions can be translated into an initial configuration of the cognitive architecture. INTRODUCTION In this chapter, we present a particular perspective on cognitive control that is motivated by an examination of sequential dependencies in human behavior. At its essence, a sequential dependency is an influence of one incidental experience on subsequent experience. Sequential dependencies arise both in naturalistic settings and in psychological experiments when individuals perform a task repeatedly or perform a series of tasks, and performing one task trial influences behavior on subsequent trials. Measures of behavior are diverse, including response latency, accuracy, type of errors produced, and interpretation of ambiguous stimuli. To illustrate, consider the three columns of addition problems in Table 1. The first column is a series of easy problems; individuals are quick and accurate in naming the sum. The second column is a series of hard problems; individuals are slower and less accurate in responding. The third column contains a mixture of easy and hard problems. If sequential dependencies arise in repeatedly naming the sums, then the response time or accuracy to an easy problem will depend on the preceding context, i.e., whether it appears in an easy or mixed block; similarly, performance on a hard problem will depend on whether it appears in a hard or mixed block. Exactly this sort of dependency has been observed (Lupker, Kinoshita, Coltheart, & Taylor, 2003): responses to a hard problem are faster but less accurate in a mixed block than in a pure Mozer, Kinoshita,& Shettel Sequential Dependencies block; similarly, responses to an easy problem are slower and more accurate in a mixed block than in a pure block of easy trials. Essentially, the presence of recent easy problems causes response-initiation processes to treat a hard problem as if it were easier, speeding up responses but causing them to be more error prone; the reverse effect occurs for easy problems in the presence of recent hard problems. Sequential dependencies reflect cortical adaptation operating on the time scale of seconds, not—as one usually imagines when discussing learning—days or weeks. Sequential dependencies are robust and nearly ubiquitous across a wide range of experimental tasks. Table 2 presents a catalog of sequential dependency effects, spanning a variety of components of the cognitive architecture, including perception, attention, language, stimulus-response mapping, and response initiation. Sequential dependencies arise in a variety of experimental paradigms. The aspect of the stimulus that produces the dependency—which we term the dimension of dependency—ranges from the concrete, such as color or identity, to the abstract, such as cue validity and item difficulty. Most sequential dependencies are fairly short lived, lasting roughly five intervening trials, but some varieties span hundreds of trials and weeks of passing time (e.g., global display configuration, Chun and Jiang, 1998; syntactic structure, Bock, 2002). Sequential dependencies may be even more widespread than Table 2 suggests, because they are ignored in the traditional psychological experimental paradigm. In a typical experiment, participants perform dozens of practice trials during which data is not collected, followed by experimental trials that are randomized such that when aggregation is performed over trials in a particular experimental condition, sequential effects are cancelled. When sequential effects are studied, they are often larger than other experimental effects explored in the same paradigm; for example, in visual search, sequential effects can modulate response latency by 100 msec given latencies in the 700 msec range (e.g., Wolfe et al., 2003). Sequential dependencies are often described as a sort of priming, facilitation of performance due to having processed similar stimuli or made similar responses in the past. We prefer not to characterize sequential dependencies using the term priming for two reasons. First, priming is often viewed as an experimental curiosity used to diagnose the nature of cognitive representations, one which has little bearing on naturalistic tasks and experience. Second, many sequential dependencies are not due to repetitions of specific stimulus identities or features, but rather to a more abstract type of similarity. For example, in the arithmetic-problem difficulty manipulation described earlier, problem difficulty, not having experience on a specific problem, induces sequential dependencies; and in language, syntactic structure induces sequential dependencies, not particular words or semantic content. COGNITIVE CONTROL We view sequential dependencies as a strong constraint on the operation of cognitive control. Cognitive control allows individuals to flexibly adapt behavior to current goals and task demands. Aspects of cognitive control include the deployment of visual attention, the selection of responses, forming arbitrary associations between stimuli and responses, and using working memory to subserve ongoing processing. At its essence, cognitive control involves translating a task specification into a configuration of the cognitive architecture appropriate for performing that task. But cognitive control involves a secondary, more subtle, ability—that of fine tuning the operation of the cognitive architecture to the environment. For example, consider searching for a key in a bowl of coins versus searching for a key on a black leather couch. In the TABLE 1. Three blocks of addition problems EASY BLOCK HARD BLOCK MIXED BLOCK 3 + 2 9 + 4 3 + 2 1 + 4 7 + 6 7 + 6 10 + 7 8 + 6 10 + 7 5 + 5 6 + 13 6 + 13