Individual differences in the Simon effect are underpinned by differences in the competitive dynamics in the basal ganglia: An experimental verification and a computational model

Abstract

Cognitive control is thought to be made possible by the activity of the prefrontal cortex, which selectively uses task-specific representations to bias the selection of task-appropriate responses over more automated, but inappropriate, ones. Recent models have suggested, however, that prefrontal representations are in turn controlled by the basal ganglia. In particular, neurophysiological considerations suggest that the basal ganglia’s indirect pathway plays a pivotal role in preventing irrelevant information from being incorporated into a task, thus reducing response interference due to the processing of inappropriate stimuli dimensions. Here, we test this hypothesis by showing that individual differences in a non-verbal cognitive control task (the Simon task) are correlated with performance on a decision-making task (the Probabilistic Stimulus Selection task) that tracks the contribution of the indirect pathway. Specifically, the higher the effect of the indirect pathway, the smaller was the behavioral costs associated with suppressing interference in incongruent trials. Additionally, it was found that this correlation was driven by individual differences in incongruent trials only (with little effect on congruent ones) and specific to the indirect pathway (with almost no correlation with the effect of the direct pathways). Finally, it is shown that this pattern of results is precisely what is predicted when competitive dynamics of the basal ganglia are added to the selective attention component of a simple model of the Simon task, thus showing that our experimental results can be fully explained by our initial hypothesis.

Introduction

Cognitive control depends on the capacity to overcome prepotent behavioral responses that interfere with those required by internal goals. In many cases, the successful resolution of this interference depends on selective attention, that is, the capacity to ignore certain features of a stimulus and instead focus on other characteristics, selected on the basis of internal task goals. However, the exact nature of the neural mechanism for resolving interference is still debated. In fact, multiple mechanisms might be recruited for cognitive control, depending on the nature and demands of the task at hand. In this paper, we suggest that the resolution of interference in a task that requires cognitive control relies on the activity of the basal ganglia, a subcortical circuit believed to be responsible for selecting which sensory information is ultimately transmitted to the prefrontal cortex (PFC). We then present an empirical test of this hypothesis and a computational model that accounts for the data presented herein.

The remainder of this paper is structured in four parts. First, we provide a brief overview of the problem of cognitive control in response selection interference. Second, we introduce the basal ganglia circuit, its role in response selection, and a behavioral task that measures the competitive basal ganglia dynamics to resolve conflict. Third, we present an experiment that demonstrates the existence of a significant (and hitherto unsuspected) correlation between performance on a cognitive control task (the Simon task: Craft and Simon, 1970, Simon, 1990) and a task that measures competitive basal ganglia dynamics (the Probabilistic Stimulus Selection task: Frank, Seeberger, & O’reilly, 2004). Lastly, we present a computational model that provides an explicit and mechanistic account of interference in the Simon task in terms of basal ganglia dynamics, which accounts for our results.

Cognitive control is typically studied through tasks that require managing interference between competing responses. For example, in the Stroop task, participants are asked to say out loud the name of the color a word is presented in while ignoring the word itself. When the word is the name of a different color (e.g., RED printed in blue), interference arises because of the tendency to name the word. This interference is manifested in longer reaction times to such incongruent trials, as opposed to congruent ones where no interference exists (e.g., RED printed in red). Another example is the Simon task (Craft and Simon, 1970, Lu and Proctor, 1995, Proctor and Lu, 1999, Simon, 1990), in which participants are asked to respond with their left and right hand to specific visual features (e.g., shape or color) of a stimulus that appears on a screen. For example, they might be asked to respond with their left hand when the stimulus is a square, and with their right hand if the stimulus is a circle (Fig. 1). Interference occurs when a stimulus is presented on the side of the screen that is contralateral to the desired response. As a result, these “incongruent” trials (Fig. 1C and D) are less accurate and take longer than “congruent” trials in which the stimulus appears on the side of the desired response (Fig. 1A and B). This extra time is supposed to reflect the additional cost necessary to resolve conflict generated by the activation of two competing responses (one for the shape, one for the position).

The precise source of interference in these tasks has been much debated and might vary across different paradigms (Liu et al., 2004, Nee et al., 2007, van Veen and Carter, 2005). In the specific case of the Simon task, interference likely occurs early on, at the moment in which the relevant and irrelevant features of the stimuli are being processed, as evidenced from both a review of the behavioral data (Lu & Proctor, 1995) and from the onset of brain oscillations in the fronto-parietal network that reliably indicate individual differences in working memory encoding (Gulbinaite, van Rijn, & Cohen, 2014).

While different authors might disagree on the source of interference for specific tasks, they tend to agree that, at the neural level, interference is resolved through mechanisms underpinned by the prefrontal cortex (PFC: Miller, 2000, Miller and Cohen, 2001). More specifically, by holding a representation of the intended goal (for instance, paying attention to the shape of the stimulus, rather than its position), PFC exerts a top-down influence that ultimately counters the prepotency of unwanted responses. The role of PFC in exerting this form of control has been verified in numerous imaging studies (Cole et al., 2013, Cole and Schneider, 2007, Kane and Engle, 2002, Koechlin et al., 2003, MacDonald et al., 2000, Miller, 2000, Ridderinkhof et al., 2004), proposed for both proactive and reactive control (Braver, 2012, Braver et al., 2009) and implemented in numerous neurocomputational models (Botvinick and Plaut, 2004, Cohen et al., 1996, Cohen et al., 1990, Herd et al., 2006). Within neural network models, this mechanism is typically implemented by adding specific “task” or “goal” representational units at the top of the network hierarchy. The activation spreading from these units ultimately provides the necessary boost to overcome the interference caused by irrelevant stimuli that are encoded by units whose synaptic weights are normally much stronger. For example, in the early influential model of the Stroop task by Cohen et al. (1990), different sets of units encode the stimulus’ word and the stimulus’ color, and both are connected to the output layer of the network, whose units represent color names. Stroop interference originates because the connectivity between input word units and output color names is much stronger than that between input color units and output color names. The model can successfully overcome interference, however, when additional input units are added that encode the task goal (that is, whether to name the color or read the word).

For such a mechanism to function properly, PFC must be able to maintain active representations of the relevant information, while at the same time discarding irrelevant information that would lead to a top-down activation of the irrelevant or incorrect responses. Indeed, single-cell recordings show that PFC neurons exhibit precisely this type of selectivity. For example, PFC neurons were found to encode only the relevant information of the stimulus (the location of a target) and not other aspects (i.e., the location of distractors; Rainer, Asaad, & Miller, 1998). Conversely, accidental encoding of irrelevant information in PFC is associated with poor performance in cognitive control tasks. For instance, compared to individuals with lower cognitive capacity (as indexed by working memory measures), individuals with higher cognitive capacity exhibit smaller event-related potentials when asked to remember the location of a visual target stimulus that is surrounded by distractors, suggesting that better cognitive abilities are associated with a greater capacity to ignore irrelevant information (Vogel, McCollough, & Machizawa, 2005). In turn, this suggests a tight relationship between PFC, working memory, and selective attention, whereby working memory depends on the successful encoding of information in PFC, which ultimately relies on selective attention to optimally allocate PFC resources to the relevant characteristics of the stimuli (Awh et al., 2006, Engle et al., 1999, Kane and Engle, 2002, Kane and Engle, 2003, McNab and Klingberg, 2008).

The theory that interference is resolved through top-down representations provides an elegant solution to the problem of cognitive control and has been largely confirmed, but it poses a second problem: How are these representations selected? In the past two decades, much evidence has accumulated suggesting that the access to PFC is largely regulated by the basal ganglia, a set of interconnected subcortical nuclei (Fig. 2) that receive converging projections from the entire cortex, but project selectively to the frontal lobes, thus operating as a “funnel” through which cortical signals must compete for access to PFC (Albin et al., 1989, Alexander et al., 1986, DeLong, 1990, Kemp and Powell, 1971). Because of their peculiar anatomy, is has been repeatedly suggested that one of the functions of the basal ganglia is to select among competing signals that need to be relayed (through the thalamus) to PFC (Gurney et al., 2001, Redgrave et al., 1999, Stephenson-Jones et al., 2011). Although this function was originally proposed to describe the selection of motor programs (Redgrave et al., 1999), more recent neurocomputational models have suggested that the same mechanism can be generalized to cognitive processes (Frank et al., 2007, Houk et al., 2007, Humphries et al., 2006, O’Reilly and Frank, 2006, Stewart et al., 2010, Stocco and Lebiere, 2014, Stocco et al., 2010). For example, O’Reilly and Frank (2006) have shown that, in a neural network model of the interactions between basal ganglia and PFC, the basal ganglia learn how to adaptively gate sensory signals to PFC for future elaboration. In other words, the basal ganglia learn to adaptively and strategically use PFC as a working memory buffer and to control which stimuli are worth memorizing (“gating”) and which, instead, can be discarded.

The most commonly held view is that the selection of sensory signals is mediated by the competition between at least two pathways within the circuit (Albin et al., 1989, DeLong, 1990). These two pathways have been traditionally indicated as the direct and the indirect pathway, and have excitatory and inhibitory effects, respectively, on thalamic neurons projecting to the PFC (Fig. 2). Specifically, the direct pathway results in an increase of thalamic inputs to the cortex, while the indirect pathway results in a net decrease of these inputs.

Crucially, the opposition between the two pathways is mediated by dopamine-driven reinforcement-learning mechanisms (Niv et al., 2005, Schultz, 1998, Schultz et al., 1993, Schultz et al., 1997), so that the basal ganglia effectively learns to optimize the selection of the most relevant signals to send to PFC based on their expected utility (O’Reilly and Frank, 2006, Stocco et al., 2010). Some models have also outlined a connection between the activity of the two pathways and specific cognitive abilities. Specifically, the opposite effects of the direct and indirect pathways on the thalamic projections has been likened to controlling a “gate” (Frank et al., 2001, O’Reilly and Frank, 2006) or a “router” (Stocco, 2012, Stocco and Lebiere, 2014, Stocco et al., 2010) to PFC, thus determining which stimuli will be maintained, and which will be updated, in working memory.

Because the indirect pathway has an inhibitory effect on thalamic projections, its role becomes more prominent as cognitive performance becomes more dependent on discarding irrelevant task features. For example, increased expression of dopamine D2 receptors, which are specific to the indirect pathway (Gerfen et al., 1990), is associated with better performance in an N-back task (Zhang et al., 2007). Because this task requires participants to maintain a moving window of the last N items in a stream of stimuli that are presented, it places severe demands on controlling which information is added to the contents of working memory. Similarly, McNab and Klingberg (2008) found that basal ganglia activity, most likely associated with the indirect pathway (because it is localized in the external globus pallidus: Fig. 2), correlates with better performance in a spatial working memory task in which participants are required to memorize the location of certain stimuli (e.g., red dots) while ignoring distractors (e.g., yellow dots). Conversely, reduced expression of D2 receptors has been associated with reduced visuospatial skills (Berman & Noble, 1995) and impaired performance on the Wisconsin Card Sorting Task (Han et al., 2008). Finally, at least one study reported that administration of cabergoline, a D2-agonist, resulted in improved performance in response inhibition and executive function tasks (Messer, 2011).

In summary, a vast amount of research suggests that the basal ganglia play a fundamental role in selecting which sensory information is ultimately encoded in PFC, and therefore, which dimension of a stimulus will be used to resolve conflict between competing responses. Furthermore, neurophysiological evidence and computational models jointly suggest that the key function of filtering out unwanted information is specific to one of two pathways within the basal ganglia circuit, the so-called indirect pathway. For this reason, we expect that individual differences in cognitive control should be related to individual differences in the effects of the indirect pathway on PFC.

Because the two pathways are interleaved throughout the basal ganglia, their competitive effects cannot be dissociated easily. However, their relative contribution can be indirectly measured through the Probabilistic Stimulus Selection (PSS) task (Frank et al., 2004). The PSS task is an iterative, implicit decision-making paradigm in which participants repeatedly choose from pairs of non-verbalizable stimuli, each of which has a different probability of yielding a reward. In the PSS task, participants are initially trained to select the most rewarding stimulus out of three different pairs (Fig. 3A and B). Note that the correct response can be made by either learning to choose the most rewarding stimulus, or by learning to avoid the least rewarding one. To distinguish between these two strategies, participants are then tested over the remaining combinations of stimuli (Fig. 3C and D), so that their sensitivity for learning to choose high-reward probability stimuli (“Choose” accuracy) and their sensitivity for learning to avoid low-reward probability stimuli (“Avoid” accuracy) can be measured independently. Multiple patient, pharmacological, and genetic studies (Frank et al., 2004, Frank and Hutchison, 2009, Frank et al., 2007) have shown that Choose accuracy reflects the contribution of the direct pathway, while Avoid accuracy reflects that of the indirect pathway.

For example, Frank et al. (2007) showed that better “Choose” accuracy was associated with polymorphisms in the DARP32 gene, known to affect the expression of D1 dopamine receptors in the direct pathway, while “Avoid” accuracy was associated with polymorphisms in the C957T gene, known to affect the expression of D2 dopamine receptors in the indirect pathway (Frank et al., 2007, Frank and Hutchison, 2009). Additional evidence comes from patients with Parkinson’s Disease, a pathology that results from the death of cells releasing dopamine in the striatum. Because dopamine excites the direct pathway but inhibits the indirect pathway, Parkinson’s Disease results in abnormal activation of the indirect pathway at the expense of the direct pathway. Which is paralleled by a corresponding increase in Avoid accuracy and decrease in Choose accuracy (Frank et al., 2004). Finally, when Parkinson’s patients were tested after taking dopamine-promoting medication (and thus increasing the activity of the direct pathway at the expense of the indirect one), their Choose accuracy increased while their Avoid accuracy plummeted.

In summary, the existing literature suggests that successful cognitive control depends on the proper allocation of attention to the relevant dimensions of a stimulus and that this process, in turn, depends at least partially on the activity of the basal ganglia, and of the indirect pathway in particular. If this assumption is correct, individual differences in cognitive control should be correlated with individual differences in the activity of the indirect pathway. To test this prediction, we conducted an experiment to correlate individual performance in the PSS task (whose accuracies are affected by the relative strengths of the direct and indirect pathway) and the size of the incongruency effect in the Simon task (which reflects a person’s capacity for cognitive control). To control for confounding factors, independent measures of working memory capacity were also collected and partialled out. Our main prediction was that, at the group level, Avoid accuracy in the PSS task should be negatively correlated with the incongruency effect in the Simon task. Furthermore, we expect that this correlation is selectively due to the capacity to filter out the irrelevant stimulus dimension (that is, spatial position) during the incongruent trials. Thus, we expect that a similar correlation exists between mean response times to incongruent trials in the Simon task and Avoid accuracy in the PSS, but not between Avoid accuracy and response times for congruent trials. Finally, because we expect that the act of processing both stimulus shape and position are highly automatic and unlikely to change over the course of a single experimental session, we do not expect to find any significant correlation between Choose accuracy and any response times in the Simon task.