Mangels Lab: Cognitive Neuroscience of Memory and Attention

Binding in Episodic Memory: The Role of Fronto-Posterior Interactions

 

Neuroimaging and neuropsychological studies consistently find that regions within the prefrontal cortex and medial temporal lobe/hippocampal complex are critical for encoding information such that it can be retrieved either in the presence of few retrieval cues (i.e., free recall), or when the retrieval of contextual details is required (i.e., source memory, temporal order, conscious recollection). The precise role of these structures and the manner in which they interact has been the subject of much debate, however. The prefrontal cortex (PFC), in particular, has been ascribed numerous roles in this process, including general strategic organization, semantic retrieval and elaboration, storage and association of spatiotemporal context, and inhibitory control of irrelevant information. Recently, in my lab and others, there has been an effort to synthesize the hypothesized role of the PFC in episodic encoding with its established involvement in working memory.

Our view of the PFC would predict that during the successful binding of information in episodic memory, prefrontal regions should be engaged in sustained communication with regions dedicated to processing various task-relevant conceptual and/or perceptual representations that need to be bound. An earlier ERP (event-related potentials) study of successful encoding under divided attention found preliminary evidence supporting this view (Mangels, Picton & Craik, 2001). In that study, concurrent, sustained (500-1500 ms) anterior frontal and inferior posterior activity predicted later recall and conscious recollection of verbal stimuli on a recognition test, but was absent for items that could only be retrieved later on the basis of familiarity or were later forgotten. Subsequent familiarity, in contrast, was associated with relatively discrete, early (200-400 ms), neural activity over left inferior temporal regions that we interpreted as evidence for the item-specific conceptual processing of the word. Nonetheless, the findings of Mangels, Craik & Picton (2001) provided only indirect support for the hypothesis that these concurrent fronto-posterior activations were indexing exchange of information in the service of successful encoding. More substantial evidence in support of this view would come from demonstration that actual functional connectivity between these regions has predictive power for encoding success. Thus, together with former graduate student, Christopher Summerfield, we sought to investigate frontal-posterior networking more directly using EEG coherence analysis (Summerfield & Mangels, 2005a, see also Summerfield & Mangels, 2005b for a study of fronto-posterior coherence during true and false retrieval). EEG coherence measures the extent to which oscillations between multiple, distant neural regions demonstrate a consistent phase relationship. In scalp-recorded EEG, this method provides a measure of long-range cortical synchrony, and thus, a putative index of neural exchange during functional coupling. Increasingly, the value of coherence as an analytic tool is being recognized for its ability to move cognitive neuroscience beyond discrete neural maps and into the dynamic, network view of brain function required by current models of cognitive processing. We have endeavored to operate at the forefront of this movement by combining this approach with multivariate statistical analysis techniques such as independent component analysis (ICA). The Matlab code used for this analysis was designed in-house by Chris Summerfield and can be found at http://www.columbia.edu/~cs2028/beast/beast.htm.

Figure 1.

We have used this approach to dissociate long-range cortical synchrony associated with successful and unsuccessful binding. In our first study, which examined binding of semantic-perceptual (word-color) associations, we found that successful encoding of word-color associations was associated with sustained bilateral fronto-posterior theta-band (4-8Hz) coherence (Figure 1; ICA component illustrating factor differentiating items later remembered with correct color from items remembered with incorrect color or forgotten). This pattern of activity was expressed particularly during the latter portion of the stimulus epoch (800 ms to the end of the stimulus presentation) (Summerfield & Mangels, 2005a). Theta is also observed from local field potentials in the hippocampus during learning, and long-term potentiation in this region may depend on the phase of theta. Thus, we speculate that theta may be the candidate code by which cortical and hippocampal regions exchange information during successful episodic encoding. Given the low spatial resolution of EEG, however, it is still unclear whether the frontal regions we observed in Summerfield & Mangels (2005) were associated with a cognitive control process (dorsolateral PFC) or with the selection of semantic information (more ventrolateral PFC). In other words, the frontal activity we observed may have indicated “site” specific activity associated with semantic processing (and its integration with posterior perceptual information), rather than sources of top-down control. Indeed, some recent evidence from our lab suggests that high-frequency gamma activity maybe specifically involved in cognitive control (see next section), a role that would be consistent with the view that gamma activity is involved in perceptual integration and predictive coding.

Figure 2.

EEG coherence has the advantage of measuring neural activity with great temporal precision, however it lacks the spatial resolution to determine what specific frontal regions are involved and whether they are specifically targeting task-relevant representations or simply modulating posterior activity more broadly. Therefore, we recently conducted an fMRI study of functional connectivity during encoding in collaboration with Joy Hirsch (Director of the Functional MRI Research Center at Columbia University) and Tor Wager (Summerfield, Green, Wager, Egner, Hirsch, & Mangels, 2006). In this task, subjects encoded associations between faces and houses, stimuli that were chosen specifically because they reliably elicit responses in separate regions of extrastriate visual cortex (i.e., fusiform face area [FFA] and parahippocampal place area [PPA], respectively). Univariate analysis confirmed that regions in these extrastriate regions were more active for successfully compared to unsuccessfully encoded face-house pairs, as were regions in the medial temporal lobe (MTL), including the left hippocampus and bilateral rhinal cortices. Many of these areas demonstrated functional connectivity with each other, such that the strength of their connectivity increased as subsequent memory performance increased (e.g., FFA-PPA, FFA-MTL). Importantly, however, when FFA and PPA were used as “seed” regions in an exploratory analysis of functional connectivity, we found a region within the left dorsolateral PFC (Figure 2) that correlated with the extent to which both of these areas predicted learning success. We suggest that the dorsolateral PFC is engaged in top-down modulation of these extrastriate regions in the service of gating the transfer of information to the medial temporal lobe. We are currently testing these ideas more formally to determine whether the PFC exerts a moderating or mediating effect on connectivity observed between FFA and PPA regions and between FFA and MTL.

Attention: top-down control and bottom-up salience

Paradigms that attempt to segregate neural activity associated with successful and unsuccessful encoding, like those discussed above, take advantage of the moment-to-moment variability in the success with which we learn new information. We subscribe to the view that a large component of this variance has to do with variations in attention. This view is supported by numerous findings demonstrating that diverting attention away from an encoding task reduces episodic memory performance (e.g., dual-task studies; Mangels, Picton, & Craik, 2001, see also Mangels, et al., 2002). However, dual-task studies only provide information about how encoding success varies under different levels of global processing resources. Contemporary theories of attention describe dissociable mechanisms for “top-down” goal-directed, endogenously-driven attention and “bottom-up” stimulus-driven, exogenously driven attention. Recent studies have applied these models toward understanding the separate and interacting roles of top-down and bottom-up attention play in encoding success, as well as in feature selection more broadly.

Prediction Error, Attention and Memory

Prediction error (discrepancies between expected and actual outcomes) strongly influence reinforcement-based associative learning in incremental learning paradigms. This part of the research program investigates how these errors influence updating of associations in long-term declarative memory.

In some studies, we have maximized the bottom-up salience of stimuli explicitly by using a small number of items that are distinctive from their background due to novelty along conceptual or emotional dimensions (i.e., “isolates”). We then modulated top-down processes by instructing subjects that the goal of the task was to make judgments either about the conceptual “fit” (conceptual orienting) or negative valence of the stimuli (emotional orienting). In terms of memory performance, emotional isolates enjoyed the expected memory advantage over control words, particularly in terms of conscious recollection (“remember” responses), whereas conceptual isolates were more likely to be retrieved on the basis of familiarity (“know” responses) than were control items. Although these patterns of memory performance were found regardless of encoding task, the encoding task modulated the ERP activity associated with stimulus detection. When top-down attention was oriented to emotional valence, a fronto-central P3 unique to emotional isolates peaked reliably earlier than when attention was oriented to conceptual fit. This speeding of the neural response paralleled a faster behavioral response to decisions about the emotional isolates. Interestingly, however, memory was worse overall in the emotional orienting condition. Thus, faster orienting did not necessarily translate to better memory.

We have also explored the influence of top-down and bottom-up attention on subsequent memory in a novel manner: by varying the temporal predictability of information in the environment (Summerfield & Mangels, 2006). In the typical memory experiment, items are presented sequentially, separated by a constant interval, so that subjects know when to expect the arrival of each to-be-remembered stimulus. Summerfield & Mangels (2006), however, interspersed to-be-learned items (words in different colors) with irrelevant stimuli (“crosshairs”), in such a manner that the arrival of the critical stimulus (word) could only be predicted with 100% certainty if two crosshairs had preceded it (i.e., capitalizing on an aging interval paradigm). We viewed highly predictable stimuli, which also had the greatest inter-trial interval, as ones in which top-down attention could be maximally engaged. In contrast, unpredictable stimuli were more likely to engage bottom-up attentional orienting (like “isolates”). In this paradigm, the more predictable stimuli were remembered better than the unpredictable stimuli. Patterns of EEG synchrony predicting subsequent memory differed as a function of predictability and offered some explanation for the memory advantage of the predictable items. For the highly predictable items, early (150 ms post-stimulus) gamma-band activity over left frontal regions was a robust predictor of later memory, whereas for the least predictable items later (>400 ms) theta-band activity over left and midline frontal cortex was increased for items later remembered. These findings are consistent with the emerging role of gamma-band activity in top-down attention and frontal midline theta in bottom-up attentional orienting. The finding that top-down control processes engage attention to the stimulus very close to the onset of the to-be-remembered information (~150 ms) may maximize the duration over which attention can be dedicated to processing the stimulus, whereas the later orienting to unexpected stimuli reduces time for more elaborative stimulus processing.

The study outlined above uses metacognitive expectancy about the stimulus sequence in order to separate top-down and bottom-up attentional processes. In another series of studies, we approached expectancy from a different perspective, the perspective of an individual’s predictions about the accuracy their own memory performance (i.e., metamemory). In collaboration with Brady Butterfield, a former graduate student in my laboratory (Butterfield & Mangels, 2003), we used ERPs in an effort to understand why negative feedback is more effective in facilitating correction of memory errors when the negative feedback is incongruent with expectation than when it is congruent (i.e., the hyper-correction of high-confidence errors, see Butterfield & Metcalfe, 2001). In these studies, subjects provided answers to general information questions (“What is the capital of Canada?”), and rated their confidence that the answer was correct. After each response, accuracy feedback was provided that either confirmed or disconfirmed expectation (as indexed by confidence), followed by the correct answer to the trivia question, which was information they could use to correct their error on two surprise retests, one immediately afterward and one a week later. We are now exploring whether the benefit to memory conferred by orienting to these “metamemory mismatches” can also be instrumental in correcting false memories, where stimulus familiarity is actually responsible for the error, and therefore, unlikely to contribute to error correction (i.e., Deese-Roediger-McDermott [DRM] false-memory paradigm).

Top-down control in visual selective attention

The idea that top-down control processes increase the efficiency of stimulus selection and processing is not specific to memory encoding. We have also investigated the influence of top-down biasing on feature selection in the face of interference (i.e., the biased competition model) using two classic visual attention paradigms, the Stroop task and visual search).

Figure 3.

The idea that top-down control processes increase the efficiency of stimulus selection and processing is not specific to memory encoding. We have also investigated the influence of top-down biasing on feature selection in the face of interference (i.e., the biased competition model) using two classic visual attention paradigms, the Stroop task and visual search). Together with former graduate student, Emily Stern, we have measured ERPs during a cued spatial Stroop task in order to examine whether top-down control can operate in advance of target presentation to bias processing of a particular stimulus dimension (Stern & Mangels, 2006). A key result is that when subjects are cued to select word information (and inhibit spatial information), the degree of sustained right anterior frontal activity manifest during this preparatory period (see electrode F6, Figure 3) is correlated with both enhancement of the neural activity related to word processing (inferior temporal negativity; ITN) and a speeding of reaction time to the word. These results provide the first direct link between preparatory, “top-down” control processes (initiated by the cue stimulus), “bottom-up” processing of the target stimulus features (word form processing), and behavior (reaction time) in a Stroop conflict task.

Using visual search as a model task, we also have investigated the neuropharamcological basis for the interaction between top-down and bottom-up attention in a recent collaboration I initiated with Todd Horowitz (Harvard Medical School), Jon Horvitz (Boston College), and Lucien Cote (Columbia University Medical Center). Dopamine is considered to play a key role in the top-down selection of task-relevant striatal inputs by amplifying the activity of those striatal cells receiving strong “bottom-up” glutamate excitation, while attenuating the impact of those receiving weaker inputs, thereby boosting signal-to-noise ratio in the flow-through to motor response systems. Correspondingly, patients with Parkinson’s disease (PD), a degenerative disease of the nigrostriatal dopamine system, often demonstrate deficits in selecting task-relevant stimuli in the presence of irrelevant stimuli. In a recent study of visual search in medicated moderate-to-severe PD patients (Horowitz, Choi, Horvitz, Cote, & Mangels, 2006), we found that PD did not impair the detection of highly salient stimuli (feature search). However, when “bottom-up” noise was added (increased distractor heterogeneity on a non-relevant dimension and conjunction search), PD performance was disrupted when top-down information was not explicitly provided by the experimenter. Interestingly, impairments were found in the intercepts (and not the slopes) of the RT x set size functions, indicating that deficit was at decision and/or response selection processes downstream of the target search itself, consistent with the role that dopamine is hypothesized to play in response selection.

Top-down control of attention: individual differences and applications

In the studies described above, top-down control was manipulated through task instruction, task design, or was rooted in the performance expectations of the individual. Recently, we have begun to examine another factor influencing top-down control, one that is often overlooked in cognitive neuroscience studies of successful learning: individual differences in achievement motivation. In cognitive neuroscience there is an emphasis on determining the average pattern of performance or neural activity, and individual differences are considered nuisance variables. Some of these individual differences, however, stem from differences in strategies, goals and beliefs that the participant might bring to the learning situation. In social cognitive neuroscience, the merger with social psychology provides tools with which to capture these individual differences more systematically. Given my particular interest in predictors of successful learning, Carol Dweck’s work on theories of intelligence and the achievement goals that arise from these beliefs presented itself as a natural complement to my work.

In her previous work, Dweck and colleagues found that students who viewed intelligence as a fixed quantity (“entity theorists”) were vulnerable to decreased performance when tasks became challenging or when they realized they were at risk of failing. In contrast, students who viewed intelligence as acquirable (“incremental theorists”) appeared better able to remain effective learners. These differences appear to be rooted in the different achievement goals that follow from holding either a fixed or an acquirable view of intelligence. Entity theorists are more likely to be concerned with proving—rather than improving—their abilities, either relative to a general benchmark (“performance goals”) and/or relative to the perceived ability of others (“normative goals”). As a result, they are especially vulnerable to negative consequences when faced with a difficult task, and are more likely to shun learning opportunities that might reveal ignorance. In contrast, incremental theorists are more likely to endorse the goal of increasing ability through effort (“learning goals”), and indicate a willingness to seek out challenging opportunities, even if failure is a possible outcome, as long as there is the potential for growth (“challenge goals”).

Figure 4.
Based on the hypothesis that beliefs about intelligence directly influence an individual’s goals during a learning task, we expected that these goals would influence how a cognitive control network deployed attention and strategic processing during learning (Mangels et al., 2006). In this study, we careful matched entity and incremental theorists for performance on an initial test of general knowledge, then looked at the extent to which these groups were able to learn from feedback and correct these errors on a later surprise retest. Entity theorists corrected significantly fewer errors, and our investigation of the ERP data time-locked to presentation of the feedback provided some interesting clues to why this occurred. First, entity theorists oriented more strongly than incremental theorists to information that conflicted with the goal of “proving” their ability, as indicated by an enhanced frontal P3a to negative, but not positive, performance-relevant feedback (see Figure 4; ERPs at Fz time-locked to negative feedback, with arrow pointing to the P3a; legend: entity theorists: black lines, incremental: grey lines; solid lines: expected errors, dotted lines: unexpected errors). A significant correlation between normative goals and the amplitude of the P3a to negative feedback in both groups suggested that the salience of this conflict was a function of concerns about how performance might appear relative to others. Yet, despite the enhanced salience of negative feedback for entity theorists, they appeared less likely than incremental theorists to subsequently engage in effortful conceptual processing of learning-relevant feedback following failure, as indicated by a reduction in the duration of left inferior temporal negativity (ITN) associated with subsequent memory performance. Given extensive evidence demonstrating that the success of associative encoding is positively related to the duration of conceptual processing, these differences may have been responsible for entity theorists’ significantly poorer performance on the surprise retest. Thus, these findings provide insight into why chronic differences in approach to learning and challenge could lead to significant differences in the success of knowledge acquisition over time.

Presently, in collaboration with Catherine Good (Barnard College), we are extending this work to the domain of stereotype threat, a phenomenon that has been shown to contribute to under-performance in target groups. The target group of current focus is women in challenging math situations. Previous research directed at investigating the potential mediators of stereotype threat effects (such as anxiety or compromised problem-solving strategies) have tended to rely on self-reports that follow performance and have yielded no clear answers. Our use of on-line, covert electrophysiological measures of attention and strategy deployment will hopefully provide some important insight into the neurocognitive mechanisms by which stereotype threat undermines learning and performance success of women in the context of mathematics. In addition, we plan to use these techniques to determine whether interventions can be successful in redirecting attention of threatened individuals toward more learning-relevant information, thereby promoting learning success even in an environment of stereotype threat.