Stimulus-dependent EEG activity reflects internal updating of tactile working memory in humans

Abstract

Despite recent advances in uncovering the neural signature of tactile working memory processing in animals and humans, the representation of internally modified somatosensory working memory content has not been studied so far. Here, recording EEG in human participants (n = 25) performing a modified delayed match-to-sample task allowed us to disambiguate internally driven memory processing from encoding-related delay activity. After presentation of two distinct vibrotactile frequencies to different index fingers, a visual cue indicated which of the two previous stimuli had to be maintained in working memory throughout a retention interval for subsequent frequency discrimination against a probe stimulus. During cued stimulus maintenance, α activity (8–13 Hz) over early somatosensory cortices was lateralized according to the cued tactile stimulus, even though the location of the stimuli was task irrelevant. The task-relevant memory content, in contrast, was found to be represented in right prefrontal cortex. The key finding was that the visually presented instructions triggered systematic modulations of prefrontal β-band activity (20–25 Hz), which selectively reflected the to-be-maintained frequency of the cued tactile vibration. The results expand previous evidence for parametric representations of vibrotactile frequency in the prefrontal cortex and corroborate a central role of dynamic β-band synchronization during active processing of an analog stimulus quantity in human working memory. In particular, our findings suggest that such processing supports not only sustained maintenance but also purposeful modification and updating of the task-relevant working memory contents.

Working memory (WM) refers to a set of operations necessary for maintenance and online processing of information in the service of behavior (1). Across primate species, and for most different types of to-be-maintained information, central WM function has been attributed to the lateral prefrontal cortex (PFC), but the precise nature of prefrontal contributions to stimulus maintenance is still discussed (for reviews, see refs. 2 and 3). Significant progress in delineating the neural signature of maintenance processing in the PFC has been achieved in the somatosensory domain using vibrotactile stimuli in delayed match-to-sample (DMTS) frequency-discrimination tasks. During frequency maintenance, single-cell firing (4) and neuronal population activity (5) in the inferior PFC of monkeys varies parametrically as a monotonic function of the to-be-maintained frequency. PFC engagement during vibrotactile DMTS maintenance also has been demonstrated in humans by task-related changes in hemodynamic responses (6) and synchronized oscillatory EEG/magnetoencephalographic (MEG) activity (7, 8). In particular, using EEG, it was found recently that the amplitude of prefrontal oscillatory activity in the upper β band (20–25 Hz) increased linearly with the frequency of the previously presented vibration (8). Such a correlate of parametric^† WM (4) in human EEG opens perspectives for noninvasive investigation of higher-level memory function, which may be difficult to assess in animal studies.

Although the crucial role of the PFC during tactile memory processing is well established, as is a less defined involvement of the secondary somatosensory and medial premotor cortices (for review, see ref. 9), it remains unclear whether the primary somatosensory cortex (SI) is engaged as well. The lack of clarity arises from casual evidence for a contribution of SI to vibrotactile memory in humans (e.g., refs. 6, 10, 11) opposed to generally negative findings of SI delay activity in monkeys (for review, see ref. 9, but also see ref. 12). To assess a potential involvement of human SI, in the present study we examined the stimulus-dependent lateralization of the EEG responses, thereby exploiting the fact that the hand representation in SI, unlike in downstream areas, is strictly contralateral (13–15). This in combination with available source reconstruction methodology, allowed us to deduce a possible engagement of the SI during WM processing.

Most theorists agree that human WM function is not restricted to the temporary perseveration or freezing of most recent sensory information, as is assessed in traditional DMTS tasks (for review, see refs. 2, 16). Rather, the contents of WM can be subject to internal manipulation and can be selected deliberately from different sensations or from long-term memory (2, 16). The question arises how a temporary representation of past sensory experience may be updated or refreshed (17) through goal-directed control over the contents of WM. Here, we sought to characterize such internal WM updating in terms of content-specific modulations of human EEG activity. We used a modified vibrotactile DMTS task (Fig. 1A), which involved active, retrospective selection of the to-be-maintained tactile information, for subsequent frequency discrimination. Our main focus was on possible parametric representations of the to-be-maintained vibrotactile frequency, and we expected these representations to be evident in prefrontal upper β oscillations (8). The key question was the extent to which such content-specific modulations of population activity in PFC may reflect the internal updating and modification of the current WM content.

Fig. 1.

Instructions to actively maintain a previously presented vibrotactile stimulus decrease oscillatory α/β activity in SI and increase α activity in the occipital cortex. (A) Cued frequency-discrimination task. In each trial, two vibrotactile stimuli of different frequencies (f1A and f1B, each varied between 12–40 Hz) were presented sequentially to subjects' index fingers. In half of the trials, f1A was presented to the left index finger, and f1B was presented to the right index finger (left/right trials, saturated colors); in the remaining trials, the sequence was laterally reversed (right/left trials, unsaturated colors). A visual cue (A, B, or O) indicated whether subsequent frequency discrimination was required against f1A (A cue), against f1B (B cue), or whether no further task had to be performed (O cue). After a retention interval of 2.5 s, the comparison frequency (f2, 8–44 Hz) was presented to either the left or the right index finger. The green bar indicates the time window of EEG data analysis. (B and C) Differential oscillatory EEG responses during cued frequency maintenance (A and B conditions) compared with control trials (O condition). (B) Group-level (n = 25) statistical parametric maps (SPMs) reflecting differences in oscillatory activity for representative posterior (Pz, Upper) and sensorimotor channels (C4, Lower). Dashed lines indicate cue onset. TF activity at the edges of the analysis window (transparent mask) was spectrally interpolated from zero-padded data. (C) (Top) Topographical statistical map of differential activity in the α band (8–13 Hz, 0.5–2 s). Color scale is as in B; dashed rectangles indicate positions of the channels shown in B. (Middle and Bottom) SPM source reconstruction of differential α activity (8–13 Hz, 0.5–2 s). Red indicates an increase in source power, and blue indicates a decrease (P < 0.001), relative to control trials.

Results

EEG was recorded from human volunteers (n = 25) performing a cued frequency-discrimination task (Fig. 1A). After the presentation of two distinct vibrotactile frequencies (f1A and f1B, varied between 12 and 40 Hz), a visual cue (“A” or “B”) indicated which of the two vibrations was to be maintained in WM throughout a retention interval for subsequent discrimination against a comparison vibration (f2; ± 4 Hz). In control trials (cue “O”), neither of the previous stimuli had to be actively maintained.

Participants correctly discriminated 68.6% of the f1A stimuli in the A condition and 72.8% of the f1B stimuli in the B condition from the comparison frequency (f2). Hence, performance was well above chance (50%) in both conditions (both P < 0.001) and was significantly better for the more recent stimulus (f1B) than for the stimulus presented earlier (f1A) (P < 0.005). Across frequencies, as expected by Weber's law (18), f1–f2 discrimination performance decreased with increasing stimulus frequency (both P < 0.001; linear trend analysis). Mean response times did not differ between the A and B conditions (520 vs. 525 ms; P > 0.05) and showed no covariation with the frequencies of f1A or f1B (both P > 0.05; linear trend analysis).

Oscillatory EEG Responses.

We examined oscillatory EEG activity during the retention interval using general linear model (GLM) analysis on a single-trial level (Fig. S1) and Random Field Theory (19, 20) to correct for family-wise errors (FWE) in the time–frequency (TF) domain. Initially, we compared all trials in which a stimulus was actively maintained (A and B, correct trials) against control trials (O cue). In line with previous EEG/MEG work on DMTS vibrotactile maintenance (7, 8), an increase in α activity (8–13 Hz) was evident (Fig. 1B Upper; 550–2,100 ms; P_cluster < 0.05; FWE) over posterior channels (Fig. 1C Upper), probably reflecting inhibition of task-irrelevant visual areas (7, 8). Furthermore, a significant decrease of activity in the α^‡ (8–13 Hz) and β (15–25 Hz) frequency ranges was observed over sensorimotor cortices (Fig. 1B Lower, 300–2,100 ms; P_cluster < 0.001; FWE). We reconstructed the sources of these effects using multiple sparse priors (21) under group constraints (22) (Fig. 1C and Materials and Methods). The increase in α power was localized in visual areas in occipital cortex (Brodmann areas 18 and 17), whereas the decrease in α power was localized to SI (Brodmann areas 2, 1, and 3b). In an additional analysis, contrasting the overall oscillatory responses in the A vs. B conditions revealed no significant effects (all P_cluster > 0.05).

Next, we determined the extent to which the cue-induced changes in activity over SI were lateralized according to the specific stimulus selected for maintenance. (Methodological details are described in Fig. 2 A and C and in SI Materials and Methods). In both memory cue conditions, the α suppression was bilateral but was most pronounced over channels contralateral to the stimulus that was cued for WM maintenance (Fig. 2A). For statistical analysis, we quantified the lateralization in each cue condition as the difference between the activity over left and right sensorimotor channels (Fig. 2B). Compared with the O condition, the suppression of α activity was significantly lateralized toward channels contralateral to f1A in the A condition (600–1,000 ms; all time bins P < 0.05) and toward channels contralateral to f1B in the B condition (300–1,050 ms and 1,450–2,100 ms; all time bins P < 0.05). A similar albeit weaker lateralization pattern was observed in the β band (15–25 Hz; Fig. 2 C and D). Additional analysis showed that such a pattern of systematic cue-induced lateralization also was evident on incorrect discrimination trials (α: 650–1,000 ms; β: 800–950 ms; all time bins P < 0.05). Thus, although the side of stimulus presentation was task irrelevant, lateralized α (and β) suppression in early somatosensory cortex systematically reflected the location of the cued stimulus, regardless of subsequent frequency-discrimination success.

Fig. 2.

α and β activity over early somatosensory cortex is lateralized according to the tactile stimulus that is actively maintained in WM. (A) Flipped topographical statistical maps of oscillatory α activity (8–13 Hz) during the retention interval (0.5–1 s) following instructions to maintain f1A (Upper) or f1B (Lower), respectively, compared with control trials (O condition). For right/left trials (unsaturated colors in Fig. 1A), the channel data were flipped along the midline (dashed black line; midline channels remained unchanged) so that for all trials the channels on the left (purple arrow) reflect activity recorded contralateral to f1B (and thus ipsilateral to f1A), and channels on the right (green arrow) reflect activity contralateral to f1A (and thus ipsilateral to f1B). (B) Cue-dependent lateralization of sensorimotor α activity. Lateralization Index for each cue condition reflects t statistics of the difference in activity between the left and right hemisphere channels outlined in A. Activity at the edges of the analysis window (transparent mask) was spectrally interpolated from zero-padded data. Colored ribbons indicate significance of deviations in the A condition (upper ribbon) and in the B condition (lower ribbon), respectively, from the lateralization in control trials (O condition). Lateralization on incorrect trials (dashed lines) was less pronounced overall, probably because of smaller trial numbers, but differed significantly between A and B trials (dashed ribbon); note that O trials always were correct, because no task was required. (C and D) As in A and B for oscillatory activity in the β band (15–25 Hz).

Parametric Modulations of Oscillatory EEG Responses.

The main goal of the present study was to examine parametric modulations of oscillatory activity as a function of the vibrotactile frequency maintained in memory. Therefore, we analyzed parametric contrasts reflecting the strength of a linear relation between TF activity and the frequency of the previously encoded stimuli across single trials (Materials and Methods and Fig. S1).

Fig. 3 A–D summarizes the parametric modulations by the frequency of the most recently attended stimulus (f1B) in the different cue conditions. In the A condition, the parametric contrast reflecting f1B revealed no significant TF clusters (Fig. 3A Top). However, a strong modulation by f1B was observed in the B condition after retrospective cueing to actively maintain f1B (B cue; Fig. 3A Middle; P_cluster < 0.005; FWE; Fig. S2C). The modulation was most pronounced in the upper β frequency range (20–25 Hz, hereinafter termed “β2”), between 900–1,300 ms after onset of the visual cue over right-frontal recording sites (Fig. 3C Left), regardless of the side on which f1B had been presented (Fig. S2D). Consistently, parametric analysis of β2 source power estimates attributed the modulation to the right prefrontal cortex (Fig. 3C Right).

Fig. 3.

β2 activity (20–25 Hz) in prefrontal cortex is parametrically modulated by the frequency of the vibrotactile stimulus that is actively maintained in WM. (A) SPMs of parametric modulations by the frequency of f1B under cued instructions to focus retrospectively on f1A (A cue), f1B (B cue), or neither tactile stimulus (O cue) for a representative channel outlined in C. Dashed gray line indicates visual cue onset. (B) Descriptive plot of the average cue-specific changes in β2 activity as a function of f1B vibration frequency for the TF window outlined by the dashed rectangle in A. (C) Topographical statistical map (Left) and SPM source reconstruction of the β2 modulation by f1B in the B condition. (D) Time courses of the parametric modulation of prefrontal β2 power by f1B for each of the three cue conditions. Colored ribbons indicate significant differences in modulation on correct trials between the B and O conditions (upper ribbon) and between the B and A conditions (lower ribbon). (E–H) As in A–D, for parametric modulations by f1A.

Early after visual cue onset, and spectrally less focal, evidence for a modulation by f1B also was found in the O condition (Fig. 3A Bottom; P_cluster < 0.005; FWE; Fig. S2E), indicating a temporary continuation of f1B processing after the O cue (which was not associated with any distracting task; see Fig. S3 for supporting analysis). Direct statistical comparison (Fig. 3D) revealed that, compared with the B condition, the delayed modulation of prefrontal β2 power by f1B was significantly reduced in the O condition (1,100–1,400 ms; all time bins P < 0.05), in which no sustained maintenance of f1B was required. Most central to the present analysis, in the A condition (Fig. 3A Top), the instructions to actively maintain a different frequency (f1A) in memory virtually eliminated any parametric representation of f1B (Fig. 3D; significant differences in A vs. B condition: 850–1,700 ms; all time bins P < 0.05). Together, these results indicate that a parametric representation of the most recently encoded stimulus in PFC, in terms of parametric upper β modulations, was strongly influenced by active control over the to-be-maintained memory content.

We proceeded with GLM analysis of potential modulations by f1A (Fig. 3 E–H). To account for the earlier presentation of f1A (always before f1B; see Fig. 1A) and for single-trial variability introduced by the intervening processing of f1B (Fig. S3), the TF data were normalized with respect to the precue interval (SI Materials and Methods). The analysis of cue-induced modulations (Fig. 3E Top) revealed that shortly after instructions to actively maintain f1A (A condition), prefrontal β2 activity was modulated parametrically by the frequency of f1A (P_peak < 0.005; FWE; see Fig. S2F for detailed cluster analysis). No evidence for cue-induced modulations by f1A was observed in the B or in the O conditions (Fig. 3E Middle and Bottom). Direct statistical comparisons (Fig. 3H) revealed a significantly enhanced β2 modulation in the A condition between 200–750 ms and between 0–750 ms, compared with B and O, respectively (Fig. 3H; all time bins P < 0.05). Note that the actual time windows of these differences might be artificially lengthened by temporal smearing and Gaussian smoothing of the TF data (Materials and Methods). Compared with the modulations by f1B reported above, the scalp topography of the modulation by f1A appeared to be slightly more right-lateralized (compare topographies in Fig. 3 C and G and in Figs. S2 and S3; see Fig. S2F for condition-specific topographies). However, parametric analysis of normalized β2 source power estimates (SI Materials and Methods) attributed the modulation by f1A to a focal right-prefrontal source (Fig. 3G) within the same region as the source of the f1B modulations (compare source renderings in Fig. 3 C and G). Interestingly, repeating the above analyses on incorrect discrimination trials revealed no evidence for any cue-induced parametric modulations (dashed lines in Fig. 3 D and H).

Discussion

To summarize, during retrospective focusing on the memory of a previous tactile event, α (and lower β) activity was suppressed over SI contralateral to the cued tactile stimulus. The task-relevant stimulus attribute, in contrast, was parametrically represented by prefrontal upper β oscillations, which varied systematically as a function of the vibration frequency held in WM.

Oscillations in the β frequency range traditionally have been associated with low-level sensorimotor processing but are implicated increasingly in a broader cognitive context as well (23–25), including WM processing (8, 26, 27). The present results corroborate recent evidence for a role of dynamic synchronization of upper β oscillations in human PFC during vibrotactile frequency retention in a standard DMTS task (8). Here, using a modified DMTS task and single-trial GLM analysis, we found such parametric representation to be linked closely to active control over the memory content: Cue instructions to focus retrospectively on a particular vibration selectively enhanced the parametric modulation by this vibration's frequency and eliminated any parametric representation of the noncued vibration. Thus, modulations of prefrontal upper β oscillations reflected not only the active maintenance of most recent sensory input but also the goal-directed internal updating of WM with earlier encoded information. Importantly, these cue-induced effects were linked to successful subsequent frequency discrimination against a probe stimulus, suggesting a critical behavioral relevance of the underlying processes.

From a neuronal perspective, the parametric modulations in prefrontal β may represent a large-scale population aspect of the complex parametric representations found previously in PFC single cells (concomitant increases and decreases in firing rate; ref. 4) and microscale populations (5) during vibrotactile frequency maintenance. The parametric (i.e., continuous) nature of such representation may correspond to an abstract internal scaling of an analog stimulus quantity, which probably is required to perform the task. One potential alternative interpretation of the present prefrontal effects would be that cognitive demands increase with increasing stimulus frequency, as might be suggested by the behavioral performance in f1–f2 discrimination. However, our finding of a positive relationship—on correct trials only—between prefrontal β and vibration frequency renders this interpretation unlikely, because we found no evidence for any increase in prefrontal β during the cognitively demanding A and B tasks compared with the O condition (Fig. 2C); these results can be taken as evidence against a direct link between prefrontal β and task demands.

Interestingly, the present cue-induced modulations of prefrontal β were not static but showed early (A condition) and late (B condition) maxima after cue presentation (Fig. 3 D and H). The specific timing of the effects suggests that prefrontal β responds in particular during the active processing of a singular sensory quantity in the current focus of attention. In the present experimental context, such refreshment (17) of the task-relevant WM content may have been required early on in the case of f1A (Fig. 3H) to counteract further decay of an already remote memory and may have been recruited only relatively late in the case of the more recent stimulus, f1B. (Fig. 3D; for a similar time course in a standard DMTS task, see ref. 8). In the domain of abstract quantity information, such active internal updating or refreshment of the WM contents may be functionally similar to the rehearsal mechanisms previously proposed in the domains of phonological (28) and visuospatial (29) information, which traditionally are assigned to the left and right PFC, respectively (30). Adding to this rough taxonomy, our results yield evidence that active WM processing of vibrotactile frequency—like auditory pitch (31)—in particular engages right PFC.

In addition to the frequency-specific responses in PFC, cued retention was associated with a suppression of focal α activity over contralateral SI, suggesting a top-down controlled engagement (7, 32, 33) of early somatosensory cortex during WM processing. Although attentional modulations of oscillatory activity in SI have been reported both during tactile sensory processing (34, 35) and in anticipation of it (13, 33), here we demonstrate systematic SI activation related to retrospective focusing on the memory of a recent tactile experience. In light of these findings, it is possible that a potential engagement of SI during stimulus maintenance might have been overlooked in previous EEG/MEG studies (7, 8) in which memory-related activity was assessed relative to a prestimulus baseline. To shed light on this possibility, we reanalyzed the EEG data reported in ref. 8, where a standard unilateral vibrotactile DMTS task was used (n = 14). Examining the lateralization of raw spectral activity (Fig. S4), we found evidence for a suppression of contralateral sensory α during stimulus maintenance and even in the prestimulus interval (Fig. S4A) in anticipation of left-hand stimulation. We infer that active WM processing of tactile information may indeed involve attentional modulations of activity in SI that were not detected previously because of different methodological procedures. Central to the present study, however, neither of our analyses yielded evidence for a behaviorally relevant parametric WM representation of the to-be-maintained stimulus attribute in SI, a finding that is in line with previous work in animals and humans (4, 8, 9).

To conclude, the present results reveal a central role of prefrontal β oscillations during WM maintenance of an analog stimulus quantity on a large-scale neuronal population level. In addition to expanding previous work on somatosensory memory processing in animals (4, 5) and humans (8), our findings demonstrate that such parametric representation of vibrotactile frequency in human PFC depends critically on active control and reflects the goal-directed internal updating of the contents of WM.

Materials and Methods

Subjects.

Twenty-seven volunteers participated in the experiment with written informed consent. Two participants were excluded from analysis because of excessive ocular and movement artifacts. The study was approved by the Ethical Committee of the Charité-Universitätsmedizin Berlin and in accordance with the Human Subjects Guidelines of the Declaration of Helsinki.

Stimuli and Behavioral Task.

Vibrotactile stimulation was delivered by two identical 16-dot piezo-electric Braille-like displays (4 × 4 quadratic matrix, 2.5-mm spacing; QuaeroSys). The displays' pins were driven by fixed-amplitude sine functions of 750 ms duration at varying frequencies (8–44 Hz). The sound of the stimulator was masked by white noise (∼90 dB) presented through loudspeakers. Each trial began with the presentation of a fixation cross centered on a 19-inch thin-film transistor screen. After a variable prestimulus interval (1,000–1,500 ms), two vibrotactile stimuli were presented sequentially, separated by a 750-ms interstimulus interval (Fig. 1A). The first vibration (f1A) was delivered to either the left or right index finger (randomly assigned), and the second vibration (f1B) always was delivered to the other index finger. The two vibration frequencies were varied pseudorandomly between 12–40 Hz in steps of 4 Hz so that f1B always was 12 Hz lower or higher than f1A. Six hundred ms after offset of f1B, the fixation cross was replaced by a visual cue (A, B, or O) indicating whether the subject should focus on f1A or f1B for the remainder of the trial or on neither of the two stimuli. The assignment of letters to the three conditions was counterbalanced across subjects. After a retention interval of 2,500 ms, a comparison vibration (f2) was delivered to either the left or right index finger (randomly assigned). In the A and B conditions, f2 was chosen to be either 4 Hz lower or higher than the cued frequency. In the O condition, f2 was randomly chosen to be either 4 Hz lower or higher than either f1A or f1B. After f2 offset, the visual cue disappeared, and responses were given with the right foot by pressing a response pedal either once or twice to indicate whether f2 was “slower” or “faster” than the frequency maintained in WM. In the O condition, subjects were free to press the pedal either once or twice but were required to give a response. After 2,000-ms response time, performance feedback was given. After 50 practice trials, each participant performed eight blocks of 60 experimental trials.

EEG Recording and Analysis.

EEG was recorded using a 64-channel active electrode system (ActiveTwo; BioSemi), with electrodes placed in an elastic cap according to the 10–20 system. Individual electrode locations were registered using an electrode-positioning system (Zebris Medical GmbH). Vertical and horizontal eye movements were recorded from four additional channels. Signals were digitized at a sampling rate of 2,048 Hz, off-line bandpass filtered (0.5–100 Hz), downsampled to 512 Hz, and average referenced. The EEG was corrected for eye movements using calibration data to generate individual artifact coefficients and adaptive spatial filtering as implemented in BESA (v5.1.8, MEGIS Software; for details, see ref. 36). Remaining artifacts were excluded from analysis by careful visual inspection; the average maximum amplitude thereby retained was 79.6 μV.

Spectral Analysis.

All further analyses were carried out using SPM8 for MEG/EEG (Wellcome Department of Cognitive Neurology, London: www.fil.ion.ucl.ac.uk/spm/) and custom MATLAB code (Mathworks). The EEG data from all correct trials were epoched (−600 to 2,500 ms relative to cue onset, to prevent any contamination by stimulator artifacts) and zero-padded. TF representations of spectral power between 5 and 40 Hz were obtained by applying a tapered sliding window Fast Fourier Transform (FFT) using a single Hanning taper and an adaptive time window of seven cycles. Exploratory analysis of higher-frequency bands (>40 Hz) using a multitapered FFT yielded no significant effects.

Statistical Analysis.

The statistical design was implemented in SPM8, using the GLM applied to each subject's single-trial TF data. To warrant conformity with the GLM under normal error assumptions, the spectral power data were converted into amplitude values using a square root transform (37) and were convolved with a 3 Hz × 300 ms (FWHM) Gaussian kernel. The GLM design matrix (Fig. S1A) consisted of three dummy variables specifying the trials' cue condition (A, B, or O), three parametric regressors coding the frequency of f1A under the respective cue conditions, and three parametric regressors analogously specifying the frequency of f1B. The vectors coding f1A and f1B were zero-centered and normalized, rendering them nearly orthogonal (average vector angle 92° after rejection of artifact trials) and thus allowing us to estimate their respective impact within the same regression model. The model was inverted using restricted maximum-likelihood estimation as implemented in SPM8, yielding β parameter estimates for each model regressor at each TF bin. TF contrasts of interest then were computed by weighted summation of individual regressors' β estimates (Fig. S1B). The individual contrast spectra were subjected to mass-univariate analysis on the group level, using one-sample t tests as implemented in SPM8. FWE in TF space were controlled using Random Field Theory (38) to determine at each channel the FWE-corrected probability that a cluster of significant TF bins might have been obtained by chance. A cluster was thereby defined as a group of adjacent TF bins that all exceeded a threshold of P < 0.01 (38), corresponding to a t value of 2.49 at 24 degrees of freedom. To account for multiple comparisons across channels, only clusters exceeding a conservative FWE-corrected threshold of P_cluster < 0.005 were considered significant (Fig. S2). If a significant TF cluster was identified in a contrast of interest, we continued with conventional statistical analysis of the average contrast activity in the respective frequency band.

Source Reconstruction.

The sources of EEG activity were modeled using source reconstruction as implemented in SPM8 (39). For each participant a forward model was constructed, using an 8,196 vertex template cortical mesh coregistered to the individual electrode positions via three fiducial markers. The forward model's lead field was computed using the three-shell boundary element method EEG head model available in SPM8. Before model inversion, the raw data were bandpass filtered around the frequency band of interest. Source estimates then were computed on the canonical mesh using multiple sparse priors (21) under group constraints (22), including the data from all conditions of interest. Trial-specific TF contrasts were used to summarize oscillatory source power for specific frequency bands and at specific times as 3D images, which then were analyzed using conventional statistical parametric mapping procedures. To account for differences in trial numbers per analysis, overall and cue-specific source analysis results were thresholded at P < 0.001 and P < 0.005, respectively.