Attentional blink (AB) describes a visuo-perceptual phenomenon in which the second of 2 targets within a rapid serial visual presentation stream is not detected. There are several cognitive models attempting to explain the fundamentals of this information processing bottleneck. Here, we used electroencephalographic recordings and the analysis of interregional phase synchronization of rhythmical brain activity to investigate the neural bases of the AB. By investigating the time course of interregional phase synchronization separately for trials in which participants failed to report the second target correctly (AB trials) and trials in which no AB occurred, and by clustering interregional connections based on their functional similarity, it was possible to define several distinct cortical networks. Analyzing these networks comprising phase synchronization—over a large spectrum of brain frequencies from theta to gamma activity—it was possible to identify neural correlates for cognitive subfunctions involved in the AB, such as the encoding of targets into working memory, tuning of attentional filters, and the recruitment of general cognitive resources. This parallel activation of functionally distinct neural processes substantiates the eligibility of several cognitive models on the AB.
Attentional blink (AB) occurs when an otherwise perfectly discriminable stimulus is unreportable if presented as the second of 2 targets within a rapid serial visual presentation (RSVP) sequence (Broadbent and Broadbent 1987; Raymond et al. 1992). In a typical task, approximately 20 visual stimuli are presented with a stimulus-onset asynchrony of around 100 ms. When both targets are spaced >500 ms apart identification of both is easy, however, with a time delay of between 200 and 500 ms, Target 1 (T1) is reportable, whereas Target 2 (T2) often is not. Two prominent models propose explanations for the AB: The “Two-Stage Model” (Chun and Potter 1995) suggests that all items in the RSVP stream are processed to the point of conceptual representation without attention (Stage 1). Attention is then utilized to consolidate these representations into a durable and reportable form. An AB is believed to result from a failure of T2 to achieve Stage 2 processing due to on-going processing of T1 (e.g., Vogel et al. 1998). The “Interference Model” (Shapiro et al. 1997) proposes that interference from each identified RSVP stimulus occurs within working memory (WM) resulting in an inability to report T2.
Recent evidence supports the Two-Stage Model's inference that AB underpins an encoding deficit with research, demonstrating that AB occurs in the absence of distractors (Nieuwenstein et al. 2009) and is proposed as a consequence of failures in serial WM encoding (Craston et al. 2008). The absence of an AB on some trials implies that T2 reaches a level of awareness at which conscious report is possible. Current research suggests that conscious awareness may relate to states of the phase that are essential to the formation of transient neuronal assemblies facilitating communication both within local neural circuits and across wide regions of the brain (Singer 1998; Rodriguez et al. 1999; Fries 2005). For example, Gross et al. (2004) dissociated AB and successful target report via increased phase synchronization between prefrontal and parieto-temporal brain regions in the beta frequency band (15 Hz). The ability of the brain to synchronize the phase of neuronal activities over long distances is related to the occurrence of AB as exemplified by research demonstrating the role of neural synchronization in long-range interarea communication within the brain (Varela et al. 2001). Therefore, we hypothesized that AB and trials on which both targets were reported successfully (no-AB trials) would show consistent differences in the activation of oscillatory activity as previously proved (Gross et al. 2004). We investigated this hypothesis through the use of a subtraction method [no-AB phase-locking value (PLV) minus AB PLV] enabling an extension of Gross et al. via a low-resolution electromagnetic tomography source analysis (LORETA; Pascual-Marqui et al. 1994) of theta (4–8 Hz), alpha (8–12 Hz), beta1 (12–20 Hz), beta2 (20–30 Hz), gamma1 (30–50 Hz), and gamma2 (50–70 Hz) frequencies. The distribution of this PLV difference measure revealed the neural networks which survive AB.
The task involved presentation of targets at 2 levels of difficulty: condition 1 presented T1 and T2 in red and condition 2 presented T1 in red and T2 in gray. In both conditions, distractors appeared in gray. We aimed to increase the episodic distinctiveness of T2 in condition 1, hypothesizing that T2 identification would be increased as opposed to condition 2. Wyble et al. (2009) suggest that the role of cognitive load may be important in AB, with less-load predicted to associate with a better target performance. The contrasting conditions used in the present research enabled examination of this proposal.
Materials and Methods
After giving informed consent, 48 undergraduate psychology students took part in the experiment. A combination of excessive ocular and muscle artifacts resulting in trial rejection and an insufficient number of trials in one or more experimental conditions reduced this number to a final sample of 32 participants [11 males; all participants were right-handed, with a normal to corrected-to-normal vision; mean age 21.5 years (SD = 4.2)]. Only participants who performed with ≥80% correct identification of T1 were included to ensure participants remained attentive throughout the RSVP sequence. Participants received a course credit for their participation. The study was approved by an ad hoc ethics committee convened by the School of Psychology, National University of Ireland Galway (NUIG).
Apparatus and Experimental Paradigm
Target and distracter stimuli were presented at central fixation on a CRT monitor. All stimuli were presented on a black background. The horizontal refresh rate of the monitor was set at 75 Hz to ensure stimuli could be presented for a fixed presentation time of 93.3 ms (e.g., for 7 frames). Timing of visual stimulation was controlled with the Presentation 0.71 software (Neurobehavioral Systems).
Participants were seated upright in a chair and instructed to fix their gaze on the center of the computer screen where the experimental stimuli were displayed. A RSVP task was run in which 26 visual stimuli were sequentially presented for 93.3 ms each (Fig. 1). The majority of stimuli were distracters, unique letters (taken from the English alphabet) with each letter appearing randomly in the sequence and in a gray color. Within the RSVP, there were also up to 2 targets (T1 and T2). T1, when present, was either a # or a § symbol. T2 was presented in each trial as a number ranging from 1 to 9. A lag of 292 ms between T1 and T2 was employed as this lag was shown to evidence a strong decrease in second target identification (i.e., an AB) in the study by Gross et al. (2004). T1 onset occurred as either the fifth, sixth, seventh, eighth, or ninth stimulus in the RSVP stream. Participants were informed that they must search for 2 targets within each RSVP, but that T1 may or may not appear within a given trial. Participants were asked to make 2 responses following an on-screen prompt at the end of each trial: specifically, they were to respond to the presence/absence of T1 and its identity followed by a response to T2 identity. The task consisted of 4 separate conditions. In the first condition (T1salient–T2salient condition, SS), T1 and T2 were presented in red (43.46 cm/d). In the second (T1salient–T2neutral condition, Sn), T1 appeared in red whereas T2 appeared in gray (44.52 cm/d). In the third condition (noT1–T2salient condition, -S), T1 was absent from the visual array, whereas T2 appeared in red. The fourth condition (noT2neutral condition, -n), contained no T1 and T2 was presented in gray. The experiment consisted of a practice block of 15 trials followed by 4 experimental blocks of 120 trials each (30 trials per condition per block). Trial-presentation order was fully randomized across blocks and separately for each participant.
Electroencephalographic (EEG) signals were recorded for 90 s eyes open and 90 s eyes closed as well as during the experimental task, and digitalized via an EEG amplifier (QuickAmp-40, Brain Products GmbH, Munich, Germany). Electrophysiological activity was referenced to the common average of all channels, and data were sampled at 1000 Hz and analogue-filtered via a 0.15 high-pass filter and a 100-Hz low-pass filter. Additionally, a notch filter at 50 Hz was applied. EEG was recorded from 30 Ag/AgCl scalp electrodes arranged according to the extended 10–20 system and mounted on an elastic cap (EASY CAP EC40, EASYCAP GmbH, Herrsching, Germany) at the following sites: Fp1, Fp2, F7, F3, Fz, F4, F8, Fc1, Fcz, Fc2, Fc5, Fc6, Tp9, C3, Cz, C4, Tp10, Cp5, Cp1, Cp2, Cp6, T7, P3, Pz, P4, T8, Po9, Po10, O1, and O2. Impedances were kept below 15 kΩ. Ocular activity was measured via EOG channels mounted at the outer canthi of the right and left eyes, and approximately 2 cm above and below the right eye, respectively.
EEG data for each subject were filtered offline with high- and low-pass filters set to 0.5 and 80 Hz, respectively. Individual subject data were visually inspected for eye and muscle artifacts prior to artifact removal via the automatic artifact rejection (Criteria: amplitude minimum −50 µV, maximum +50 µV) and the Gratton et al. (1983) ocular artifact correction. Data were segmented into epochs with correct (no-AB) and incorrect (AB) T2 identification. Segment length was −300 to 700 ms in respect to T1 onset (see Supplementary Fig. 2 in Supplementary Material as a summary of analysis step). EEG was band-pass filtered according to 6 relevant frequencies of interest: 4–8 Hz (theta), 8–12 Hz (alpha), 12–20 Hz (beta1), 20–30 Hz (beta2), 30–50 Hz (gamma1), and 50–70 Hz (gamma2). Filtered single-trial EEG activity was transformed into source space using LORETA as implemented in BrainVision Analyzer 2 (Brain Products). LORETA estimates current source density for a realistic head model consisting of 6430 cortical voxels under the assumption that neighboring voxels have similar activity. The main purpose for transferring surface EEG into current source estimates in 3D space was to attenuate spurious phase synchronization based on volume conduction. Current source densities for 26 regions of interest (ROIs; see Supplementary Table 1 and Fig. 1) from filtered single trials were then exported to MATLAB 7.0 where interregional phase coherence was estimated via calculation of PLV [implemented with in-house scripts based on a method suggested by Lachaux et al. (1999)]. These ROIs were chosen in accord with recommendations made within the AB literature (Hommel et al. 2006). Phase synchronization was calculated for values in the range of 0–1, where 0 represents no reliable alignment of phase and 1 represents perfect phase synchronization between ROIs for a given frequency.
Phase coupling was calculated for all 325 ROI pairs resulting from 26 LORETA ROIs for each frequency band and for each time point between −300 and +700 ms relative to T1 onset. The first and last 100 ms from the analyzed time interval was trimmed to prevent spurious results due to filter artifacts at the edge of the analyzed segments. The PLV from remaining 800 ms (−200 to 600 ms in respect to T1) was averaged into 16 time windows of 50 ms duration each. This enabled a mapping of the temporal evolution of connections between ROI which evidenced a significant difference between AB and no-AB trials.
For each ROI pair (across all frequencies), the difference between no-AB and AB (16 time windows) was subjected to a non-parametric Friedman test (Friedman 1937) to identify ROI pairs exhibiting a PLV time course with a significant difference between no-AB and AB trials. Additionally, PLVs collapsed over the 16 time windows were compared between the 2 trial types using non-parametric Wilcoxon tests for each ROI pair at each frequency to identify pairs showing a general (time-independent) difference between AB and no-AB. A Ward's hierarchical cluster analysis, employing the Squared Euclidean Distance to determine the distance between clusters and furthest neighbor method for clustering, was performed on the PLV difference (ΔPLV) between no-AB minus AB for all ROI pairings (at any given frequency) exhibiting either a significant Friedman test (P < 0.01) or Wilcoxon test. ROI pairs that did not show any difference in the PLV between no-AB and AB were not included in hierarchical cluster analysis. The reasoning behind this analysis approach was the following one: Phase synchronization differences between AB and no-AB trials might be very similar for multiple ROI pairs. There might even be very similar time course of AB-related phase synchronization at particular ROI pairs at multiple frequencies. Instead of considering these very similar effects at multiple ROI pairs and multiple frequencies on their own, we applied the above described cluster analysis. This should cluster those ROI pairs at those frequencies together that show the most similar AB-related effect of phase synchronization, and therefore, are supposed to be functionally related.
Inspection of cluster coefficients from the agglomeration schedule suggested a solution with 10 clusters. This was further reduced to 5 clusters after a criterion of a minimum of 5 connections between ROI pairings, in cluster formation, was employed to simplify analyses.
In the cluster analysis, the time course of ΔPLVs was important given that ROI pairings exhibiting a similar temporal evolution of PLV differences between no-AB and AB should cluster together. Subsequently, and for each cluster separately, all ROI pairings involved (across all frequencies) were averaged to gain the total network synchronization for each cluster. A second analysis step was performed on these averaged data to assess the consistency of identified AB-related synchronization patterns at the sample level. The statistics on network-cluster level were equivalent to the precluster analysis statistics, namely a Friedman test on the difference between no-AB and AB trials over 16 time intervals. Wilcoxon tests between trial types at each time interval separately served as post hoc tests to dissociate the time course where conditions significantly differed [these were corrected for multiple comparisons using the false discovery rate procedure as suggested by Benjamini and Hochberg (1995)].
The control conditions in the present research, that is the SS, -S, and -n trials, were also subjected to the same statistical analysis as described above, to better demarcate the role of single versus dual target processing, and the exact influence of the salient versus nonsalient targets in the RSVP. Since there was no AB observable for any of these 3 conditions, PLV between no-AB trials as described above and the control conditions was compared. The difference in the PLV between no-AB and each of the 3 control conditions was averaged for the 5 clusters obtained from the cluster analysis in the main analysis described above and submitted to Friedman tests.
Table 1 summarizes the mean detection responses to T2, which was close to ceiling performance in all conditions except the Sn condition, where T2 performance was reduced as a result of correct T1 identification (Table 1). A typical AB pattern was observed in which there was a decrease in T2 detection given the correct identification of T1 within a close window in time; 292 ms.
|SS (salient T1, salient T2 present)||82|
|Sn (salient T1, neutral T2 present)||63*|
|-S (T1 absent, salient T2)||90|
|-n (T1 absent, neutral T2)||86|
|SS (salient T1, salient T2 present)||82|
|Sn (salient T1, neutral T2 present)||63*|
|-S (T1 absent, salient T2)||90|
|-n (T1 absent, neutral T2)||86|
A two-way repeated-measures ANOVA with factors T1 PRESENCE (present and absent) and T2 SALIENCE (salient and neutral) was conducted to evaluate the effect of T1 presence and target salience. The dependent variable was the percentage of correct responses. There were significant main effects for T1 Presence (F1,31 = 174.56, P < 0.0001, partial η2 = 0.849) and Target Salience (F1,31 = 75.274, P < 0.0001, partial η2 = 0.708). There was also a significant T1 Presence × T2 Salience interaction effect (F1,31 = 3.630, P < 0.0001, partial η2 = 0.066). Reference to Table 1 indicates that, as hypothesized, when T1 was present and T2 appeared in a neutral color, correct dual target report was reduced, falling to 63%. Post hoc t-tests confirmed a significant difference between SS and Sn conditions: t(31) = −8.53, P = 0.000, whereas SS did not significantly deviate from -n and -S.
Phase Synchronization Results
Statistical analysis of the no-AB minus AB PLV (ΔPLV) was carried out using Friedman tests, which revealed a significant difference between trial types for all 5 identified networks (clusters). Post hoc Wilcoxon tests demarcated the time course at which the networks showed a significant difference in connectivity between no-AB and AB trials. All Wilcoxon tests were corrected for multiple comparisons according to the method suggested by Benjamini and Hochberg (1995).
Cluster 1 comprised exclusively of theta activity synchronizing left prefrontal and right posterior ROI (Fig. 2). The PLV showed a significantly different time course for the AB and no-AB conditions [Friedman test: X2(15) = 58.86, P = 0.00]. Wilcoxon tests revealed significant differences at the following intervals post T1 presentation, in the dual target report conditions: 50 ms [Z = −3.05, P = 0.00], 100 ms [Z = −2.87, P = 0.00], 150 ms [Z = −2.58, P = 0.01], and 200 ms [Z = −2.24, P = 0.03], always indicating a stronger PLV for AB trials compared with no-AB.
Cluster 2 consisted mainly of left frontal to bilateral parietal connectivity between ROIs at alpha frequency. Frontal to left temporal ROI coupling was also apparent at alpha frequency. Theta activity synchronized temporal ROIs bilaterally. Gamma1 and gamma2 exhibited synchronization between left temporal and parietal, as well as left sensory motor and right temporal ROI, respectively (Fig. 3). For cluster 2, the Friedman test carried out on PLV difference between no-AB and AB trials was significant [X2(15) = 90.55, P = 0.00]. Wilcoxon tests revealed significantly a higher PLV for AB compared with no-AB at −150 ms (Z = −2.26, P = 0.02), −100 ms (Z = −3.8, P = 0.00), and −50 ms (Z = −3.42, P = 0.00) prior to T1 onset as well as at 100 ms (Z = −3.68, P = 0.00), and 150 ms (Z = −3.10, P = 0.00) after T1 presentation. A stronger PLV in no-AB trials compared with AB was obtained at 300 ms (Z = −2.39, P = 0.02) and 350 ms (Z = −2.34, P = 0.02), thus after T2 was presented.
Cluster 3 involved synchronization of left prefrontal to right posterior ROIs at theta, alpha, and gamma2 frequencies. Frontal to angular, supramarginal gyrus, precuneus, and occipital ROIs were also synchronized at theta and beta1 frequencies. Synchronization at gamma2 frequency was obtained between right frontal, left temporal, and parietal ROIs (Fig. 4). The PLV significantly differed between the AB and no AB conditions [Friedman test: X2(15) = 65.58, P = 0.00]. Wilcoxon tests revealed trials significantly varied at the following intervals post T1 presentation: 50 ms (Z = −2.56, P = 0.01), 100 ms (Z = −3.55, P = 0.00), 150 ms (Z = −3.29, P = 0.00), 250 ms (Z = −2.26, P = 0.02), 300 ms (Z = −3.96, P = 0.00), 350 ms (Z = −3.65, P = 0.00), 400 ms (Z = −3.27, P = 0.00), and 450 ms (Z = −2.2,7 P = 0.01), always indicating a higher PLV for AB trials compared with no-AB trials.
Cluster 4 involved significant synchronization at theta frequency connecting left frontal to frontal medial and right precuneus ROIs in addition to coupling between left parietal and right temporal ROIs at the same frequency. Alpha synchronization between right precuneus and right superior parietal ROIs was also evident, combined with connectivity at beta1 frequency linking right sensory motor cortex to right temporal ROI (Fig. 5). The Friedman test on PLV difference between no-AB and AB trials indicated a significant temporal evolution of ΔPLV [X2(15) = 67.81, P = 0.00]. Wilcoxon tests determined that the trial types significantly differed at the following intervals prior to and upon T1 onset: −200 ms (Z = −2.34, P = 0.02), −150 ms (Z = −2.47, P = 0.01), −100 ms (Z = −2.84, P = 0.00), −50 ms (Z = −3.42, P = 0.00), and 0 ms (Z = −3.35, P = 0.00). A PLV differed significantly between AB and no AB trials also at the following intervals post T1 presentation: 350 ms (Z = −3.29, P = 0.00), 400 ms (Z = −3.40, P = 0.00), 450 ms (Z = −3.83, P = 0.00), and 500 ms (Z = −3.2, P = 0.00). As can be seen in Figure 5, the PLV was higher in AB trials than in no-AB at all these time intervals.
Cluster 5 consisted of synchronization coupling left frontal to right posterior ROIs at theta, alpha, beta1, beta2, and gamma2 frequencies. The Friedman test on PLV difference between no-AB and AB trials indicated a significant temporal evolution of PLV [X2(15) = 67.04, P = 0.000]. Wilcoxon tests determined that AB trials exhibited a stronger PLV than no-AB trials in the 200-ms interval pre T1: −200 ms (Z = −3.24, P = 0.00), −150 ms (Z = −2.54, P = 0.01), −100 ms (Z = −3.35, P = 0.00), and −50 ms (Z = −2.67, P = 0.01). Trials also significantly differed at the following intervals post T1 presentation: 100 ms (Z = −2.82, P = 0.01), 150 ms (Z = −3.31, P = 0.00), 200 ms (Z = −4.08, P = 0.00), 250 ms (Z = −3.98, P = 0.00), 300 ms (Z = −2.47, P = 0.01), 350 ms (Z = −3.26, P = 0.00), and 400 ms (Z = −2.82, P = 0.01; Fig. 6). A PLV was higher in AB trials than in no-AB at all time intervals.
Statistical comparisons between no-AB trials of the critical dual target condition (Sn), salient dual target condition (SS), and single target conditions (T1 absent; -S, -n) were conducted using Friedman tests, which revealed the following results at the 5 identified cluster networks.
In cluster 1 (see Supplementary Fig. 3), for the SS condition, a PLV showed no significantly different time course than the no-AB condition [Friedman test: X2(15) = 12.01, P = 0.69]. In the -S condition, a PLV was significantly different in time course compared with no-AB trials [Friedman test: X2(15) = 30.99, P = 0.009], with a lower PLV particularly toward T2. Finally, for the -n condition, ΔPLV showed no significant effect [Friedman test: X2(15) = 19.61, P = 0.188].
In cluster 2, for the salient dual target condition, SS showed a significantly different time course of PLV than no-AB trials [Friedman test: X2(15) = 34.46, P = 0.003] as did the T1 absent conditions -n and -S [Friedman tests: X2(15) = 26.17, P = 0.036 and X2(15) = 31.64, P = 0.007, respectively]. For all 3 control conditions, there was a lower PLV obtained in the advent of T2 and in the time windows following T2 presentation compared with no-AB trials (see Supplementary Fig. 3).
There was a significantly different temporal evolution of PLV in the -n control condition compared with no-AB for cluster 3 [Friedman test: X2(15) = 26.88, P = 0.031], with a decrease of PLV between T1 and T2 presentation in -n, whereas in no-AB trials an increase of PLV was obtained (see Supplementary Fig. 3). SS and -S conditions did not significantly differ from no-AB trials regarding PLV time course [Friedman tests: X2(15) = 19.69, P = 0.186 and X2(15) = 23.60, P = 0.072, respectively].
At cluster 4, PLV displayed a significantly different time course for each of the 2 single target control conditions compared with no-AB trials [Friedman tests -S: X2(15) = 79.44, P = 0.000; and -n: X2(15) = 83.96, P = 0.000], with a lower PLV following T1 presentation (see Supplementary Fig. 3). Condition SS did not significantly differ from no-AB trials, though [SS: X2(15) = 24.03, P = 0.065].
Similar to cluster 4, control condition SS did not significantly differ in the PLV from no-AB trials in cluster 5 [Friedman test: X2(15) = 21.22, P = 0.13]. However, both single target conditions were demarcated by a significant smaller increase in PLV than the no-AB trial [Friedman tests -S: X2(15) = 34.67, P = 0.003 and -n: X2(15) = 47.9, P = 0.000; see Supplementary Fig. 3].
We examined differences in the synchronization of oscillatory brain activity between trials in which it proved impossible to report 2 targets correctly (i.e., AB) and trials in which participants were able to report both targets (no-AB). Single target report conditions were also analyzed (i.e., T1 absent and T2 present trials) to further explicate AB. A hierarchical cluster analysis was used to classify phase synchronization between paired ROIs across 6 frequency bands. This was achieved by clustering PLV differences (ΔPLV) calculated by subtracting the PLVs of no-AB minus AB trials for each frequency band. The analysis partitioned 5 clusters representing 5 separate, functionally distinct, cortical networks. The demonstration of distributed cortical networks active during dual target report is in accordance with earlier findings, suggesting that stimulus awareness during RSVP is dependent on the dynamic interaction between a number of widely distributed neural processes, rather than on the modulation of one single process or component (Gross et al. 2004; Sergent et al. 2005; Kranczioch et al. 2007). We posit that each of the identified distributed networks, across several frequency bands, is responsible for separate cognitive components involved in AB. Specifically, analyses revealed separate attentional and WM networks for both target and distractor processing enabling a determination of the contribution of both processes to AB, which is important for testing cognitive models of AB.
A Network for Target Encoding and Maintenance
Cluster 1 (Fig. 2) is postulated to represent a lone theta network synchronized across long-range connections from left frontal to right posterior-parietal ROIs. The network appears target oriented being related to successful target encoding and maintenance. The role of theta synchronization in memory is well established: an impressive body of research links increased theta power with memory processes such as rehearsal, short-term memory retention, encoding, and recognition of new episodic information (c.f., Klimesch 1999). The fronto-parietal topography demonstrated in this attentional network is reminiscent of previous research (e.g., Sarnthein et al. 1998; Sauseng et al. 2007) reporting anterior–posterior theta synchronization. These studies have established that such activity is likely to be involved in a memory-related binding mechanism connecting prefrontal regions—online information is stored here for continuous updating—with posterior regions where sensory information is stored (Klimesch et al. 2010). Two distinct trends are evident in the opposing synchronization patterns for AB and no-AB trials. In AB, a lack of synchronization in the 200 ms prior to T1 onset is contrasted by a sharp and significant increase at 200 ms post T1 (Fig. 2). This may reflect effortful T1 encoding processes still engaged upon T2 onset and manifest in a failure to report T2 correctly. This pattern is contrasted by reduced synchronization post T1 onset in the no-AB trials, followed by a steady increase in the interval prior to T2 onset. This pattern probably represents the system's ability to use the increased bottom-up salience of T1 to announce the T1 presence, thereby maintaining deployment of top-down attention for T2 onset and successful target report.
The analyses comparing SS, -S, and -n conditions with no-AB support this interpretation. There is no significant difference in temporal evolution of phase synchronization between no-AB and the conditions which either required top-down attention—due to the target lacking salience (-n)—or where the T2 needed to be top-down selected between 2 saliently presented targets (SS). However, if only one salient target appeared (-S), bottom-up processing of this stimulus did not require increased fronto-parietal theta synchronization. Consequently, phase synchronization in the -S condition was significantly lower in the T2 time window compared with the no-AB condition.
Tuning of an Attentional Filter
Cluster 2 is dominated by alpha synchronization (Fig. 3), which we interpret to reflect optimal utilization of an attentional filter in no-AB trials compared with AB trials. Synchronization is significantly lower than in AB trials until T2 onset where an increase in synchronization appears essential to successful T2 report. In contrast, despite increased synchronization throughout the pre and post T1 interval, T2 onset is accompanied by a significant reduction in synchronization in AB trials. The network's temporal evolution is similar to the distractor network outlined by Gross et al. (2004), where synchronization to distractors was typical of a failure to successfully report T2. We interpret the contrasting synchronization patterns as representing optimal use of the attentional filter. In the successful dual target report, one can detect the salient T1 stimulus by relying on its strong bottom-up signal and only tuning the attentional filter to T2 onset (this is accompanied by the increase in synchronization seen in no-AB trials prior to T2 onset). In contrast, in AB trials, the attentional filter is tuned to T1—an inefficient strategy—and upon T1 onset, a reduction in synchronization occurs. However, following a similar model outlined by Di Lollo et al. (2005), the filter is then tuned exogenously to distractor stimuli, leading to an increase in synchronization again and a failure to reset in time for T2 onset. An internal filter tuned to targets can again be established milliseconds later, but too late to ensure successful T2 report. The pattern here is also corroboratory of the attentional strategies evidenced in clusters 1 and 4, whereby activation to the bottom-up salience of T1 is optimized maintaining sufficient resources to engage top-down attention to T2 onset. A strategy of over investment of attentional resources to T1 is a predominant explanation for AB (Shapiro et al. 1997; Vogel et al. 1998; Bowman and Wyble 2007; Ghorashi et al. 2007; Olivers and Meeter 2008).
Comparisons between the no-AB condition and the control conditions -n, -S, and SS show that for all 3 control conditions, phase synchronization was significantly lower around T2 onset. This supports the idea that cluster 2 represents the tuning of an internal attentional filter necessary for efficient T2 processing after T1 presentation. In the conditions lacking a T1, this fine tuning would not be necessary. In conditions with a salient T2 (-S and SS), bottom-up processing will reduce the necessity for tuning such an internal attentional filter. Both processes result in lower phase synchronization for this cluster in the control conditions.
The topography of the alpha activity represented in this network is also reminiscent of the right-lateralized frontal attention system outlined by Corbetta and Shulman (2002). This network has been described as being involved in the preparation and application of top-down, goal-directed behavior such as the identification and selection of target stimuli in the present experiment (c.f., Mangun 2012). Shulman et al. (2010) have also shown that right frontal to left inferior temporal and bi-lateral connections are important in the detection of behaviorally relevant stimuli, especially when salient, as with the onset of T1 in the present paradigm. To ensure successful target report, both systems may be required to couple specifically at alpha frequency serving to inhibit distractors and ensure sufficient resources remain for detection and encoding of targets occurring at theta frequency. The left inferior temporal, middle temporal, and superior parietal regions implicated in this network may represent the verbalized nature of the task, whereby T2 was always a number. Increased alpha power and phase are known to correlate with a reduced perceptual report (Hanslmayr et al. 2005; Mathewson et al. 2009). Such evidence maps on well with the opposing pattern presented here in AB and no-AB trials. Hanslmayr et al. (2011) have also proposed a theoretical account for AB based on the impact of alpha activity upon the attentional and awareness system. The authors note that the propensity to present targets at 10 Hz in the AB literature is critically important to AB occurrence: stimulation at 10 Hz is known to induce steady-state evoked potentials, increasing alpha amplitude in occipital brain areas [Vialatte et al. (2010)]. The difficulty to encode T1 drives an increase in alpha power (and phase) corresponding to an internal state of attention, whereby in a paradigm such as AB the participant is more likely to misreport the second target. In contrast, periods of reduced alpha correspond to a state of external awareness and an increase in the likelihood of good attentional (and perceptual) performance. Alpha power and phase are known to vary in strength across time periods (Klimesch et al. 2007; Mathewson et al. 2009), and this may provide explanatory evidence for the finding that participants show AB on certain trials and not on others. Following on from this logic, Zauner et al. (2012) report results that alpha entrainment, as measured by the amplitude of the alpha evoked response, and the extent of alpha phase concentration, is larger for AB than for no AB trials and interpret this as a probable cause of AB.
A Salience Network—Deployment of Processing Resources
We interpret cluster 3 (Fig. 4) to represent the deployment of processing resources during the task, within a salient-specific network. In AB, a sustained increase in synchronization is apparent in the time window post T1- to T2-onset. This may represent overinvestment of resources to T1, observable in oscillatory activity, occurring at theta, alpha, beta1, and gamma2 frequencies, and resulting in a failure to report T2. In contrast, in no-AB trials, only a short burst of phase synchronization following presentation of the salient T1 was obtained.
When no-AB trials were compared with the control conditions (SS,-S, and -n), no significantly different temporal evolution of phase synchronization was obtained if there was at least one salient target in the rapid serial presentation stream (conditions SS and -S). With only neutral stimuli presented (-n), significantly lower phase synchronization was found compared with no-AB trials (in which T1 was salient), hence, suggesting that cluster 3 is associated with the processing of salient stimuli. AB trials show very high synchronization after T1, with a strong peak after T2 presentation, suggesting that the participants tried to identify T2 based on salience. Since T2 was not presented in a different color in the AB condition, this leads to a failure to report T2. The topography of the cluster 3 network is consistent with the salience network reported in neuroimaging studies, including anterior cingulate cortex, dorsolateral PFC, and posterior-parietal cortex (c.f., Buckner et al. 1996; Luck et al. 1997; Braver et al. 2001, 2003; Asplund et al. 2010; Mangun 2012; Sadaghiani et al. 2012; also see evidence presented in the discussion on the attentional filter, above).
A Network for Salience-Based Processing of T1
Cluster 4 (Fig. 5) provides evidence of an additional, separate attentional network. Optimal use of the salient T1 in a bottom-up manner is represented by significantly lower synchronization in the pre T1 interval followed by an increase in coherence prior to T2 onset in no-AB trials. The reverse pattern is evidenced in AB trials resulting in attention failing to encode T2 due to on-going T1 processing. Target encoding is reflected in the topography of this predominant fronto-temporo-posterior theta network, representing top-down frontal support for target detection and selection.
This interpretation is supported by the comparison between no-AB trials and the SS, -S, and -n conditions. Those control conditions lacking a T1 (-S and -n) differ significantly from no-AB regarding phase synchronization during the time windows that (would have) followed T1. SS, the control condition, containing a (salient) T1 did not show any significantly different temporal evolution of phase synchronization, indicating that in this condition (similar to no-AB) salience was optimally used for processing T1.
A Network for Tuning of an External Attentional Filter
The temporal synchronization evident in cluster 5 (Fig. 6) indicates that increased synchronization here is linked to a failure to maintain sufficient resources for T2 report. Extensive synchronization occurs across theta, alpha, beta1, beta2, and gamma2 frequencies. In no-AB trials, synchronization within this cluster increases steadily but very moderately until T1 had been presented and then decreases again. In contrast, in AB trials, there is a sharper increase of synchronization toward T1, followed by a short dip and then a strong enhancement of phase synchronization toward T2. We assume that this cluster might reflect processes described in the external attentional filter in the model suggested by Di Lollo et al. (2005). Due to optimal usage of the internal attentional filter (cluster 2) and optimal processing of the salient T1 (see cluster 4), there is modest requirement for the external attentional filter in no-AB trials. A lack of adequate tuning of the internal attentional filter (reflected by cluster 2) in the advent of T2, however, is substituted by very strong synchronization within cluster 5 in AB trials.
The comparison between no-AB trials and the control conditions suggests that whenever T1 is missing (-S and -n), phase synchronization in cluster 5 increases slowly, and moderately, without plateauing. Without a T1, there is reduced requirement for the external attentional filter being tuned than in no-AB trials. In the SS condition, very similar processes can be observed as in the no-AB trials, including tuning of an external attention filter as reflected by cluster 5.
The topography of cluster 5 includes synchronization between right inferior frontal and right inferior temporal cortex, resembling the right hemispheric ventral fronto-parietal attention network in Corbetta and Shulman's (2002) attention model (the stimulus-driven attentional network). While cluster 2 is more similar to the dorsal fronto-parietal attentional network associated with the voluntary control of attention. This further supports the idea that these 2 networks may represent the external and the internal attentional filter in Di Lollo et al.'s (2005) AB model, respectively.
Our results argue that to successfully overcome the AB may require activation of a number of distinct, parallel cortical networks, in which attention and WM are deployed. The analytical approach applied here allowed the separation of the electrophysiological correlates of cognitive processes involved in the AB. We identified parallel cortical connectivity patterns most likely associated with (1) target encoding into WM, (2) tuning of an attentional filter, (3) target selection processes, and (4) the recruitment of cognitive resources during the task.
There are numerous theories explaining the cognitive basis of AB. Our data support several of these models, suggesting that the cognitive processes described by these models can in parallel contribute to target identification in the AB paradigm. Activity of a distributed theta network indicated by cluster 1 in our study concur well with the Two-Stage Model (Chun and Potter 1995) as well as with the Interference Model (Shapiro et al. 1997); both of them postulating nil- or attenuated processing capacities for T2 while WM encoding takes place for T1. Our results suggest that when a left prefrontal to right parietal and temporal theta network is highly synchronized in response to T1, indicating intensive processing of T1 in WM, this network will desynchronize when T2 is presented and therefore T2 will not be encoded properly in WM. Contrastingly, trials in which T2 can be reported exhibit low synchronicity of this fronto-temporo-parietal theta network after T1 onset, but increased synchronization when T2 is presented, thus, indicating proper processing of T2 in WM without interference from T1.
A more recent model of AB has been proposed by Di Lollo et al. (2005). This model suggests that an endogenous attentional filter can be tuned to T1, so that the first target within the RSVP stream will be identified. After detection of T1, attentional filter switches to an exogenous mode which is constantly updated by features of currently processed stimuli (e.g., distracters between targets). Therefore, this exogenous attentional filter will not be tuned to properties of T2, which will not be detected. This is exactly what is reflected by synchronization patterns of cluster 2 in our study. In trials where T2 was not reported correctly, there was a transient increase in synchronization of a predominantly right fronto-parietal alpha network prior to T1. This indicates the establishment of an endogenous attentional filter [network topography and involved frequencies are well in line with literature on the cortical basis of attentional control, see, e.g., Corbetta and Shulman (2002) and Sauseng et al. (2005)]. Post T1 onset, there is a swift desynchronization of this alpha network followed by a resynchronization during distracter presentation. This might indicate a switch from an endogenous to an exogenous attentional filter as predicted by Di Lollo et al.'s model. Since this endogenous attention filter is not tuned to T2, the second target will be missed which is reflected by a strong desynchronization of the alpha network during T2 presentation. Re-establishment of an endogenous attention filter is only achieved about 200 ms after T2 onset—too late to be effective. In trials where T1 detection was based more on bottom-up processing, an endogenous attentional filter was tuned for T2 detection as indicated by increased synchronization in the relevant alpha network prior to T2 presentation leading to successful T2 report. These results can be considered as direct neural evidence for Di Lollo et al.'s model of attention filter tuning.
Another recent model on AB, the episodic Simultaneous Type/Serial Token model (eSTST; Bowman and Wyble 2007; Wyble et al. 2009), is supported by our data too. eSTST proposes that the visual system is designed to flexibly mediate the allocation of attention through dynamic excitation and inhibition of the mechanism responsible for the deployment of temporal allocation. The mechanism is triggered by potentially relevant stimuli (visual “types” corresponding to stimulus identity) and once instigated a process of “tokenization” (corresponding to a temporal order) begins creating an episodic memory representation in WM. By this account, AB is the outcome of a cognitive strategy to ensure each visual type remains episodically distinct. Such a theory not only provides an explanation for why serial target report is possible under RSVP conditions (e.g., Di Lollo et al. 2005), but also explains why dual target report is often not possible if there is a non-target interval between targets. For example, in a target–target–target stream, the suppression elicited during the encoding of T1 is overcome by the excitation associated with the trailing targets. In contrast, in target–distractor–target paradigms, the intervening distractor or blank screen (e.g., Nieuwenstein et al. 2009) suppresses attention producing a failure to report T2. The target excitation and encoder inhibition processes vary systematically in strength over trials explaining why the AB occurs on some trials and not on others (Wyble et al. 2009). The present paradigm served to exploit the eSTST account of the AB through the provision of a condition where T1 appeared in a salient contrast while T2 did not. This modification was utilized to test the hypothesis that manipulating T1 salience would increase its episodic distinctiveness in the task. We have interpreted that, during dual target report, observers may conserve top-down attention for T2 search by relying on the strong bottom-up signal evoked by T1 during successful report. The behavioral data substantiated this conclusion with the critical Sn condition evidencing a drop to 63% accurate responding. In contrast, in the SS condition, both targets appeared in a salient contrast—boosting episodic distinctiveness—and dual target report was 82%.
To conclude, the present study has outlined 5 distinct cortical networks that are synchronized at various frequencies during dual target report reflecting changes in the attentional demands of the task directly related to behavioral performance. Analyses enabled a dissociation of the involved processes and a determination of their contribution to AB which we suggest as being important for testing cognitive models of AB in future research.
This work was supported by an Irish Research Council (IRC) scholarship awarded to M.G.
Conflict of Interest: None declared.