Abstract

Visual working memory (VWM) sustains information online as integrated object representations. Neuronal mechanisms supporting the maintenance of feature-specific information have remained unidentified. Synchronized oscillations in the gamma band (30–120 Hz) characterize VWM retention and predict task performance, but whether these oscillations are specific to memorized features and VWM contents or underlie general executive VWM functions is not known. In the present study, we investigated whether gamma oscillations reflect the maintenance of feature-specific information in VWM. Concurrent magneto- and electroencephalography was recorded while subjects memorized different object features or feature conjunctions in identical VWM experiments. Using a data-driven source analysis approach, we show that the strength, load-dependence, and source topographies of gamma oscillations in the visual cortex differentiate these memorized features. Load-dependence of gamma oscillations in feature-specific visual and prefrontal areas also predicts VWM accuracy. Furthermore, corroborating the hypothesis that gamma oscillations support the perceptual binding of feature-specific neuronal assemblies, we also show that VWM for color–location conjunctions is associated with stronger gamma oscillations than that for these features separately. Gamma oscillations hence support the maintenance of feature-specific information and reflect VWM contents. The results also suggest that gamma oscillations contribute to feature binding in the formation of memory representations.

Introduction

Visual working memory (VWM) sustains visual information online for subsequent actions and cognitive operations and is one of the elementary cognitive abilities. Both theoretical work (Baddeley 1996, 2003) and functional magnetic resonance imaging (fMRI) studies (Prabhakaran et al. 2000; Rowe et al. 2000; Munk et al. 2002; Pessoa et al. 2002; Sakai et al. 2002; Linden et al. 2003; Mohr et al. 2006) suggest that VWM comprises both central executive and visual-sensory components. Additionally, behavioral data suggest that VWM retains visual information in the form of integrated object representations rather than as collections of sensory features (Luck and Vogel 1997; Wheeler and Treisman 2002; Zhang and Luck 2008). Whereas the visual system constructs and represents visual object information (Kravitz et al. 2013), fronto-parietal (FP) attentional networks coordinate executive and attentional functions (Fuster 2000; Kastner and Ungerleider 2000; Corbetta and Shulman 2002).

Gamma-band (γ, 30–120 Hz) neuronal synchronization is one putative mechanism that could support the integration of anatomically distributed feature processing into neuronal assemblies representing perceptual objects. Single-unit and field-potential recordings from cat and monkey visual cortex show that neuronal spike synchronization together with gamma-band oscillations correlate with perceptual feature binding (Singer and Gray 1995; Singer 2013). However, a number of invasive studies also link beta- (14–30 Hz) and gamma-band synchronization with the coordination of attentional interactions in large-scale networks including both visual and FP regions (Fries et al. 2002; Fries 2005; Buschman and Miller 2007; Gregoriou et al. 2009; Bosman et al. 2012). Hence, synchronized beta and gamma oscillations could maintain both the representations of sensory information and executive functions governing the maintenance of that information in VWM.

Indeed, human source-reconstructed magnetoencephalography (MEG) and electroencephalography (EEG), as well as invasive intracranial recordings, show that large-scale (Axmacher et al. 2008; Palva et al. 2010) and local (Jokisch and Jensen 2007; Medendorp et al. 2007; Sauseng et al. 2009; Morgan et al. 2011; Palva et al. 2011; Roux et al. 2012) gamma-band synchronization characterizes VWM maintenance in both visual and FP areas. Furthermore, the VWM load-dependence of gamma-band amplitudes in visual but also in FP regions predicts individual behavioral accuracy in delayed match-to-sample (DMS) VWM tasks (Sauseng et al. 2009; Palva et al. 2011; Roux et al. 2012). This correlation has been taken to imply that the amplitudes of beta and gamma oscillations reflect the sensory information sustained in VWM and hence the “contents” of VWM, although there is little direct evidence supporting this hypothesis. It thus remains unclear whether gamma oscillations reflect the visual information per se maintained in VWM or more generic attentional and executive functions that may also be predictive of individual VWM performance (Magen et al. 2009).

Evidence supporting the role of gamma-band synchronization in maintaining sensory information, rather than underlying attentional and executive functions, is apparent at the sensor-level EEG (Tallon-Baudry and Bertrand 1999) and MEG (Kaiser et al. 2008, 2009) investigations, which imply that local gamma-band oscillations reflect the WM maintenance of both coherent visual objects and sound features. Along these lines, a study using MEG and source modeling shows that the strength of gamma oscillations is correlated with perceptual integration demands during VWM in posterior parietal cortex (Morgan et al. 2011). However, a recent EEG sensor-level study shows that alpha- (8–12 Hz) rather than beta- and gamma-band oscillations are predictive of VWM contents (Anderson et al. 2014).

Hence, if beta and gamma oscillations are to reflect VWM contents and the representations of maintained visual information, these oscillations should reflect the memorized features with specificity to those features, within the visual cortical regions where these features are processed. Additionally, these oscillations should be the predominant feature-specific characteristic during the VWM retention period and predict the behavioral accuracy in the memorization of the features.

To test the hypothesis that gamma oscillations reflect the maintenance of distinct sensory features and the binding of feature information into integrated memory representations, we recorded ongoing brain activity with concurrent MEG and EEG (M/EEG) and used a parametric VWM task where the subjects memorized different object features or their conjunctions while the physical stimuli and their statistical properties remained identical. With these data, using M/EEG source reconstruction that retains individual neuroanatomical localization accuracy and a data-driven analysis approach, we asked whether feature-specific oscillation amplitude effects within the 3- to 120-Hz frequency range would characterize the VWM retention interval, and whether such effects would be observed in the visual cortical regions conceivably processing these memory representations.

Methods

Task and Stimuli

The VWM experiments comprised DMS tasks where, in each trial, the subjects were presented a “Sample” stimulus (S1, duration 150 ms) containing randomly either 2 or 4 objects (i.e., VWM loads 2 and 4) either to the left or right side of a fixation dot at the center of the display. The objects were 3- to 6-sided random polygons, which could appear at 15 different locations in each hemifield and which had easily separable but not primary or secondary colors. In separate experimental blocks of Experiment 1, the subjects memorized the shapes, colors, or spatial locations of the objects in S1. In Experiment 2, the subjects memorized the color–location conjunctions of these objects. In each trial, for both experiments, S1 was followed by a 2.05-s retention period after which a “Test” stimulus (S2, duration 500 ms) containing one object was presented in the same hemifield to S1. Subjects indicated whether the task-relevant feature of any of the objects in S1 was the same as or different from that feature in S2 (Fig. 1A) with a forced-choice left- or right-hand thumb twitch counterbalanced across subjects. The intertrial interval varied randomly from 2 to 3 s. The features were automatically and randomly generated separately for each trial such that each S1 and each object therein were unique but statistically identical across all conditions. In Experiment 1, the features of the object in S2 always had an independent 50% probability of being identical with any of the features in S1, whereas in Experiment 2, the conjunction of features had a 40% probability of being identical with those in any one of the objects in S1. The size of the area where the objects were presented was 7.3° × 7.3° and the objects on average spanned an area of 0.65° × 0.65°. The minimum center-to-center distance between the objects was 1.29° and maximum 3.87°. Prior to each experiment, the subjects received task instructions and a practice session.

Figure 1.

Gamma and beta oscillation amplitudes reflect object features maintained in VWM. (A) A schematic of the VWM task with a load of 4 objects. (B) HRs and RTs (mean ± SD) for all tasks (Sh, shape; Co, color; Lo, location; Con, conjunction). Lines indicate significant differences across loads and conditions (***P < 0.001; **P < 0.01; *P < 0.05; Bonferroni-corrected with N = 10 tests. Light gray indicates uncorrected significance values). (C) TFRs indicate for each TF element the fraction of brain regions (P, color scale) with a statistically significant effect in the comparison of baseline-corrected oscillation amplitudes between shape, color, and location conditions (one-way ANOVA, P < 0.05, FDR-corrected). The ANOVA showed a feature effect in beta (β, 14–30 Hz) and low-gamma (lγ, 40–72 Hz) band amplitudes during the retention interval. Colored rectangles indicate the TF ROIs selected for visualization of the cortical topography of these effects (see D). (D) Visualization of the lγ- and β-band amplitude ANOVA effects on inflated flattened cortical surfaces shows that the strength of oscillations in visual and frontal cortices is feature specific. Colors indicate the fraction of significant TF elements in the TF-ROI (P). The colored lines define the boundaries of visual cortex subdivisions and the grayscale lines the attentional systems defined by intrinsic connectivity patterns (Fig. 2C).

Figure 1.

Gamma and beta oscillation amplitudes reflect object features maintained in VWM. (A) A schematic of the VWM task with a load of 4 objects. (B) HRs and RTs (mean ± SD) for all tasks (Sh, shape; Co, color; Lo, location; Con, conjunction). Lines indicate significant differences across loads and conditions (***P < 0.001; **P < 0.01; *P < 0.05; Bonferroni-corrected with N = 10 tests. Light gray indicates uncorrected significance values). (C) TFRs indicate for each TF element the fraction of brain regions (P, color scale) with a statistically significant effect in the comparison of baseline-corrected oscillation amplitudes between shape, color, and location conditions (one-way ANOVA, P < 0.05, FDR-corrected). The ANOVA showed a feature effect in beta (β, 14–30 Hz) and low-gamma (lγ, 40–72 Hz) band amplitudes during the retention interval. Colored rectangles indicate the TF ROIs selected for visualization of the cortical topography of these effects (see D). (D) Visualization of the lγ- and β-band amplitude ANOVA effects on inflated flattened cortical surfaces shows that the strength of oscillations in visual and frontal cortices is feature specific. Colors indicate the fraction of significant TF elements in the TF-ROI (P). The colored lines define the boundaries of visual cortex subdivisions and the grayscale lines the attentional systems defined by intrinsic connectivity patterns (Fig. 2C).

M/EEG recordings were carried out in randomly ordered blocks so that each block comprised 160 trials and 5 such blocks were collected for each subject and memorized feature in Experiment 1. Experiment 2 was otherwise identical, but the subjects memorized the conjunctions of colors and locations and 10 blocks of 160 trials were collected for each subject. The order of blocks was randomized in both experiments. The total number of trials collected for each feature condition was 800 in Experiment 1 and 1600 in Experiment 2. After rejection of trials contaminated by eye movement artifacts, 656 ± 139 (mean ± standard deviations, SDs), 550 ± 211, 610 ± 153, and 1358 ± 205 trials remained in shape, color, location, and conjunction conditions, respectively and the memory loads combined. Both experiments spanned a minimum of 2 recording sessions performed on separate days. Within each recording day, subjects were given ample breaks to rest between the blocks.

This study was approved by the Ethical Committee of Helsinki University Central Hospital and conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from each subject prior to the experiment.

Subjects and Recordings

Concurrent M/EEG was recorded during VWM tasks from 13 healthy right-handed volunteers [29 ± 6 years (mean ± SD), 7 females] in Experiment 1 and from 8 subjects (28 ± 4 years, 3 females) from the same cohort in Experiment 2. Subjects had normal or corrected-to-normal vision and none of the subjects had a history of neurological or neuropsychiatric disorders. M/EEG data were recorded with 204 planar gradiometers, 102 magnetometers, and 60 EEG electrodes (Elekta Neuromag Ltd, Finland) at a 600-Hz sampling rate. The thumb-twitch responses were recorded with electromyography (EMG) and ocular artifacts with electrooculogram. The Maxfilter software (Elekta Neuromag Ltd) was used to suppress extracranial noise and to colocalize the recordings from different sessions and subjects in signal space. T1-weighted anatomical MRI scans for cortical surface reconstruction models were obtained at a resolution of a ≤1 × 1 × 1 mm with a 1.5-T MRI scanner (Siemens, Germany).

Behavioral Performance

Hit rates (HRs) were estimated as the proportion of correct responses from all trials. Reaction times (RTs) were estimated from the onset of S2 to the onset of the EMG activity exceeding 7 SD of baseline EMG level. The EMG recordings had a high signal-to-noise ratio and reliably indicated the onset of muscle activity at each thumb twitch. HRs and RTs were then computed separately for each feature condition and memory load (Fig. 1B). Only artifact-free trials with accepted behavioral responses were included in the analysis.

Source Reconstruction

M/EEG source reconstruction, surface parcellations, filtering, and data analysis followed earlier procedures and was performed in a data-driven manner (Palva et al. 2010, 2011; Rouhinen et al. 2013) to exclude the possibility that a hypothesis bias could influence the outcome of the analysis. Briefly, FreeSurfer (http://surfer.nmr.mgh.harvard.edu/) was used for automatic volumetric segmentation of the MRI data, surface reconstruction, flattening, and cortical parcellation, as well as neuroanatomical labeling with the Destrieux atlas (Dale et al. 1999; Fischl et al. 2002). MNE Suite (http://www.nmr.mgh.harvard.edu/martinos/userInfo/data/sofMNE.php) was used in M/EEG-MRI colocalization, to create the three-layer boundary element head models and the cortically constrained gray-matter surface source models, and for the preparation of the forward and minimum-norm-estimate (MNE) inverse operators. MEG and EEG data (204 MEG planar gradiometers, 102 MEG magnetometers, and 60 EEG electrodes) were hence integrated at the inverse transform stage of data analysis (Hämäläinen and Sarvas 1989; Hämäläinen and Ilmoniemi 1994). The source models had dipole orientations fixed to the pial surface normals and a 7-mm interdipole separation throughout the cortex, which yielded models containing 6000–8000 source vertices. The preprocessed M/EEG time series data from each separate channel were Morlet-wavelet filtered into 31 frequency bands, fmin = 3 Hz to fmax = 120 Hz with the Morlet time–frequency compromise parameter m, being m = 5. Inverse operators were prepared separately for each wavelet filter frequency by using the pre-S1 baseline period of all trials for estimating the noise covariance matrix.

Oscillation Amplitude Analysis

To reconstruct ongoing cortical amplitude dynamics, we used the frequency band-specific inverse operators to transform filtered complex single-trial M/EEG time series to source vertex time series and then collapsed these to time series of cortical parcels in a 400-parcel collection. The collapsing was performed with sparse-weighted collapse operators that maximized the reconstruction accuracy in each subject's source space (Korhonen et al. 2014). We then obtained the band amplitude envelopes from complex parcel time series and averaged the envelopes across trials (Tallon-Baudry et al. 1996; Palva et al. 2005) for each feature and load condition. Trial-averaged oscillation amplitudes were decimated into mean amplitudes of 32 × 300 ms time windows with 200 ms overlap from 0.7 s prior to S1 onset to 2.7 s after S1 onset. Prior to statistics, we collapsed the 400 parcel data to a coarser parcellation of 198 parcels to improve statistical stability, and hence the amplitude data comprised of a volume of 13 subjects, 198 parcels, 8 conditions, 31 frequency bands, and 32 time windows.

Statistical Analysis

The group statistics across subjects and between conditions or between the baseline and the retention period were performed separately for each frequency, time window, and cortical parcel. False discovery rate (FDR) correction was then performed for multiple comparisons in the time window parcel plane separately for each frequency. Both the 400- and the 198-parcel parcellations were derived by iteratively splitting the largest parcels of the Destrieux atlas at their most elongated axis at the group level (Palva et al. 2010, 2011), which hence yields neuroanatomical labels and an anatomical frame common to all subjects while completely retaining the individual localization accuracy.

All statistical tests were performed after a baseline subtraction in each condition of the interval from −0.5 to −0.1 s to S1. The influence of the memorized feature on oscillation amplitudes was estimated with one-way analysis of variance (ANOVA) across shape, color, and location conditions (P < 0.05, FDR-corrected) (Fig. 1C). The VWM task effect against baseline, VWM load effect between 4- and 2-object memory loads, and task effect between conditions (Figs 2, 3, and 5–7, as well as Supplementary Figs 2–4) were tested with paired t-tests after baseline subtraction (P < 0.05, FDR-corrected). To estimate the correlation between individual VWM performance and oscillation amplitude dynamics, we first obtained the difference of 2- and 4-load condition HRs for each individual and then estimated the correlation of these differences with the oscillation amplitude load effects (“4” − “2”) with a Pearson correlation test (P < 0.05, FDR-corrected, Fig. 4).

Figure 2.

Beta (β)- and gamma (γ)-band amplitudes in feature-specific visual cortical regions and attentional systems are correlated with memory load. (A) TFR of load effects for the shape, color, and location VWM tasks (t-test for the difference of baseline-corrected 4- and 2-object conditions, P < 0.05, FDR-corrected). Low-gamma (lγ)- and high-gamma- (hγ, 80−120 Hz) band amplitudes were positively correlated with memory load for shape VWM during the late retention and for color VWM during early retention, while α- and β-band amplitudes were negatively correlated with load especially in color and location conditions. Red-yellow colors define the fraction of cortical parcels where the amplitude was greater (positive tail, P+) and blue colors where the amplitude was smaller (negative tail, P−) for the 4- than 2-object memory load. (B) The cortical localization of load-dependent modulations in TF selections for shape, color, and location VWM tasks. Load-dependent strengthening of hγ and lγ of shape and color VWM reveal modulations in the visual and frontal cortical regions. Load-dependent strengthening of α–β (9–20 Hz) for shape VWM and β (20–35/30 Hz) for color and location VWM, respectively, are observed in early visual regions. Warm colors define the fraction of positive and cold colors the fraction of negative TF elements. (C) The colored lines define boundaries for functional ROIs [blue line, ventral; pink, dorsal; yellow, V4, V8; dark green, V1–V3; green, LO (lateral occipital); Tootell and Hadjikhani 2001; Hansen et al. 2007; Kravitz et al. 2013]. Gray, dark gray, and black lines define DA, VA, and FP networks, respectively.

Figure 2.

Beta (β)- and gamma (γ)-band amplitudes in feature-specific visual cortical regions and attentional systems are correlated with memory load. (A) TFR of load effects for the shape, color, and location VWM tasks (t-test for the difference of baseline-corrected 4- and 2-object conditions, P < 0.05, FDR-corrected). Low-gamma (lγ)- and high-gamma- (hγ, 80−120 Hz) band amplitudes were positively correlated with memory load for shape VWM during the late retention and for color VWM during early retention, while α- and β-band amplitudes were negatively correlated with load especially in color and location conditions. Red-yellow colors define the fraction of cortical parcels where the amplitude was greater (positive tail, P+) and blue colors where the amplitude was smaller (negative tail, P−) for the 4- than 2-object memory load. (B) The cortical localization of load-dependent modulations in TF selections for shape, color, and location VWM tasks. Load-dependent strengthening of hγ and lγ of shape and color VWM reveal modulations in the visual and frontal cortical regions. Load-dependent strengthening of α–β (9–20 Hz) for shape VWM and β (20–35/30 Hz) for color and location VWM, respectively, are observed in early visual regions. Warm colors define the fraction of positive and cold colors the fraction of negative TF elements. (C) The colored lines define boundaries for functional ROIs [blue line, ventral; pink, dorsal; yellow, V4, V8; dark green, V1–V3; green, LO (lateral occipital); Tootell and Hadjikhani 2001; Hansen et al. 2007; Kravitz et al. 2013]. Gray, dark gray, and black lines define DA, VA, and FP networks, respectively.

Figure 3.

Normalized amplitudes and standard error of means (SEMs) for load-dependent modulations across functional ROIs for shape (black), color (red), and location (gray) VWM conditions for high gamma (hγ), low gamma (lγ), and beta (β; here 17–35 Hz) bands for both early (0.5–1.2 s) and late (1.2–2 s) retention intervals. Lines above the bars indicate significant differences in the mean amplitudes between ROIs for each VWM condition. Lines below bars indicate significant differences in the mean amplitudes between conditions for each ROI. Solid lines indicate significant differences across loads and conditions (t-test, P < 0.05; Bonferroni-corrected with N = 10 tests for the lines above, N = 3 tests for the lines below). Dashed lines indicate uncorrected significance values. See Figure 2C for the ROI selections.

Figure 3.

Normalized amplitudes and standard error of means (SEMs) for load-dependent modulations across functional ROIs for shape (black), color (red), and location (gray) VWM conditions for high gamma (hγ), low gamma (lγ), and beta (β; here 17–35 Hz) bands for both early (0.5–1.2 s) and late (1.2–2 s) retention intervals. Lines above the bars indicate significant differences in the mean amplitudes between ROIs for each VWM condition. Lines below bars indicate significant differences in the mean amplitudes between conditions for each ROI. Solid lines indicate significant differences across loads and conditions (t-test, P < 0.05; Bonferroni-corrected with N = 10 tests for the lines above, N = 3 tests for the lines below). Dashed lines indicate uncorrected significance values. See Figure 2C for the ROI selections.

Figure 4.

Individual gamma (γ)-band load effects predict the load-related decline in behavioral accuracy. (A) TFRs of the correlation of load-dependent amplitude modulations with the change of behavioral accuracy between the 2- and 4-object VWM loads for shape VWM revealed that low gamma (lγ) and high gamma (hγ) oscillations were correlated with individual load-dependent decline in behavioral accuracy. (B) Correlation of the load-dependence of amplitudes and the decline in task performance for color VWM. (C) The load-dependence of hγ and lγ (here 30–60 Hz) amplitudes predicted good behavioral VWM performance in ventral stream object-representation regions as well as in FP and VA networks. (D) For color VWM, the change in behavioral accuracy was positively correlated with the load-dependence of hγ amplitudes in right-hemispheric ventral stream regions and negatively with β–lγ (18–40 Hz) amplitudes in early and ventral visual areas.

Figure 4.

Individual gamma (γ)-band load effects predict the load-related decline in behavioral accuracy. (A) TFRs of the correlation of load-dependent amplitude modulations with the change of behavioral accuracy between the 2- and 4-object VWM loads for shape VWM revealed that low gamma (lγ) and high gamma (hγ) oscillations were correlated with individual load-dependent decline in behavioral accuracy. (B) Correlation of the load-dependence of amplitudes and the decline in task performance for color VWM. (C) The load-dependence of hγ and lγ (here 30–60 Hz) amplitudes predicted good behavioral VWM performance in ventral stream object-representation regions as well as in FP and VA networks. (D) For color VWM, the change in behavioral accuracy was positively correlated with the load-dependence of hγ amplitudes in right-hemispheric ventral stream regions and negatively with β–lγ (18–40 Hz) amplitudes in early and ventral visual areas.

Time–Frequency and Anatomical Visualization of the Fractions of Significant Effects

We used statistical testing across all brain regions, frequency bands, and time windows to not only test the primary hypothesis, but also to concurrently test all alternative hypotheses so that missing a robust but unhypothesized phenomenon would not lead to misleading interpretations of the results (Palva et al. 2010, 2011; Rouhinen et al. 2013). That is, this data-driven approach is intended to reveal not only whether the hypothesized phenomenon is observed in the data but also how robust that phenomenon is by comparison to other significant phenomena at the level of the whole cortex. To obtain a comprehensive and unbiased overview of all results for a given test, we used time–frequency representations (TFRs) showing for each time–frequency element the fraction of parcels out of all parcels where a statistically significant positive and/or negative effect was observed. To reveal the brain areas accounting for the most prominent effects revealed by these TFRs, we selected a time–frequency (TF) regions of interest (ROIs) and displayed the fraction of significant TF elements of all elements in that TF ROI for each parcel. These data were visualized on an inflated flattened cortical surface, which allows the presentation of all data, and specifically of all visual cortical regions distributed in both medial and lateral surfaces, in a single anatomical plot, as opposed to a three-dimensional plot where a large fraction of cortical surface is hidden regardless of orientation.

Functional Landmarks

To support the interpretation of the results in terms of the identities of functional subdivisions of the visual cortex, we used prior fMRI studies (Tootell and Hadjikhani 2001; Hansen et al. 2007; Kravitz et al. 2013) to divide the visual cortex into regions that putatively underlie the processing and representation of location, color, and object information: dorsal, V4 and V8, and ventral areas, respectively. We also show here the tentative boundaries for early (V1–V3) and lateral-occipital (LO) visual areas. To identify the brain regions putatively supporting attentional or executive control functions, we used the boundaries of dorsal attention (DA), ventral attention (VA), and FP networks defined by fMRI intrinsic connectivity patterns (Yeo et al. 2011) and colocalized with our individual subjects' cortical surfaces with Freesurfer.

To corroborate the parcel resolution analyses with that of the level of visual system ROIs, we obtained for each subject the means of parcel amplitudes in the 5 visual regions (early, V4 and V8, LO, ventral, and dorsal areas, Fig. 2C), normalized these amplitudes, and compared them among ROIs and conditions (t-test, P < 0.05, FDR-corrected, Figs 3 and 7B).

Results

Psychophysics Suggest That Feature-Specific Mechanisms Play a Role in VWM Maintenance

The main objective of this study was to investigate whether the amplitudes of gamma oscillations reflect the maintenance of feature-specific information in VWM, which would strongly suggest that gamma oscillations play a direct role in supporting the maintenance of VWM contents. To this end, we used a feature-selective DMS VWM task where in each trial, subjects memorized the shapes, colors, or locations of 2 or 4 objects in a “Sample” stimulus (S1) and were then presented a single object in a “Test” stimulus (S2) after a 2.05-s memory retention interval. The subjects reported whether the task-relevant feature in S2 was the same or different than that of any object in S1 (Fig. 1A). Both physical and statistical experimental parameters were hence identical throughout the shape, color, and location VWM tasks, recorded in randomized separate short sessions. This paradigm could hence reveal content-specific VWM maintenance while simultaneously controlling the attentional/executive mechanisms of VWM maintenance and excluding variability in experimental parameters as confounding factors. We measured behavioral performance with accuracy (or HR) and mean RT for each task condition and memory load separately (Fig. 1B). The RTs increased and the HRs decreased with increasing memory load in each task (t-test, P < 0.05, FDR-corrected). Significant, albeit small, differences in HRs and RTs between the shape and the color or location VWM tasks suggest that the shape task was slightly more difficult (Fig. 1B, t-test, P < 0.05, Bonferroni-corrected). Similar mean HRs for the 4-object load in color and 2-object load in shape task show that these tasks were equal in difficulty (color HR: 70.4%, RT: 0.79 s; shape HR: 70.1%, RT: 0.72 s; P > 0.8 for HR and P < 0.05 for RT). We then used Pearson product–moment correlation tests to estimate whether individual behavioral performance was correlated between 2- and 4-object loads in each task (Supplementary Fig. 1A) or between tasks (Supplementary Fig. 1B). This analysis revealed strong correlations between 2- and 4-object HRs for all conditions (Supplementary Fig. 1A, P < 0.001, FDR-corrected, r > 0.8). However, within-subject correlations between tasks were rare and observed only between color and location tasks for 2 object condition (Supplementary Fig. 1B, P < 0.05, FDR-corrected, r = 0.7), as well as between the load-dependent decline in performance between the shape and location tasks (P = 0.01, uncorrected, r = −0.7). The near absence of significant interfeature correlations indicated that, in addition to a central trait-like WM capacity (Heck et al. 2014), distinct and feature-specific neuronal mechanisms also contributed to the maintenance of feature-specific VWM and the individual memory capacity limits therein.

Retention Period Beta and Gamma Amplitudes Reflect the Maintenance of Stimulus-Specific Features in VWM

The main objective of the ensuing analysis was to investigate whether gamma oscillations reflect the maintenance of distinct features in VWM. We first addressed this question by estimating the dependence of oscillation amplitude on the features maintained in memory with one-way ANOVA for the 2-object VWM load conditions as the performance therein was good for all features and hence did not include difficulty as a confounder. To avoid the hypothesis per se biasing the data analysis through affecting the selection of cortical regions or frequencies of interest (Kriegeskorte et al. 2009), we used a data-driven approach where group-level statistical analysis was performed to all time windows from pre-S1 baseline to the onset of S2, 198 parcels covering the complete cortical surface, and frequency bands from 3 to 120 Hz. This approach hence not only tests the primary hypothesis, but intrinsically also allows to weight that hypothesis's outcome against many alternative hypotheses. To illustrate the globally most prominent effects, these data were first summarized into TFRs showing the fraction of parcels out of all 198 parcels that show either a significant positive (P+, or where applicable, a negative, P−) test effect for oscillation amplitudes in each time and frequency bin after FDR correction for multiple comparisons. Hence, this “fraction-of-significant” visualization reveals the most salient phenomena in the dataset irrespective of whether they support the original hypothesis. As first data-driven evidence for our primary hypothesis, the ANOVA for features (Fig. 1C) revealed that beta- (14–30 Hz) and low-gamma (30–60 Hz) oscillation amplitudes were sensitive to features maintained in VWM, whereas significant effects were not observed in theta (θ, 3–7 Hz) or alpha (α, 8–14 Hz) frequency bands. These ANOVA data hence indicated that, in up to 20% of cortical parcels, the amplitudes of beta and gamma oscillations differentiated one or any of the memorized features.

To visualize the cortical loci of these effects and assess whether they originate from the visual or alternatively some other brain systems, we selected 2 salient TF selections of interest from the TFRs (Fig. 1C) for retention interval beta and low-gamma effects, and then estimated for each parcel the fraction of all TF elements in the selection showing a significant positive (P+, or later a negative P−) effect. To facilitate functional inferences, we overlaid the cortical topographies of M/EEG observations with fMRI-based boundaries of visual cortical subdivisions and functional systems. We first divided the visual cortex into early visual (V1–V3), LO, V4 and V8, as well as ventral and dorsal parts (Tootell and Hadjikhani 2001; Hansen et al. 2007; Kravitz et al. 2013); of which, the ventral part underlies object and the dorsal part location processing (Riesenhuber and Poggio 2002; Grill-Spector and Malach 2004; Kravitz et al. 2013), and V4, V8 color processing (Tootell and Hadjikhani 2001; Kravitz et al. 2013; Fig. 2C). To delineate also the attentional and executive control systems that putatively comprise the FP, VA, and DA resting-state networks, we used intrinsic functional connectivity analyses from the 1000-subject cohort (Yeo et al. 2011; Fig. 2C). In the beta- and low-gamma-TF ROIs of the feature ANOVA data, we observed feature effects in many visual regions including the ventral, dorsal, and early visual regions, and also in V4, V8 areas in the beta band (Fig. 1D). However, both low-gamma- and beta-band amplitudes were feature selective also in frontal areas of the FP and VA networks. The ANOVA feature effect hence suggests that both visual and frontal gamma oscillations reflect the maintenance of feature-specific information. Because these ANOVA findings did not reveal which feature-feature amplitude relationships contributed to the observed effects, we next set out to identify the time–frequency–anatomy patterns that were specific to the memorized features.

VWM Load Effects Are Spectrally, Temporally, and Anatomically Unique to Memorized Features

To further address the role of ongoing oscillations in feature-specific VWM, we separately investigated amplitude modulations in the shape, color, and location VWM tasks. We first inspected whether oscillation amplitudes were dependent on memory load, that is, on the amount of feature information maintained in VWM. These data revealed that greater oscillation amplitudes for 4- than for 2-object memory load in the low- and high-gamma (80–120 Hz) bands throughout the late retention (1.2–2 s) period for shape VWM, and only in the early retention (0.5–1.2 s) for color VWM (Fig. 2A). In color VWM, beta-band amplitudes were also load-dependently enhanced throughout the retention period (0.5–2 s). Intriguingly, we did not observe essentially any load-dependent increases in the oscillation amplitudes for location VWM. Hence, the analyses of each condition separately revealed clearly distinct TF patterns of oscillation amplitude load effects for shape, color, and location VWM. Furthermore, these data showed that, in contrast with beta and gamma, alpha amplitudes were load-dependently suppressed in all tasks (cf. Fig. 2A). Overall, these narrow-band load effects were concurrent with broad-band suppression below baseline levels of oscillation amplitudes in all frequency bands between 3 and 40 Hz in all tasks (Supplementary Fig. 2).

To reveal the sources of these load effects, we chose TF selections for the most robust effects and plotted the fractions of significant TF elements in each cortical parcel. In shape VWM, the gamma load effects originated from ventral visual areas, V1–V3 cortices, and FP, DA, and VA networks (Fig. 2B). In color VWM, early load-dependent strengthening of high-beta to low-gamma amplitudes was localized primarily to early visual areas, but also to the putative color-processing areas V4, V8, as well as to the ventral visual areas. Interestingly, in the location VWM task, the apparently minor load-dependent increase in beta-band oscillation amplitudes was observed specifically in bilateral early visual areas. These results hence show that, in parallel with distinct TF patterns, the load effects in beta- and gamma-band amplitudes originated from also partially distinct cortical areas. To further corroborate and illustrate the feature-specific variability of beta and gamma oscillations in the visual cortex, we averaged the amplitudes separately for parcels with significant modulation for each cortical ROI. These visualizations showed that the mean amplitudes had a similar frequency profile throughout the visual cortex ROIs, and hence that the low- and high-gamma- band components in Figure 2 were produced by large-scale broad-band activity rather than by superposition of anatomically distributed narrow-band phenomena (Supplementary Fig. 3). Taken together, the results both from the analysis of variance for feature effects and from the assessment of oscillation memory load-dependence are in line with the behavioral data in indicating that task demands for maintaining feature information in VWM are met by distinct neuronal mechanisms that specifically involve beta- and gamma-band oscillations.

Regional ROI Analysis Corroborates Data-Driven Load Effect Observations

To corroborate the feature-specific differences in the cortical localization of beta- and gamma-band load effects among shape, color, and location conditions, we obtained mean amplitude load effects for the 5 visual regions supporting processing of distinct visual features (Fig. 2C; early, V4 and V8, LO, ventral, and dorsal). This analysis confirmed the earlier results of high-gamma amplitudes being largest for shape VWM in ventral stream areas during both early and late retention intervals (Fig. 3). Furthermore, high-gamma amplitudes in the ventral ROI were also larger for shape VWM than for location VWM. For color VWM, low-gamma band amplitudes during the early retention interval were largest in the early visual areas when compared with other areas or with shape or location VWM tasks. Beta-band amplitudes, on the other hand, were similar in all visual areas, whereas in the dorsal stream, the amplitudes were suppressed and significantly smaller than in other visual areas. For spatial VWM, beta-band amplitudes were largest in early visual areas where they were larger than in ventral or dorsal ROIs. Thus, the amplitude load effects for the high-gamma band for shape VWM and the low-gamma band for color VWM were distinctive in those functional subdivisions of visual cortex corresponding to the shape and color processing. Importantly, these results corroborated the earlier TF analyses and showed that load-dependent low-gamma and high-gamma- band amplitudes were distinctive to memorized features and were also modulated in feature-specific visual cortices. In summary, gamma oscillation amplitude effects took place in feature-specific visual cortices differentially for shape, color, and location VWM, and these oscillations might hence reflect the contents of VWM.

Load-Dependence of Beta and Gamma Oscillations Predict Behavioral Performance

To next address the possible behavioral relevance of retention period gamma amplitude modulations and thereby their mechanistic contribution to VWM maintenance, we estimated the correlation between the load effects on amplitude modulations and HRs across subjects. The TFR of the fraction of parcels with a significant correlation shows that the individual gamma amplitude load effect throughout shape VWM retention interval (Fig. 4A) and in the early color VWM retention (Fig. 4B) was predictively correlated with the individual decline in HR from 2- to 4-object VWM load. Hence, the stronger the load-dependent amplitude enhancement was, the less HR decreased with increasing memory load.

Crucially, the correlation between the load-dependent amplitude increase and decline in HR for shape VWM was most pronounced in ventral stream object-representation areas, albeit observable also in color and dorsal stream regions (Fig. 4C). For color VWM, this correlation was significant in the left hemispheric ventral stream and color areas (Fig. 4D). Interestingly, for shape VWM, load-dependence of gamma oscillations also in FP and VA networks were predictive of the load-dependent decline in behavioral accuracy (Fig. 4C). Hence, this analysis showed that especially gamma-band oscillations in both feature-specific cortical regions and attentional networks were behaviorally relevant for shape and color VWM and might thus reflect the maintenance of multiple behaviorally relevant features in VWM.

Gamma Oscillations Specifically Enhanced for Shape VWM

The results this far showed that specifically the strength of gamma oscillations was correlated with feature-specificity of VWM as well as that sustained and load-dependently strengthened gamma oscillations, albeit transiently present during color VWM, were strongest during shape VWM. However, as shape VWM was more difficult than the color or location tasks, the gamma-band load effects could be related to task difficulty. To address this possibility, we compared the 2-object shape with 4-object color VWM that were equal in difficulty. A TFR of the fraction of parcels with significant differences between the conditions (t-test, P < 0.05, FDR-corrected) revealed that oscillation amplitudes in high-beta and -gamma bands were greater during shape than color VWM retention even with balanced task difficulty (Fig. 5A). This supports the notion that sustained gamma oscillations reflect the processing demands that are pronounced in the shape VWM, although being present also for color VWM. Importantly, this gamma effect was seen in cortical areas that underlie the processing object information (Riesenhuber and Poggio 2002; Grill-Spector and Malach 2004; Kravitz et al. 2013), that is, in the ventral visual cortex, LO, and in the FP network (Fig. 5B). To ask whether the apparently broad-band gamma oscillations in the TFR (Fig. 5A) would reflect the superposition of narrow-band oscillations in different brain areas, we again averaged the amplitudes across significant parcels separately for each visual ROI. Unlike in the case of the widespread broad-band load effect (Supplementary Fig. 3), the high-gamma oscillations between 80 and 120 Hz were more characteristic to V1–V3, V4, and V8, as well as LO, whereas a lower 40- to 80-Hz gamma effect was observed in later ventral and dorsal stream regions (Supplementary Fig. 4).

Figure 5.

At balanced task difficulty, shape VWM shows stronger gamma oscillations than color VWM. (A) TFR for the difference between 2-object load shape condition and 4-object load color condition (t-test, P < 0.05, FDR-corrected). Red-yellow and blue colors indicate for each TF element the fraction of parcels where the amplitude in shape VWM was greater (P Sh > Co) or smaller (P Sh < Co), respectively, than in color VWM. (B) The difference in low gamma (lγ; here 30–45 Hz) and high gamma (hγ; here 72–120 Hz) amplitudes was salient in ventral stream regions and also in the FP network of the left hemisphere (colors and axes as in Fig. 2).

Figure 5.

At balanced task difficulty, shape VWM shows stronger gamma oscillations than color VWM. (A) TFR for the difference between 2-object load shape condition and 4-object load color condition (t-test, P < 0.05, FDR-corrected). Red-yellow and blue colors indicate for each TF element the fraction of parcels where the amplitude in shape VWM was greater (P Sh > Co) or smaller (P Sh < Co), respectively, than in color VWM. (B) The difference in low gamma (lγ; here 30–45 Hz) and high gamma (hγ; here 72–120 Hz) amplitudes was salient in ventral stream regions and also in the FP network of the left hemisphere (colors and axes as in Fig. 2).

Gamma Oscillations Support Feature Binding for Integrated Object Representations

Taken together, the first experiment showed that retention period gamma-band oscillations support feature-specific VWM maintenance in the visual cortex. Intriguingly, even with matched difficulty, these gamma oscillations were more pronounced for VWM of the binding-demanding shape features than for colors and locations that are likely to have simple feed-forward representations (Singer 1995; Herrmann et al. 2004). These observations would thus be in line with the hypothesized role for gamma oscillations in support of the integration of distributed feature processing into coherent memory representations (Singer 1999; Tallon-Baudry and Bertrand 1999; Jensen et al. 2007). To test this idea explicitly, we recorded M/EEG with a conjunction VWM task where the subjects' memorized conjunction of color and location information with the stimuli identical to those used Experiment 1. We hypothesized that if gamma oscillations contributed to feature binding and the maintenance of these integrated representations in VWM, color–location conjunctions should be associated with enhanced gamma-band activity in both dorsal and ventral visual areas in comparison with the memorization of these features separately.

The behavioral performance in the conjunction task was similar to that in the color and location tasks performed separately (Fig. 1B), and hence memorizing feature conjunctions did not affect task difficulty as confirmed by previous findings in a distinct paradigm (Luck and Vogel 1997). TFRs of significant load effects showed, as hypothesized, that color–location conjunction VWM was associated with load-dependent strengthening of both beta- and low-gamma- band oscillations throughout the retention period (Fig. 6A). In addition, the oscillation amplitudes during conjunction VWM were greater than the mean of amplitudes in the separate color and location tasks both in low- and high-gamma bands (Fig. 6B). Importantly, the load as well as the conjunction effect in gamma-band amplitudes was observed in both ventral and dorsal visual areas (Fig. 6C,D), supporting the idea that retention period gamma oscillations indeed support the integration of information distributed to these processing streams in the color–location conjunction VWM. Taken together, these data were in line with the hypothesis that gamma oscillations support the integration of the conjunction of feature-specific information and support the integration of distributed feature processing into coherent memory representations.

Figure 6.

VWM of color–location conjunctions, unlike VWM of these features separately, is associated with load-dependent and sustained gamma (γ) oscillations. (A) TFR of significant load effects for conjunction VWM (t-test for the difference of baseline-corrected 4- and 2-object conditions, P < 0.05, FDR-corrected, cf. Fig. 1C). (B) Gamma-band amplitudes are significantly stronger during the VWM retention interval for the color–location conjunctions than for these features separately (t-test for the difference of the conjunction and the average of color and location conditions, P < 0.05, FDR-corrected, cf. Fig. 1C). (C) Low-gamma (lγ)- and beta (β)-band amplitudes were positively correlated with memory load in both color- and location-related visual cortical regions, as well as in the ventral stream areas. (D) Similarly, γ-oscillation (here 40–120 Hz) amplitudes in these regions were stronger in the conjunction task than in the color and location tasks on average. (E) Amplitude changes averaged in visual cortex subdivisions (dark red, conjunction load effect and purple, conjunction minus mean of color and location) for hγ (80–120 Hz), lγ (40–72 Hz), and β-bands (17–35 Hz) averaged over the whole retention period (0.5–2 s). Dashed lines indicate uncorrected significance values across loads and conditions (t-test, P < 0.05).

Figure 6.

VWM of color–location conjunctions, unlike VWM of these features separately, is associated with load-dependent and sustained gamma (γ) oscillations. (A) TFR of significant load effects for conjunction VWM (t-test for the difference of baseline-corrected 4- and 2-object conditions, P < 0.05, FDR-corrected, cf. Fig. 1C). (B) Gamma-band amplitudes are significantly stronger during the VWM retention interval for the color–location conjunctions than for these features separately (t-test for the difference of the conjunction and the average of color and location conditions, P < 0.05, FDR-corrected, cf. Fig. 1C). (C) Low-gamma (lγ)- and beta (β)-band amplitudes were positively correlated with memory load in both color- and location-related visual cortical regions, as well as in the ventral stream areas. (D) Similarly, γ-oscillation (here 40–120 Hz) amplitudes in these regions were stronger in the conjunction task than in the color and location tasks on average. (E) Amplitude changes averaged in visual cortex subdivisions (dark red, conjunction load effect and purple, conjunction minus mean of color and location) for hγ (80–120 Hz), lγ (40–72 Hz), and β-bands (17–35 Hz) averaged over the whole retention period (0.5–2 s). Dashed lines indicate uncorrected significance values across loads and conditions (t-test, P < 0.05).

To further test whether mean amplitudes would differ across cortical regions, we averaged amplitudes separately over all visual areas and used regional ROI analysis to test amplitude differences across regions. In ventral and early visual areas, low-gamma amplitudes were larger than in LO for load-dependent effects (t-test, P < 0.05, Fig. 6E). This analysis did not reveal any other systematic effects in the conjunction-specific amplitudes across cortical regions, which was expected as VWM of conjunction of color and location information should involve both dorsal and ventral stream visual regions. All in all, the memorization for the conjunction of color and location information induced sustained retention period gamma oscillations that were stronger for the VWM of feature conjunctions than for the VWM of these features separately.

Alpha Oscillations Are Similarly Enhanced in all Tasks Across the Visual Hierarchy

Taken together, our data show that specifically gamma oscillation amplitude modulations are correlated with the maintenance of feature-specific information in VWM, especially when binding of anatomically distributed processing is demanded. In contrast to gamma oscillations, alpha oscillation amplitude modulations have been suggested to underlie the attentional and executive functions of VWM (Sauseng et al. 2009; Palva et al. 2011) as well as to protect VWM against distracting information (Bonnefond and Jensen 2012; Roux et al. 2012; Park et al. 2014).

To investigate the role of alpha oscillations in the maintenance of specific features in the present task, we investigated whether the load-dependent decrease as well as slight increase observed for alpha oscillations would be localized into feature-specific or feature-irrelevant cortical areas. To that end, we selected an alpha-band TF selection from the Figure 2A and estimated cortical localization of this effect. For shape VWM, load-dependence of alpha amplitudes was observed in early visual and V4, V8 areas and for color and location VWM in early visual areas (Fig. 7A). To test whether these differences across different visual ROIs were significant, we used regional ROI analysis for the mean load-dependence of amplitudes across ROIs. This analysis revealed that, during the late retention interval, load-dependent mean alpha amplitudes were indeed higher in early visual and LO regions than dorsal or ventral regions for shape, color, and location VWM (Fig. 7B). Importantly, we did not observe increased alpha amplitudes in dorsal and hence task-irrelevant areas either for shape or color VWM. For this reason, these data showed that the load-dependence of alpha amplitudes did not take place in the feature-specific nor in the feature-irrelevant regions, but rather always in early visual areas. As gamma oscillations have been suggested to signal active neuronal processing, we also investigated the functional significance of alpha oscillations by testing whether the alpha and gamma amplitudes would be correlated across visual regions. However, we did not observe any significant correlations between alpha and gamma oscillation amplitudes (t-test, P > 0.2), indicating that the amplitude of alpha oscillations is not dependent on the amplitude of gamma oscillations in the present task.

Figure 7.

Alpha (α)-band oscillations are load-dependently suppressed in widespread cortical regions, but strengthened in early visual cortex. (A) Cortical localization of the α- (8–13 Hz) amplitude load effect during the early (0.5–1.2 s) and late (1.2–2 s) parts of the retention interval for the TF ROIs indicated in Figure 2A (for the shape, color, and location conditions) and Figure 6B (conjunction condition). (B) Normalized mean amplitudes (±SEM) for different comparisons and ROIs. During the late retention period, mean α amplitudes in all conditions were stronger in V1–V3 and LO than in visual regions higher in cortical hierarchy. In dorsal ROI, α amplitudes were more suppressed for the color than for location VWM and in dorsal ROI more for the location than shape VWM.

Figure 7.

Alpha (α)-band oscillations are load-dependently suppressed in widespread cortical regions, but strengthened in early visual cortex. (A) Cortical localization of the α- (8–13 Hz) amplitude load effect during the early (0.5–1.2 s) and late (1.2–2 s) parts of the retention interval for the TF ROIs indicated in Figure 2A (for the shape, color, and location conditions) and Figure 6B (conjunction condition). (B) Normalized mean amplitudes (±SEM) for different comparisons and ROIs. During the late retention period, mean α amplitudes in all conditions were stronger in V1–V3 and LO than in visual regions higher in cortical hierarchy. In dorsal ROI, α amplitudes were more suppressed for the color than for location VWM and in dorsal ROI more for the location than shape VWM.

Discussion

We used source-reconstructed M/EEG and state-of-the-art data analyses to investigate the significance of cortical oscillations in the maintenance of visual feature information in VWM. Specifically, we hypothesized that beta- and gamma-band oscillations could support the maintenance of feature-specific information in visual regions underlying the neuronal processing of the memorized features. We found dominant load-dependent and behaviorally significant beta and gamma oscillations to be a characteristic property of VWM maintenance at the whole cortex level. The feature-specific strengths, temporal patterns, and sources of gamma oscillations in the visual cortex together suggest that gamma oscillations could indeed directly reflect the maintenance of visual representations and hence reflect the contents of VWM. In addition, the prevalence of gamma oscillations for shape and color–location conjunction VWM in comparison with color or location VWM separately implies that local gamma synchronization reflects integration of features into object-based memory representations.

Gamma Amplitudes Reflect the Maintenance of Feature-Specific Information

Although several sensor-level EEG (Tallon-Baudry et al. 1998; Deiber et al. 2007; Gruber et al. 2008; Sauseng et al. 2009) and also MEG studies using source reconstruction (Jokisch and Jensen 2007; Medendorp et al. 2007; Palva et al. 2011; Roux et al. 2012) have reported enhanced beta and gamma amplitudies during VWM retention period as well as evidence for the behavioral significance of such enhancements (Sauseng et al. 2009; Palva et al. 2011; Roux et al. 2012), it has remained unclear whether these oscillations subserve the binding and representation of feature-specific sensory information maintained in VWM or the attentional and executive processes coordinating and driving this maintenance.

After testing all cortical parcels and all frequency bands from 3 to 120 Hz for VWM load effects, we observed that memory load correlated positively with oscillation amplitudes to a large extent in beta and especially in gamma frequency bands. Albeit present in widespread cortical structures, the load-dependent modulations in beta and gamma amplitudes were prominent in the visual cortical regions. The analysis based on parcel resolution suggested that the load effects in the visual cortex would further be localized to regions specific to the processing of the memorized feature. This notion was corroborated and extended by the ROI-based analysis that showed that the beta- and gamma-load effects are strongest in visual cortical areas underlying the processing of the memorized feature. Load-dependent modulations of shape VWM were largest in the ventral visual stream that underlies object processing (Riesenhuber and Poggio 2002; Grill-Spector and Malach 2004; Kravitz et al. 2013), whereas for the color VWM, gamma-band load effects were strongest in the early visual areas, and beta-band effects in ventral and V4, V8 areas that participate to the processing of color information (Tootell and Hadjikhani 2001; Hansen et al. 2007; Kravitz et al. 2013). This distinction in the localization to early and late cortical processing areas for color and shape VWM, respectively, might reflect the transition of neuronal representations from primary sensory to a more abstract level and the generalization of the representations during the VWM retention. This notion would be supported also by the early and late, respectively, retention interval gamma activations. Importantly, although the amplitudes were in general load-dependently suppressed for the location VWM, beta amplitudes were nevertheless increased in the early visual areas, suggesting that activity in this small area could be sufficient to represent the location information. Taken together, these data suggest that both beta and gamma oscillations reflect the maintenance of feature-specific information in VWM. These results hence extend prior studies that have observed feature-specific effects in gamma amplitudes in the human parietal cortex (Morgan et al. 2011) as well as beta-band synchronization between monkey frontal and parietal cortices (Salazar et al. 2012), and hence in regions underlying attentional and executive control (Fuster 2000; Kastner and Ungerleider 2000; Corbetta and Shulman 2002). Furthermore, as our data-driven analysis approach did not reveal other oscillatory correlates of feature-specific VWM, local beta and gamma oscillations are among the key system-level mechanisms to support feature-specific VWM.

A mechanistic role for gamma oscillations in VWM maintenance was supported by the observation that the load-dependent increase in gamma amplitudes predicts a load-dependent decline in behavioral accuracy for shape and color VWM. Load-dependent decline in behavioral accuracy was most strongly predicted by gamma amplitudes in the ventral stream, but observed also in color and dorsal stream regions. Indeed, shape processing is known to engage also the dorsal stream areas, while the functional roles of V4, V8 areas are complex and extend beyond color processing (Tootell and Hadjikhani 2001; Kravitz et al. 2013). On the other hand, gamma- band oscillations predicted performance in the color VWM only weakly, but this might have been caused by only a minor load-dependent decline in the HR decreasing hence the power of the analysis.

Distinct Temporal Profiles and Patterns of Gamma Oscillations Reflect Feature-Binding Demands

Shape, color, and location VWM were associated with different temporal and spectral profiles of load-dependent oscillations. The perception of simple features such as spatial locations and colors is likely to be accomplished by feed-forward processing and automatic feature representations, but the precise neuronal representations of novel shapes and feature conjunctions necessarily require the integration of anatomically distributed feature information (Singer 1995; Fries et al. 2002; Palva et al. 2002; Herrmann et al. 2004). The earlier occurrence of load-dependent gamma oscillations in color VWM suggests that color memory is indeed faster than the processing of shapes for which the load-dependent oscillations occurred during the late retention period. These salient differences in the frequency profiles of the retention interval show together with weak correlations of the individual behavioral performance across tasks that these partially different neuronal mechanisms may indeed support the maintenance of these features.

Our observations of stronger gamma oscillations with more complex features are also in line with the “temporal correlation” hypothesis that posits gamma-band synchronization to signal the perceptual relatedness of distributed cell assemblies and therefore support feature integration (Singer and Gray 1995; Singer 1999). Most importantly, this hypothesis was supported by data showing stronger gamma oscillations for the VWM of conjunction of spatial and color information, than for the VWM of these features separately. Importantly, enhanced gamma amplitudes for maintaining color–location conjunctions in VWM were localized both to the ventral and dorsal stream visual regions that are associated with the processing of object–color and location information, respectively (Tootell and Hadjikhani 2001; Grill-Spector and Malach 2004; Hansen et al. 2007; Kravitz et al. 2013). Thus, these results provide strong support for the idea that gamma oscillations underlie the integration of feature-specific information into integrated object representations maintained in memory.

The Roles of Frontal Gamma oscillations

In addition to the feature-specific visual cortical regions, we also observed strengthened gamma oscillations in FP and VA networks even under conditions with balanced task difficulty. These results indicate that the amplitude modulations in these regions do not merely reflect increased executive and attentional demands by increased VWM load. Importantly, FP network has been shown to underlie top-down coordination of the occipito-temporal activity during VWM maintenance (Herrmann et al. 2004; Zanto et al. 2011; Gazzaley and Nobre 2012), whereas VA network responds to behaviorally relevant objects (Corbetta et al. 2008). As object features are suggested to be bound to coherent object representations only when attended (Treisman and Gelade 1980; Reynolds and Desimone 1999; Wheeler and Treisman 2002; Treisman and Zhang 2006), we suggest that gamma oscillations in FP and VA networks might underlie the active effort required to bind the task-relevant object features into coherent VWM representations. Furthermore, as the shape VWM task was slightly more difficult compared with the other tasks, load-dependence of gamma oscillations for shape VWM in FP and VA networks might partially also reflect enhanced demands for attentional and executive functions.

Functional Role of Alpha Oscillations in Feature-Specific VWM

In addition to gamma oscillations, prior studies have also observed a correlation between alpha amplitudes and behavioral performance in visual (Palva et al. 2011) and somatosensory (Haegens et al. 2010) working memory. We observed similar load-dependent increases in alpha amplitudes in the early and LO visual regions for all conditions as well as widespread load-dependent large-scale suppression that was more pronounced for color and location than for shape and conjunction VWM.

The strength of posterior alpha oscillation amplitudes has been associated with the suppression of distracting information (Sauseng et al. 2009; Bonnefond and Jensen 2012; Roux et al. 2012). According to this hypothesis, alpha amplitudes should be stronger in visual cortical regions underlying the processing of irrelevant memory features, whereas gamma oscillations should be observed in the task-relevant neuronal circuitry. However, in our analyses examining this possibility, we observed that alpha amplitudes were less load-dependently suppressed in the visual cortical areas for the memorized feature and more suppressed for the irrelevant features. Furthermore, we did not observe either positive or negative correlations between the load-dependent strengths of alpha and gamma oscillations, which is in line with a recent study showing that cell-specific alpha- and gamma-band synchronizations in monkey V4 are not mutually exclusive (Vinck et al. 2013). Complementing the lack of feature-specific load effects, we did not observe alpha-band effects in direct comparisons of the 2-object conditions in an ANOVA (Fig. 1). While the late retention emergence of positive alpha-band load effects in early visual areas may reflect memory protection through blanket inhibition of new visual inputs, these results do not support the hypothesis that posterior alpha oscillations would suppress distracting information in task-irrelevant cortical regions nor the hypothesis that alpha oscillations are predictive of VWM contents per se (Anderson et al. 2014).

Several prior studies have proposed that alpha oscillations could contribute to the coordination of attentional and central executive functions of VWM especially at supracapacity memory loads (Palva et al. 2011; Klimesch 2012; Roux et al. 2012). However, as our experimental paradigm was designed to have as identical attentional and executive demands between conditions as possible, it does not yield insights on the role of alpha oscillations in these functions. Furthermore, in our task in which performance was largely within the individual VWM capacity limitations, we observed only a minor load-dependent alpha amplitude increase supporting the idea that alpha oscillations especially in the FP network are indeed a compensatory mechanism for VWM loads exceeding individual capacities (Palva et al. 2011). Taken together, these data hence suggest that alpha and gamma oscillations reflect independent mechanistic aspects of active VWM maintenance.

Conclusions

Our results strongly suggest that retention period gamma oscillations reflect the maintenance of visual feature-specific information in VWM and support the integration of features into integrated memory representation (Treisman and Gelade 1980; Luck and Vogel 1997; Treisman 2006). Local gamma oscillations may thereby underlie the maintenance of VWM contents.

Supplementary Material

Supplementary material can be found at: http://www.cercor.oxfordjournals.org/.

Funding

This work was supported by University of Helsinki Research Grants (grant nos 788/51/2010, 195/111.00/2011); Academy of Finland (grant nos SA 266402, SA 267030, and SA 253130), and the Sigrid Juselius Foundation. The authors declare no competing Financial interests.

Notes

We thank Tom Campbell for comments on an earlier version of this manuscript. Conflict of Interest: None declared.

References

Anderson
DE
Serences
JT
Vogel
EK
Awh
E
.
2014
.
Induced alpha rhythms track the content and quality of visual working memory representations with high temporal precision
.
J Neurosci
 .
34
:
7587
7599
.
Axmacher
N
Schmitz
DP
Wagner
T
Elger
CE
Fell
J
.
2008
.
Interactions between medial temporal lobe, prefrontal cortex, and inferior temporal regions during visual working memory: a combined intracranial EEG and functional magnetic resonance imaging study
.
J Neurosci
 .
28
:
7304
7312
.
Baddeley
A
.
2003
.
Working memory: looking back and looking forward
.
Nat Rev Neurosci
 .
4
:
829
839
.
Baddeley
A
.
1996
.
The fractionation of working memory
.
Proc Natl Acad Sci USA
 .
93
:
13468
13472
.
Bonnefond
M
Jensen
O
.
2012
.
Alpha oscillations serve to protect working memory maintenance against anticipated distracters
.
Curr Biol
 .
22
:
1969
1974
.
Bosman
CA
Schoffelen
JM
Brunet
N
Oostenveld
R
Bastos
AM
Womelsdorf
T
Rubehn
B
Stieglitz
T
De Weerd
P
Fries
P
.
2012
.
Attentional stimulus selection through selective synchronization between monkey visual areas
.
Neuron
 .
75
:
875
888
.
Buschman
TJ
Miller
EK
.
2007
.
Top-down versus bottom-up control of attention in the prefrontal and posterior parietal cortices
.
Science
 .
315
:
1860
1862
.
Corbetta
M
Patel
G
Shulman
GL
.
2008
.
The reorienting system of the human brain: from environment to theory of mind
.
Neuron
 .
58
:
306
324
.
Corbetta
M
Shulman
GL
.
2002
.
Control of goal-directed and stimulus-driven attention in the brain
.
Nat Rev Neurosci
 .
3
:
201
215
.
Dale
AM
Fischl
B
Sereno
MI
.
1999
.
Cortical surface-based analysis. I. Segmentation and surface reconstruction
.
Neuroimage
 .
9
:
179
194
.
Deiber
MP
Missonnier
P
Bertrand
O
Gold
G
Fazio-Costa
L
Ibanez
V
Giannakopoulos
P
.
2007
.
Distinction between perceptual and attentional processing in working memory tasks: a study of phase-locked and induced oscillatory brain dynamics
.
J Cogn Neurosci
 .
19
:
158
172
.
Fischl
B
Salat
DH
Busa
E
Albert
M
Dieterich
M
Haselgrove
C
van der Kouwe
A
Killiany
R
Kennedy
D
Klaveness
S
et al
2002
.
Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain
.
Neuron
 .
33
:
341
355
.
Fries
P
.
2005
.
A mechanism for cognitive dynamics: neuronal communication through neuronal coherence
.
Trends Cogn Sci
 .
9
:
474
480
.
Fries
P
Schroder
JH
Roelfsema
PR
Singer
W
Engel
AK
.
2002
.
Oscillatory neuronal synchronization in primary visual cortex as a correlate of stimulus selection
.
J Neurosci
 .
22
:
3739
3754
.
Fuster
JM
.
2000
.
Executive frontal functions
.
Exp Brain Res
 .
133
:
66
70
.
Gazzaley
A
Nobre
AC
.
2012
.
Top-down modulation: bridging selective attention and working memory
.
Trends Cogn Sci
 .
16
:
129
135
.
Gregoriou
GG
Gotts
SJ
Zhou
H
Desimone
R
.
2009
.
Long-range neural coupling through synchronization with attention
.
Prog Brain Res
 .
176
:
35
45
.
Grill-Spector
K
Malach
R
.
2004
.
The human visual cortex
.
Annu Rev Neurosci
 .
27
:
649
677
.
Gruber
T
Tsivilis
D
Giabbiconi
CM
Müller
MM
.
2008
.
Induced electroencephalogram oscillations during source memory: familiarity is reflected in the gamma band, recollection in the theta band
.
J Cogn Neurosci
 .
20
:
1043
1053
.
Haegens
S
Osipova
D
Oostenveld
R
Jensen
O
.
2010
.
Somatosensory working memory performance in humans depends on both engagement and disengagement of regions in a distributed network
.
Hum Brain Mapp
 .
31
:
26
35
.
Hämäläinen
MS
Ilmoniemi
RJ
.
1994
.
Interpreting magnetic fields of the brain: minimum norm estimates
.
Med Biol Eng Comput
 .
32
:
35
42
.
Hämäläinen
MS
Sarvas
J
.
1989
.
Realistic conductivity geometry model of the human head for interpretation of neuromagnetic data
.
IEEE Trans Biomed Eng
 .
36
:
165
171
.
Hansen
KA
Kay
KN
Gallant
JL
.
2007
.
Topographic organization in and near human visual area V4
.
J Neurosci
 .
27
:
11896
11911
.
Heck
A
Fastenrath
M
Ackermann
S
Auschra
B
Bickel
H
Coynel
D
Gschwind
L
Jessen
F
Kaduszkiewicz
H
Maier
W
et al
2014
.
Converging genetic and functional brain imaging evidence links neuronal excitability to working memory, psychiatric disease, and brain activity
.
Neuron
 .
81
:
1203
1213
.
Herrmann
CS
Munk
MH
Engel
AK
.
2004
.
Cognitive functions of gamma-band activity: memory match and utilization
.
Trends Cogn Sci
 .
8
:
347
355
.
Jensen
O
Kaiser
J
Lachaux
JP
.
2007
.
Human gamma-frequency oscillations associated with attention and memory
.
Trends Neurosci
 .
30
:
317
324
.
Jokisch
D
Jensen
O
.
2007
.
Modulation of gamma and alpha activity during a working memory task engaging the dorsal or ventral stream
.
J Neurosci
 .
27
:
3244
3251
.
Kaiser
J
Heidegger
T
Wibral
M
Altmann
CF
Lutzenberger
W
.
2008
.
Distinct gamma-band components reflect the short-term memory maintenance of different sound lateralization angles
.
Cereb Cortex
 .
18
:
2286
2295
.
Kaiser
J
Rahm
B
Lutzenberger
W
.
2009
.
Temporal dynamics of stimulus-specific gamma-band activity components during auditory short-term memory
.
Neuroimage
 .
44
:
257
264
.
Kastner
S
Ungerleider
LG
.
2000
.
Mechanisms of visual attention in the human cortex
.
Annu Rev Neurosci
 .
23
:
315
341
.
Klimesch
W
.
2012
.
Alpha-band oscillations, attention, and controlled access to stored information
.
Trends Cogn Sci
 .
16
:
606
617
.
Korhonen
O
Palva
S
Palva
JM
.
2014
.
Sparse weightings for collapsing inverse solutions to cortical parcellations optimize M/EEG source reconstruction accuracy
.
J Neurosci Methods
 .
226C
:
147
160
.
Kravitz
DJ
Saleem
KS
Baker
CI
Ungerleider
LG
Mishkin
M
.
2013
.
The ventral visual pathway: an expanded neural framework for the processing of object quality
.
Trends Cogn Sci
 .
17
:
26
49
.
Kriegeskorte
N
Simmons
WK
Bellgowan
PS
Baker
CI
.
2009
.
Circular analysis in systems neuroscience: the dangers of double dipping
.
Nat Neurosci
 .
12
:
535
540
.
Linden
DE
Bittner
RA
Muckli
L
Waltz
JA
Kriegeskorte
N
Goebel
R
Singer
W
Munk
MH
.
2003
.
Cortical capacity constraints for visual working memory: dissociation of fMRI load effects in a fronto-parietal network
.
Neuroimage
 .
20
:
1518
1530
.
Luck
SJ
Vogel
EK
.
1997
.
The capacity of visual working memory for features and conjunctions
.
Nature
 .
390
:
279
281
.
Magen
H
Emmanouil
TA
McMains
SA
Kastner
S
Treisman
A
.
2009
.
Attentional demands predict short-term memory load response in posterior parietal cortex
.
Neuropsychologia
 .
47
:
1790
1798
.
Medendorp
WP
Kramer
GF
Jensen
O
Oostenveld
R
Schoffelen
JM
Fries
P
.
2007
.
Oscillatory activity in human parietal and occipital cortex shows hemispheric lateralization and memory effects in a delayed double-step saccade task
.
Cereb Cortex
 .
17
:
2364
2374
.
Mohr
HM
Goebel
R
Linden
DE
.
2006
.
Content- and task-specific dissociations of frontal activity during maintenance and manipulation in visual working memory
.
J Neurosci
 .
26
:
4465
4471
.
Morgan
HM
Muthukumaraswamy
SD
Hibbs
CS
Shapiro
KL
Bracewell
RM
Singh
KD
Linden
DE
.
2011
.
Feature integration in visual working memory: parietal gamma activity is related to cognitive coordination
.
J Neurophysiol
 .
106
:
3185
3194
.
Munk
MH
Linden
DE
Muckli
L
Lanfermann
H
Zanella
FE
Singer
W
Goebel
R
.
2002
.
Distributed cortical systems in visual short-term memory revealed by event-related functional magnetic resonance imaging
.
Cereb Cortex
 .
12
:
866
876
.
Palva
JM
Monto
S
Kulashekhar
S
Palva
S
.
2010
.
Neuronal synchrony reveals working memory networks and predicts individual memory capacity
.
Proc Natl Acad Sci USA
 .
107
:
7580
7585
.
Palva
S
Kulashekhar
S
Hämäläinen
M
Palva
JM
.
2011
.
Localization of cortical phase and amplitude dynamics during visual working memory encoding and retention
.
J Neurosci
 .
31
:
5013
5025
.
Palva
S
Linkenkaer-Hansen
K
Näätänen
R
Palva
JM
.
2005
.
Early neural correlates of conscious somatosensory perception
.
J Neurosci
 .
25
:
5248
5258
.
Palva
S
Palva
JM
Shtyrov
Y
Kujala
T
Ilmoniemi
RJ
Kaila
K
Näätänen
R
.
2002
.
Distinct gamma-band evoked responses to speech and non-speech sounds in humans
.
J Neurosci
 .
22
:
RC211
.
Park
H
Lee
DS
Kang
E
Kang
H
Hahm
J
Kim
JS
Chung
CK
Jensen
O
.
2014
.
Blocking of irrelevant memories by posterior alpha activity boosts memory encoding
.
Hum Brain Mapp
 .
35
:
3972
3987
.
Pessoa
L
Gutierrez
E
Bandettini
P
Ungerleider
L
.
2002
.
Neural correlates of visual working memory: fMRI amplitude predicts task performance
.
Neuron
 .
35
:
975
987
.
Prabhakaran
V
Narayanan
K
Zhao
Z
Gabrieli
JD
.
2000
.
Integration of diverse information in working memory within the frontal lobe
.
Nat Neurosci
 .
3
:
85
90
.
Reynolds
JH
Desimone
R
.
1999
.
The role of neural mechanisms of attention in solving the binding problem
.
Neuron
 .
24
:
19
29
.
111–25
.
Riesenhuber
M
Poggio
T
.
2002
.
Neural mechanisms of object recognition
.
Curr Opin Neurobiol
 .
12
:
162
168
.
Rouhinen
S
Panula
J
Palva
J
Palva
S
.
2013
.
Load dependence of beta and gamma oscillations predicts individual capacity during attentive tracking
.
J Neurosci
 .
27
:
19023
19033
.
Roux
F
Wibral
M
Mohr
HM
Singer
W
Uhlhaas
PJ
.
2012
.
Gamma-band activity in human prefrontal cortex codes for the number of relevant items maintained in working memory
.
J Neurosci
 .
32
:
12411
12420
.
Rowe
JB
Toni
I
Josephs
O
Frackowiak
RS
Passingham
RE
.
2000
.
The prefrontal cortex: response selection or maintenance within working memory?
Science
 .
288
:
1656
1660
.
Sakai
K
Rowe
JB
Passingham
RE
.
2002
.
Active maintenance in prefrontal area 46 creates distractor-resistant memory
.
Nat Neurosci
 .
5
:
479
484
.
Salazar
RF
Dotson
NM
Bressler
SL
Gray
CM
.
2012
.
Content-specific fronto-parietal synchronization during visual working memory
.
Science
 .
338
:
1097
1100
.
Sauseng
P
Klimesch
W
Heise
KF
Gruber
WR
Holz
E
Karim
AA
Glennon
M
Gerloff
C
Birbaumer
N
Hummel
FC
.
2009
.
Brain oscillatory substrates of visual short-term memory capacity
.
Curr Biol
 .
19
:
1846
1852
.
Singer
W
.
2013
.
Cortical dynamics revisited
.
Trends Cogn Sci
 .
17
:
616
626
.
Singer
W
.
1995
.
Development and plasticity of cortical processing architectures
.
Science
 .
270
:
758
764
.
Singer
W
.
1999
.
Neuronal synchrony: a versatile code for the definition of relations?
Neuron
 .
24
:
49
65
.
111–25
.
Singer
W
Gray
CM
.
1995
.
Visual feature integration and the temporal correlation hypothesis
.
Annu Rev Neurosci
 .
18
:
555
586
.
Tallon-Baudry
C
Bertrand
O
.
1999
.
Oscillatory gamma activity in humans and its role in object representation
.
Trends Cogn Sci
 .
3
:
151
162
.
Tallon-Baudry
C
Bertrand
O
Delpuech
C
Pernier
J
.
1996
.
Stimulus specificity of phase-locked and non-phase-locked 40 Hz visual responses in human
.
J Neurosci
 .
16
:
4240
4249
.
Tallon-Baudry
C
Bertrand
O
Peronnet
F
Pernier
J
.
1998
.
Induced gamma-band activity during the delay of a visual short-term memory task in humans
.
J Neurosci
 .
18
:
4244
4254
.
Tootell
RB
Hadjikhani
N
.
2001
.
Where is “dorsal V4” in human visual cortex? Retinotopic, topographic and functional evidence
.
Cereb Cortex
 .
11
:
298
311
.
Treisman
A
.
2006
.
How the deployment of attention determines what we see
.
Vis Cogn
 .
14
:
411
443
.
Treisman
A
Zhang
W
.
2006
.
Location and binding in visual working memory
.
Mem Cogn
 .
34
:
1704
1719
.
Treisman
AM
Gelade
G
.
1980
.
A feature-integration theory of attention
.
Cogn Psychol
 .
12
:
97
136
.
Vinck
M
Womelsdorf
T
Buffalo
EA
Desimone
R
Fries
P
.
2013
.
Attentional modulation of cell-class-specific gamma-band synchronization in awake monkey area v4
.
Neuron
 .
80
:
1077
1089
.
Wheeler
ME
Treisman
AM
.
2002
.
Binding in short-term visual memory
.
J Exp Psychol Gen
 .
131
:
48
64
.
Yeo
BT
Krienen
FM
Sepulcre
J
Sabuncu
MR
Lashkari
D
Hollinshead
M
Roffman
JL
Smoller
JW
Zollei
L
Polimeni
JR
et al
2011
.
The organization of the human cerebral cortex estimated by intrinsic functional connectivity
.
J Neurophysiol
 .
106
:
1125
1165
.
Zanto
TP
Rubens
MT
Thangavel
A
Gazzaley
A
.
2011
.
Causal role of the prefrontal cortex in top-down modulation of visual processing and working memory
.
Nat Neurosci
 .
14
:
656
661
.
Zhang
W
Luck
SJ
.
2008
.
Discrete fixed-resolution representations in visual working memory
.
Nature
 .
453
:
233
235
.