Abstract
Resting-state fluctuations are ubiquitous and widely studied phenomena of the human brain, yet we are largely in the dark regarding their function in human cognition. Here we examined the hypothesis that resting-state fluctuations underlie the generation of free and creative human behaviors. In our experiment, participants were asked to perform three voluntary verbal tasks: a verbal fluency task, a verbal creativity task, and a divergent thinking task, during functional magnetic resonance imaging scanning. Blood oxygenation level dependent (BOLD)-activity during these tasks was contrasted with a control- deterministic verbal task, in which the behavior was fully determined by external stimuli. Our results reveal that all voluntary verbal-generation responses displayed a gradual anticipatory buildup that preceded the deterministic control-related responses. Critically, the time–frequency dynamics of these anticipatory buildups were significantly correlated with resting-state fluctuations’ dynamics. These correlations were not a general BOLD-related or verbal-response related result, as they were not found during the externally determined verbal control condition. Furthermore, they were located in brain regions known to be involved in language production, specifically the left inferior frontal gyrus. These results suggest a common function of resting-state fluctuations as the neural mechanism underlying the generation of free and creative behaviors in the human cortex.
Introduction
Spontaneous, ultra-slow, activity fluctuations, also termed resting-state fluctuations, are among the most ubiquitous and widely studied cortical phenomena. These slow (typically <0.5 Hz) intrinsic fluctuations were found across the entire human cortex, using a range of methods, from functional magnetic resonance imaging (fMRI, Biswal et al. 1995; Nir et al. 2006; Fox and Raichle 2007) to single neuron and IEEG recordings (Nir et al. 2008). Resting-state fluctuations are of great interest, and have been the focus of a vast number of studies. This extensive research has demonstrated that these fluctuations are organized in functional networks, which in turn were linked to task-related activation patterns and individual traits, experiences, and pathologies (Smith et al. 2009; Harmelech and Malach 2013; Tavor et al. 2016; Grossman et al. 2019). However, despite this massive research effort, the functional role that the resting-state fluctuations may play in human cognition and behavior remains a mystery. Here, we would like to present data supporting the hypothesis that resting-state fluctuations constitute a common cortical mechanism underlying human free and creative behaviors.
Free or voluntary behaviors constitute a unique and important class of behaviors that are internally generated, and are not directly driven by an external stimuli or instructions (Brass and Haggard 2008; Haggard 2008). Note that this definition is purely operational, and is not directly related to the deep and ongoing philosophical debate concerning free will. This operational distinction between deterministic and free behaviors is illustrated, for example, in studies that contrast voluntary self-initiated versus cue-triggered or externally determined movements (Libet et al. 1983; Schurger et al. 2012). Another commonly used paradigm demonstrating this distinction is verbal fluency (VF), in which participants perform freely paced internal generation of verbal exemplars from a certain defined category (either semantic or phonemic, e.g., animals or words that start with the letter “a”). These blocks are compared with a deterministic control condition, for example, wherein participants repeat a specific given word (Schlosser et al. 1998; Crosson et al. 2001; Knecht et al. 2003; Birn et al. 2010).
It may be argued that such category or domain-restricted behaviors are not completely free (e.g., a specific semantic category such as tools in a VF task). However, the critical difference between determined and free behaviors lies in the lack of a direct casual chain leading from an external cue to the specific behavior performed. Haggard (2008) defined three main components of this aspect: the “what” element, that is, the instruction to repeat a specific word (“rest” for example) in the deterministic control condition, in contrast to freely recalling or selecting any specific tool from the entire category in a VF condition; the “when” element, that is, repeating the word in response to an external cue in the deterministic condition, as compared with an internally determined pace in free VF; and the “whether” element—the actual occurrence of a behavior directly stems from an external trigger in the control condition, versus free internally driven generation in VF. This break in the external-causal chain which typifies the free condition raises the question of mechanism. For example, what neuronal process may induce the participants to select a specific tool from the entire category, that is, to select a specific decision out of the numerous potential alternatives?
While such voluntary behaviors are of critical importance to the sense of spontaneity and freedom which are essential elements of human wellbeing, not less important is the realization that the phenomena of creative insights and original thinking all emerge, by definition, through a free and internally generated process, that can never be instructed deterministically from outside. Thus, insight into a neuronal mechanism underlying free and creative behaviors may bear a direct impact on our understanding of a fundamental cognitive ability underlying human progress in all scientific and cultural domains. The topic of the brain mechanisms linked to creativity has been of great interest, as well as controversy regarding how to probe this “impossible” quest of studying creativity inside the laboratory (Fink et al. 2007; Dietrich and Kanso 2010; Jung et al. 2013; Benedek, Jauk, et al. 2014b; Beaty et al. 2019). The brain networks involved in such behaviors, specifically in the commonly employed divergent thinking tasks, have been extensively mapped, demonstrating among other regions the consistent involvement of the lateral prefrontal cortex, the anterior cingulate cortex (ACC) and the default mode network (DMN) (Fink et al. 2009; Abraham et al. 2012; Gonen-Yaacovi et al. 2013; Mayseless et al. 2015; Wu et al. 2015; Beaty et al. 2016). A few studies have examined more directly the brain activity preceding a creative idea-generation or problem solving “a-ha” moment, revealing different activation patterns during and prior to riddle solving by an “insight” experience as compared with noninsight (Jung-Beeman et al. 2004; Kounios et al. 2006).
However, the specific neural mechanisms and dynamics driving these moments of sudden “out-of-the-blue” creative, perhaps world-changing ideas, are still an unresolved and intriguing question. Indeed, compared with the vast research effort that has been directed towards controlled, deterministic behaviors, far fewer studies attempted to decipher the neuronal mechanism underlying free and creative behaviors.
An important and pioneering line of research began with the seminal discovery of the “readiness potential” (Kornhuber and Deecke 1965). This slow buildup of EEG signal prior to motor movements was shown to specifically precede voluntary movements rather than externally determined ones. Elegant subsequent work by Libet et al. (1983), as well as more recent works (Soon et al. 2008; Fried et al. 2011; Schultze-Kraft et al. 2016) demonstrated that this buildup occurs mainly below the threshold of awareness, leading to a lively philosophical and psychological debate regarding free will (Frith et al. 2000; Wegner 2017; Maoz et al. 2019). Most studies considered the readiness potential as reflecting a subconscious goal-directed deliberation process. However, an intriguing alternative has been introduced by Schurger et al. (2012), who hypothesized that the readiness potential actually corresponds to a stochastic fluctuation generated by accumulation of neuronal noise.
Interestingly, this gradual anticipatory signal is not solely a motor-related effect, and is also observed a few seconds before free recall of previously displayed visual stimuli, but not before their direct viewing (Polyn et al. 2005; Gelbard-Sagiv et al. 2008; Norman et al. 2019). Extending on Schurger et al. (2012), we have hypothesized (Moutard et al. 2015) that in fact the entire range of free behaviors, from motor decisions regarding when to act, to moments of creative insight, all depend on slow stochastic fluctuations as their driving mechanism.
It should be noted that if this hypothesis is correct, then these stochastic fluctuations must be an extremely ubiquitous brain phenomena, since free behaviors can be found in a widely diverse set of tasks ranging from motor movements, visual imagery, memory recall, creative verbal generation tasks, and even music improvisation, to name a few (e.g., Libet et al. 1983; Finke 1996; Limb and Braun 2008; Benedek, Beaty, et al. 2014a; Norman et al. 2019). A second, obvious requirement of such a neural mechanism is that it should be active spontaneously in the absence of external stimuli or deterministic instructions. Intriguingly, a ready neuronal candidate fulfilling these requirements is glaringly evident in the phenomena of the resting-state activity fluctuations.
(A) A scheme of the hypothesis: resting-state fluctuations drive the formation of a new verbal idea. This effect is manifested as a gradual signal buildup, leading to the crossing of a threshold, and by so bringing the idea into conscious thought. (B) Schematic representation of BOLD resting-state fluctuations and average event-related voluntary responses with slow dynamics (left column) and fast dynamics (right column). (C) A scheme of the three experimental task blocks and the control block. VF blocks were initiated with a visual cue instructing the participant what category of words to generate. The participants then had to covertly generate words that belonged to the specific category, and press a button immediately each time an idea came to mind. They continued trying to generate relevant words until block termination after ~150 s, indicated by the cue “break,” followed by a 30-s break. AU blocks were designed similarly to the VF blocks, but were longer and lasted ~210 s. The participants’ task was to think of creative AU to an everyday object that was specified at the beginning of each block. Instances task blocks also had the same structure as the two former tasks. Here, the participants needed to generate as many instances of common concepts as possible, until the block terminated. The control blocks were identical in all three experiments. They were initiated by the instruction to covertly repeat a specific word that was different in each block, after each time an auditory cue was heard. After the covert repetition of the word, the participants pressed the button. Control blocks were ~120 s long, and the timings of the auditory cues replayed the participants’ performance in the 120 final seconds of the previous voluntary block.
Previous studies have shown that different resting-state activity characteristics of individual participants, including functional connectivity patterns, fractional amplitude of low-frequency fluctuations (fALFF), entropy and regional homogeneity measures are correlated with their creative behavior abilities, insight solutions quantity, and free verbal-generation performance levels (Kounios et al. 2008; Takeuchi et al. 2012; Beaty et al. 2014; Wei et al. 2014; Yin et al. 2015; Beaty et al. 2018; Shi et al. 2019; Sun et al. 2019). However, despite this extensive research, the functional role that the resting-state fluctuations may play in these free behaviors remains a mystery.
Here, we propose that resting-state activity fluctuations drive the generation of free and creative ideas, by raising the neural activity in task-relevant networks above the decision bound and thus eliciting the free behavior events. This hypothesis is illustrated schematically in Figure 1A. Specifically, we hypothesize that resting-state fluctuations can explore different possible verbal solutions by stochastically activating a diverse range of behaviorally relevant networks. When such stochastic activity crosses an activation threshold, this leads to the conscious emergence of a word or a verbal idea in the mind. Because of the inherently slow nature of resting-state fluctuations, the stochastic accumulation leading to the threshold crossing will appear as a slow, gradual build-up of activity preceding the event onset. A crucial prediction here is that if indeed a resting-state fluctuation actually constitutes the anticipatory buildup, then there should be a tight correlation between the waveforms of resting-state fluctuations, measured during rest, and the anticipatory buildup found prior to the free behavior. Thus, if there are substantial individual differences in the shape or dynamics of resting-state fluctuations, then these individual differences should also be reflected in the buildup appearing prior to free behaviors. This prediction is schematically illustrated in Figure 1B: individuals whose spontaneous fluctuations display very slow dynamics, will also show a very slow and sluggish anticipatory buildup prior to the voluntary event, as visualized in the left column time-courses in Figure 1B. In contrast, individuals whose spontaneous fluctuations are faster and steeper, will also show a sharper buildup prior to voluntary behaviors (right column in Fig. 1B).
We further hypothesize that different free behaviors, including creative insights, are all similarly driven by resting-state fluctuations. Hence, we predict to see a link between free behavior-related activations and resting-state fluctuations across all tasks tested in the study, but, importantly, not between resting-state fluctuations and externally determined behavior-related activations.
In the present study, we examined this hypothesis in the verbal-language domain, in a blood oxygenation level dependent (BOLD)-fMRI experiment. Participants underwent a resting-state scan, after which they completed three different free verbal tasks: VF, creativity, and divergent thinking tasks, as well as an externally driven verbal repetition task. By allowing long trial durations and extracting verbal events that are isolated in time, we were able to examine the neural activity that precedes the creative event onset, and compare it with the control, externally determined events as well as to resting-state fluctuations of individual participants.
Materials and Methods
Participants
In total, 24 healthy, right-handed participants (11 female, mean age 27.56 ± 3.89) with normal vision participated in the study that included three different tasks in two separate scanning sessions. One participant was excluded due to excessive motion. Another participant was excluded from one scanning session due to excessive motion, while his second session was maintained. All participants provided written consent and received payment for their participation. All procedures were approved by the local ethics committee. The experiment included three different tasks: a VF task, an alternative uses (AU) task and an instances of common concepts task (INST). Twenty-two participants participated in the VF task (10 female, mean age 27.5 ± 3.97), 20 in the AU task (10 female, mean age 27.7 ± 4.16), and 15 in the INST task (8 female, mean age 27.47 ± 4.75). Fifteen participants completed all three tasks, four completed only the VF and AU tasks, three completed only the VF, and one completed only the AU.
The Experimental Tasks and Design
Our study included a total of three different verbal tasks: VF, AU, and INST tasks, which were completed in two separate scanning sessions. One session included five runs of the VF task, and the other session included two AU runs and two INST runs. Session and run order were counterbalanced and randomized. All experimental sessions were initiated with a short training period outside magnet, in order to familiarize the participants with the task/tasks that were to be performed in the current scanning session. Next, the participants entered the scanner and underwent an 8-min resting-state scan. They were instructed to rest with open eyes, while maintaining their gaze in the center of a gray screen. Then, the participants either completed five VF task runs, or two AU and two INST runs. All tasks were performed in a covert manner, in order to avoid movement-related artifacts due to overt speech, specifically artifacts during the time period preceding the idea onset and report. An anatomical scan was also completed after two experimental runs. Figure 1C depicts the three tasks as well as the control condition, which was the same in all experimental runs.
Each VF block was initiated with a visual cue instructing the participants to covertly generate words from a specific phonemic or semantic category. The semantic categories were tools, birds, fish, vegetables, and USA states; the phonemic categories were words that begin with the Hebrew letters “yud,” “tzadik,” “pay,” “vav,” and “tet.” The participants were instructed to report exemplar generation events by immediately pressing a button every time they thought of a new word from the relevant category, after which a short auditory cue was heard. Each VF block lasted 2.5 min and was terminated by the visual cue “break,” followed by a 30-s break period.
Control blocks began with the visual instruction “repeat the word ‘__’ when the auditory cue is heard,” with a different specified word to be repeated on each block. Here, the participants needed to wait until they heard the auditory cue. Once it was played, they were instructed to covertly repeat the instructed word, and immediately afterwards press the button. Importantly, the auditory cues in the control blocks replayed the participants’ performance in the VF blocks, specifically reconstructing the participants’ inter-response intervals during the 100 final seconds of these blocks. This manipulation allowed us to compare similarly spaced control and VF events during the data analysis stages. Control blocks were 100 s long, and were also terminated with the visual cue “break,” followed by a 30-s break. Each experimental run lasted ~10 min, and included two VF blocks, one of a semantic category and one phonemic, and two control blocks, in random order. In both the control and VF conditions, the initial visual instructions for each block appeared for 2 s, after which a blank gray screen was presented for the rest of the block. Participants were instructed to maintain their gaze in the middle of the screen for the entire block duration.
The AU task, a divergent thinking task used commonly in creative thinking studies, required the participants to think of creative uses for everyday objects: a button, a paperclip, a barrel, and a cup (Guilford 1950; Kaufman et al. 2008; Arden et al. 2010; Dietrich and Kanso 2010; Beaty et al. 2014). AU blocks began with the instruction to generate creative uses for a specific object. The participants pressed the button every time a new idea came to their mind, after which the auditory “beep” was played. AU blocks were terminated after 210 s with the cue “break,” followed by a 30-s break period. Control blocks were similar to those in the VF experiment described above, requiring “passive” word repetition that replayed the participants’ AU block performance, with 120-s long blocks. Each AU run lasted ~13 min, and included two AU blocks and two control blocks, in random order.
The final task, the INST task, had a similar structure to the two previous tasks. Here, the participants needed to generate exemplars for common instances, including things that are heavy, loud, round and tall (Silvia 2011), and press a button whenever a new exemplar was generated. INST blocks were 210 s long, and control blocks were 120 s. Each run lasted ~13 min, and included two INST blocks and two control blocks in random order.
By setting long block durations in all three tasks, we were able to obtain voluntary and control verbal events that were separated in time with a relatively long “no-event” period preceding them. Specifically, only events that occurred 12 s or later from the previous event were further examined. This allowed us to inspect the “clean” neural activity leading up to the subjective experience of a spontaneous emergence of an idea, as compared with the control word repetition in response to an external cue, without contamination of previous event-related responses.
At the end of each experimental session, the participants filled out questionnaires reporting all the words and creative ideas they remembered generating for each category that was probed. This was done in order to ensure they had understood the tasks correctly, and generally sample their responses.
MRI Setup
Sixteen participants (of which all 16 completed the VF experiment, and 15 completed the AU and the INST experiments) were scanned in the 3 Tesla MRI scanner (Magnetom Prisma, Siemens), at the Weizmann Institute of Science, using a 20-channel receive head/neck coil. Functional images of BOLD contrast were obtained using a T2*-weighted gradient echo planar imaging (EPI) sequence (TR = 2000 ms, TE = 30 ms, flip angle = 75°, FOV 210 mm, voxel size 3.0 × 3.0 × 3.3 mm, 32 slices, tilted to the ACPC plane). Whole-brain T1-weighted anatomical images were acquired for each participant using a 3D magnetization prepared rapid acquisition gradient echo (MPRAGE) sequence (TR = 2300 ms, TE = 2.32 ms, TI = 900 ms, flip angle = 8°, voxel size 0.9 × 0.9 × 0.9 mm, with integrated parallel acquisition (iPAT) acceleration factor of 2).
The additional seven participants (of which six completed the VF experiment and five the AU experiment), were scanned in the 3 Tesla MRI scanner (Tim Trio, Siemens) at the Weizmann Institute of Science, using a 12-channels head matrix coil. Functional images of BOLD contrast were obtained using a T2*-weighted gradient EPI sequence (TR = 2000 ms, TE = 30 ms, flip angle = 75°, FOV 216 mm, voxel size 3 × 3 × 3 mm, 32 slices, tilted to the ACPC plane). Whole-brain T1-weighted anatomical images were acquired for each participant using a 3D MPRAGE sequence (TR = 2300 ms, TE = 2.98 ms, TI = 900 ms, flip angle = 9°, voxel size 1 × 1 × 1 mm).
Please note that for the VF and AU tasks the data were combined from two different scanners. This may be problematic, due to differences in signal-to-noise and additional across-scanner variability. However, scanner differences should have affected both the control and the voluntary task conditions, and hence were less likely to substantially impact the difference between these two conditions, which was our main finding. To further inspect this issue, we examined separately only the Prisma-scanned participants in the VF task, which contained small but sufficient number of participants (n = 16). The main results remained unchanged.
Pupil Size and Eye-Tracking Acquisition, Preprocessing, and Analysis
Pupil size and eye-movements of the participants’ dominant eye were recorded continuously throughout the experimental scanning sessions, using an Eyelink-1000 eye-tracking device (SR Research), at a sampling rate of 500 Hz. The pupillary data of three participants in the AU and INST sessions were not recorded, due to technical issues. Furthermore, scans in which the amount of missing data, due to technical issues or excessive blinking, were larger than 20% were also removed, resulting in the exclusion of two additional participants from the INST task analysis. In total, pupillary analyses included 22 participants for the VF task, 17 participants in the AU task, and 10 participants in the INST task.
The scaled index estimate of the pupil diameter, as recorded by the Eyelink system, was preprocessed using custom-made MATLAB scripts, following previously reported pipelines (Yellin et al. 2015; Broday-Dvir et al. 2018). Points of missing pupil data or unlikely pupil size (3 standard deviations [SDs] from mean pupil size within a trial) were removed, along with their neighboring data points 80 ms before the onset and after the offset of the detected segments. The resulting gaps of missing data were replaced using an inverse-distance weighted interpolation (Howat, University of Washington, 2007). The entire pupillary time-course of each run was then band-pass filtered, removing frequencies outside the 0.015–5 Hz range, using the least-squares FIR filter adapted from the EEGLab MATLAB toolbox (Delorme and Makeig 2004). Next, the time-courses underwent z-score normalization.
Pupil event-related response onsets were defined as the moment of the button press, by which the participants reported generating a new verbal exemplar or idea in the voluntary conditions, or repeating a specific word in the control condition. The pupillary time-courses for each event were extracted from −5 s before event onset, to 6 s after the event onset. Only creative and control events that were preceded by at least a 5-s event-free period were further analyzed, while the events that were not “isolated” in time, as defined, were discarded. Additionally, we ensured no eye-movement differences existed between the voluntary and control conditions, ruling out their potential effects on the pupil dilation differences between the conditions (see details in “Eye-movement control analysis in the Supplementary Methods and Analyses”). On average, 65.54 ± 21.32(SD) VF events, 21.59 ± 13.66 AU events, and 44.8 ± 10.42 INST events were obtained for each participant.
Next, pupil event-related responses were averaged across individual trials in each participant, separately for the voluntary and the control conditions in each of the three experiments, and the responses were then averaged across all participants. In order to examine the differences between the voluntary and control conditions, separately in each experiment, we used a paired two-tailed t-test at each time point, and corrected for multiple comparisons using FDR correction, according to Benjamini–Hochberg method (α = 0.05) (Benjamini and Yekutieli 2001).
fMRI Preprocessing
MRI data processing was achieved using FSL 5.0.4 (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/) and in house MATLAB codes (R2016b, The MathWorks). The functional task data were preprocessed using FEAT version 6 (FMRIB’s expert analysis tool), and included the following steps: removal of the first two volumes from each functional scan; motion correction using FMRIB’s Linear Image Registration Tool MCFLIRT (Jenkinson et al. 2002); brain extraction using BET (Smith 2002); high pass temporal filtering of 100 s; and spatial smoothing using a Gaussian kernel of FWHM = 6 mm. Similar preprocessing was done for the resting-state data, except for the smoothing. Tissue-type brain segmentation was run using FAST (Zhang et al. 2001), resulting in white-matter and ventricle masks for each participant. Functional images were aligned to the high-resolution anatomical volumes in each participant, initially using linear registration (FLIRT) and then optimized with Boundary-Based Registration (Greve and Fischl 2009). Anatomical images were transformed to MNI space using FMRIB’s Nonlinear Image Registration Tool (FNIRT), and the resulting wrap fields were applied to the functional images in order to allow the projection of all participants onto a common brain template.
Additionally, in the resting-state data, non-neuronal contributions to the BOLD signal were removed by linear regression of motion parameters and ventricle and white-matter time courses from the unsmoothed data, for each participant, with no further global regression performed on the data (Fox et al. 2009; Hahamy et al. 2014). Following this step, the resting-state data were spatially smoothed, with a Gaussian kernel of FWHM = 6 mm.
Whole Brain GLM Analysis
Whole brain statistical maps were computed using a general linear model (GLM), separately for each of the three tasks. Voluntary and control whole blocks were used as the regressors, as well as the 6 standard and 18 extended motion parameters, estimated by FSL MCFLIRT, in order to regress out motion artifacts. The regressors were convolved with the canonical double-gamma hemodynamic response function (HRF), attaining a model of the expected hemodynamic responses (Boynton et al. 1996). Multiple linear regression was performed for each run of each participant, obtaining estimates of the response amplitudes (beta values) in each voxel, for each of the conditions. The comparisons that were contrasted included the voluntary blocks (VF, AU, and INST) versus baseline, the control blocks versus baseline, and the voluntary blocks versus control blocks. Resulting contrast images were entered into a second-level analysis per participant, in order to combine the functional runs of each individual participant (fixed effects model). Next, group analysis was run using FMRIB local analysis of mixed effects (FLAME) (Smith et al. 2004), using the beta values of each participant as the dependent variables in a paired t-test in order to estimate intersubject random effects. Resulting Z-statistic maps were corrected for multiple comparisons using cluster correction, determined by a voxel threshold of Z > 3.1 and a corrected cluster significance threshold of P = 0.05 (Eklund et al. 2016; Greve and Fischl 2018). Statistical maps were projected on inflated and flat cortical surfaces in MNI space, constructed using Freesurfer 5.3 (Dale et al. 1999; Fischl et al. 1999).
An additional GLM analysis was run in order to define individual participants’ regions of interest (ROIs), for further ROI analyses. It was identical to the GLM described above, only instead of modeling the entire block durations with a single predictor per block (for both voluntary and control blocks, in all three experiments), we created separate predictors for the initial 26 s of each block, and predictors for the remaining latter part of each block. First level contrast images were entered into a second-level fixed-effect analysis, combining the separate functional runs of each participant. Individual ROI definition was based on the contrast of the initial 26 s of the voluntary blocks versus baseline.
ROI Definition and Time Course Extraction
ROIs were defined in order to carry out ROI analyses, examining the average responses and amplitude changes during the voluntary and control events, as well as the fALFF values during the verbal responses and resting-state time courses, in the three different experiments. ROIs known to be involved in language processing, as well as in the three tasks employed here (VF, AU, and INST), were defined individually for each participant, including the left inferior frontal gyrus (IFG) and left middle frontal gyrus (MFG), separated to an anterior region (LH Pars Triangularis) and a posterior region (LH Pars Opercularis); the left precentral gyrus (LH PreCG); the paracingulate gyrus (PCG) and ACC; and the DMN regions (Schlosser et al. 1998; Costafreda et al. 2006; Beaty et al. 2014). All ROIs were defined individually for each participant, based on their activation maps in conjunction with atlas-based anatomical masks, in order to take into account the intersubject variability in anatomical coordinates of functional regions. Critically, individual subjects’ activation maps were based on the contrast of the first 26 s of the voluntary blocks (VF/AU/INST) > baseline (breaks between blocks), while all further ROI-based analyses included only the events that occurred 30 s or later into the blocks, for both voluntary and control events. Thus, we avoid circularity in the definition and the analyses of the ROI data. This division of the blocks to an early 26 s ROI-definition period, and a latter analyzed period, was based on the behavioral patterns of the subjects (shown in Fig. 2C). In general, subjects began all three types of voluntary blocks with a relatively high verbal production rate on average, which decreased exponentially with time, with a clear drop in rate 30 s after the beginning of the block. Because the generation rate was initially high, the intergeneration time was too short to allow proper analysis of the events in the first 26 s. However, it is possible that the neuronal activity patterns differ between the early and the latter time periods of the block, because of, for example, serial order effects, and that these differences may affect ROI definition, and accordingly the results. Therefore, we ran an additional control, for the VF task, in which the ROIs were defined based on the latter part of the blocks (see Control ROI definition in the Supplementary Methods and Analyses).
Anatomical masks for the ROI definition were based on the Harvard-Oxford Cortical Structural Atlas (Desikan et al. 2006), and created for the LH Pars Triangularis, LH Pars Opercularis, LH precentral gyrus, LH insular cortex, PCG gyrus and anterior cingulate gyrus combined, and for the DMN regions, the precuneus and the posterior cingulate cortex (PCC) combined. An additional DMN region, the inferior parietal lobule (IPL), was defined based on the Julich histological atlas (Caspers et al. 2008). All ROIs, except for the two DMN regions, were defined as the voxels that showed increased activation in each subject’s contrast of the first 26 s of the voluntary blocks (VF/AU/INST) > baseline, thresholded at P < 0.0001, in conjunction with the relevant anatomical masks. DMN ROIs were defined as the voxels that showed decreased activations in this contrast, using the same threshold. If no voxels survived the thresholding, it was gradually lowered until P < 0.01, to obtain a minimal ROI of ~30 voxels. If still no voxels survived, the ROI region was not defined for that participant. The LH anterior insula ROI was defined only for the VF task, and not for the AU and instances task, as it did not survive the thresholding in most participants in these two tasks.
ROI time courses were extracted for each ROI in each participant, for the experimental runs as well as for the resting-state scans. Before the extraction of the ROI time courses, further motion correction procedures were done: in addition to regressing out the 6 standard and 18 extended motion parameters estimated by FSL MCFLIRT, that was previously run (see fMRI preprocessing and Whole brain GLM analysis), we also identified and excluded TRs with head movements using a scrubbing procedure (Power et al. 2012). This step was performed mainly since our block durations were relatively very long (150 s for the VF task, and 210 s for the AU and INST), thus increasing the possibility of small head movements and related motion artifacts with time. Post-hoc control analysis was also run to ensure that no differences in head movement patterns were found between the voluntary and control conditions during the button presses and auditory cue presentations, in all three tasks (see Ruling out possible head motion related artifacts in Supplementary Methods and Analyses).
Following these steps, we averaged the BOLD signals across all voxels in each ROI, computing the time courses for each ROI in each participant.
ROI Event-Related Response Analysis
In order to examine the event-related ROI responses, we first normalized the ROI BOLD time courses to percent signal change, relative to the mean BOLD amplitude across each entire run. The BOLD responses were locked to the onset of the voluntary and control verbal events, indicated by the button presses, excluding events that occurred in the first 30 s of each block, as explained in the “ROI Definition and Time Course Extraction” section above. Furthermore, only events that were separated by at least 12 “event-free” seconds from the previous event were analyzed, for both the voluntary and control conditions, thus allowing examination of the activity preceding these events that is not contaminated by residual activity from previous events. On average, 21.68 ± 5.79(SD) VF events, 13.06 ± 5.38 AU events, and 16.07 ± 4.29 INST events were obtained for each participant. Event-related responses were averaged across trials and experimental runs for each participant, separately for the voluntary and control conditions in the three experimental tasks, from −8 to +12 s relative to the button press. The BOLD responses were then averaged across all participants. The differential time courses of the two conditions were also calculated, defined as the within subject average voluntary response minus the control response, averaged across participants. Paired t-tests were conducted in each ROI and condition, comparing the BOLD amplitude at each time point to the baseline amplitude, defined as the average amplitude at times −6 and −4 relative to event onset.
fALFF Analysis
In order to quantify and compare the time-frequency dynamics during the resting state, to those of the voluntary and control responses, we calculated the fALFF (Zou et al. 2008; Zuo et al. 2010) in these conditions for each participant. The fALFF values of the voluntary and control responses were calculated from the average event-related response of each participant, for all the ROIs defined, separately for the three tasks (VF, AU, and INST). These individual averaged event-related responses were obtained as described in the ROI Event-Related Response Analysis section, following the preprocessing steps described in fMRI preprocessing and ROI definition and time-course extraction, and percent signal change normalization. These mean responses, for each participant, each ROI, and each condition, were transformed to the power-frequency domain using the Fourier frequency transform. Next, the sum of the square root of the power amplitudes across the 0.01–0.1 Hz frequency range was divided by the sum of the square root of the power across the entire frequency range measured (0–0.25 Hz). This resulted in an fALFF value for each subject and each ROI, for the mean VF, AU, and INST responses, and their matching control responses. Similarly, we obtained the fALFF values of the resting-state time course of each subject, for the two resting-state scans that were performed (prior to the VF experiment, and prior to the AU and INST experiments, as explained in The Experimental Tasks and Design section). Resting-state data first underwent preprocessing, including regressing out the motion parameters, white matter and ventricle time-courses, as described in “fMRI preprocessing” and ROI definition and time course extraction. Next, the resting-state time-courses were transformed to the power-frequency domain, by Fourier Transform, and the square root of the power was calculated for each frequency. Then, the sum of the square root amplitudes across 0.01–0.1 Hz was divided by the sum across the entire frequency range, resulting in an fALFF value for each participant and each ROI.
Finally, the resting-state fALFF values were correlated with the voluntary and the control fALFF values, separately for the VF, AU, and INST tasks, using Spearman correlation. All correlation P-values were derived from a nonparametric permutations test, with 10 000 random subject-wise permutations, followed by FDR correction at α = 0.05, to correct for multiple comparisons across the different ROIs.
Results
In this study, we aimed to compare the neural activity associated with spontaneous emergence of a freely generated idea or a verbal exemplar, with the activity during an externally determined condition, and their relationship to resting-state dynamics. To that end, participants underwent resting-state scans, and completed three different voluntary verbal tasks inside the fMRI: a VF task (n = 22), an AU task (n = 20), and a common instances (INST) task (n = 15). These voluntary tasks were compared with a control condition of repeating an externally given word. Figure 1C depicts the three tasks as well as the control condition, which was the same in all experimental runs.
Behavioral Results
A number of parameters of the participants’ behavior were examined, including their performance, quantified as the total number of words or creative ideas they produced, and the temporal dynamics of their production rate across the experimental blocks duration. Participants produced on average 20.64 ± 5.0 exemplars in each VF block, 9.75 ± 6.22 ideas in each AU block, and 23.13 ± 9.91 items in every instances block. One-way ANOVA analysis followed by post-hoc comparisons revealed that the number of words produced in the AU blocks was significantly lower than in the other two tasks, and that the performance in the VF and INST tasks was not statistically different (F(2,61) = 14.81, P < 0.0001; Bonferroni post-hoc comparisons: VF > AU, P < 0.001; INST > AU, P < 0.001; VF ≅ INST, P = 0.875). Supplementary Figure 1 presents the average number of words generated in each specific category, for the three experimental tasks.
Next, we correlated the participants’ performance, quantified as the average number of words they produced in an experimental block, across the three tasks. Significant positive correlations in the participants’ performance were found between all three tasks, as shown in Figure 2A (Pearson’s R and P-values displayed in the figure). Figure 2B depicts the distribution of the interevent time intervals (IEIs) between consecutively produced words in the three experimental tasks. The IEI occurrence numbers were averaged across participants, using 2-s bins in the VF and instances task, and 4-s bins in the AU task. The mean IEIs between consecutive word productions for each of the three tasks, averaged across the entire block duration, are also presented in Figure 2B. The IEIs distribution showed a common trend in all three tasks, with a high frequency of short IEIs that decreases for longer IEIs. Furthermore, the IEI distributions fit a power law distribution, linearly fitting the log–log function, as shown in the insets in Figure 2B. The vertical dashed lines in the plots mark the 12-s IEI cutoff for event-related analyses: only events that were preceded by an IEI that was at least 12 s or longer were further analyzed.
(A) Correlations between the average number of words produced by individual participants in the three different tasks. Pearson’s R correlation coefficients and their P-values are specified. (B) The IEI distribution, averaged across participants. Error bars denote the count average across participants ± SE, for each IEI bin (2-s bins in the VF and INST plots, and 4-s bins in the AU plot). The total mean IEI value ± SE is indicated for each of the three tasks. The dashed lines indicate the 12-s mark, the IEI cutoff for further event-related analysis. The insets show the log–log plot and its’ linear fit for each task. (C) Mean number of words produced during 10- or 15-s time bins (VF and AU/INST, in accordance) along the trial, averaged across all participants. Error bars denote the mean ± SE. The total mean number of words produced across the entire block duration ± SE is indicated for each of the three tasks. The insets show the log–log plot and its’ linear fit for each task.
In order to inspect the temporal dynamics of word generation across the block duration, we divided the blocks to 10- (for the VF task) or 15-s (for the AU and INST tasks) time bins. Next, we counted the number of words generated in each bin, averaged across blocks for each participant, and finally averaged across participants. Interestingly, as demonstrated in Figure 2C, there was remarkable similarity in the behavior across the three tasks, with an initially high word production rate, that decreased along block duration. As in the IEI distribution, the mean performance rate across time also followed the power law, linearly fitting a log–log function, shown in the insets in Figure 2C.
Anticipatory Buildup in Pupillary Dilation
Previous studies have shown that pupil dilation is a reliable measure for task engagement and cognitive effort, even during covert behaviors (Einhauser et al. 2008; de Gee et al. 2014; Yellin et al. 2015; Broday-Dvir et al. 2018). Therefore, we measured eye movements and pupillary dilations during the three different experimental tasks, and compared the voluntary conditions with the externally driven control conditions. Average pupil size was locked to the word generation onsets, as indicated by the button press, for events that were separated by at least 5 s from the previous report, in order to avoid confounding signals from the previous event. Figure 3 depicts the pupillary dilation response during the free versus determined control events for the three tasks. A two-tailed paired t-test was calculated for each time point, comparing pupil size between the voluntary and the control conditions. As can be seen in the left panel, pupil size was significantly larger between times ~−1.7 s and ~0 s relative to the word generation onset during the VF events, as compared with the control events, indicated by the yellow line (P < 0.05, two-tailed paired t-test, fdr corrected). Similarly, pupil dilation was higher prior to the AU events, as compared with the control events, during the time period between ~−2.8 and −0.3 s preceding the verbal event generation (P < 0.05, two-tailed paired t-test, fdr corrected). The INST task showed a similar pattern of higher activity ~1 s prior to the voluntary events, as compared with the control events (P < 0.05, two-tailed paired t-test, uncorrected), though the effect was weaker, likely due to the smaller number of participants with pupil measurements in this task. This task also showed a higher peak amplitude for the control response as compared with the instances mean response (P < 0.05, two-tailed paired t-test, uncorrected), which could be related to the unpredicted sudden appearance of the external auditory cue (a similar trend, though insignificant, is seen in the AU vs. control responses). Thus, a 1–3 s slow anticipatory buildup was evident in pupil diameter preceding the voluntary, but not the externally determined verbal events in all three experiments.
Group average pupil response during VF, AU, and INST events and control verbal events, locked to the button presses. Transparent borders indicate the mean ± SE. Yellow lines indicate a significant difference between the two conditions (two-tailed paired t-test, P < 0.05, fdr corrected). Orange line indicates P < 0.05, uncorrected (though note the smaller number of participants in the INST task).
Whole Brain BOLD GLM Results
Next we examined the BOLD-fMRI activations associated with the three tasks. Figure 4 depicts the whole brain GLM analysis contrasting the voluntary and control entire blocks for the VF (A), AU (B), and instances tasks (C). Unfolded cortices are presented on the left, showing the t-value group maps, with the pale-transparent colors reflecting the zero-threshold t-values, in order to allow full comparison between the three tasks. Overlaid, bright-colored and outlined activations, in warm colors (for increased activations) or cool colors (reflecting decreased activations) shown on the unfolded cortex, as well as on the inflated cortices on the right-hand side, depict the cluster corrected activations (voxel threshold of Z > 3.1 and a corrected cluster significance threshold of P < 0.05).
Whole brain t-value group contrast maps, generated by a random effects GLM analysis, for (A) VF > control whole blocks, (B) AU > control whole blocks, and (C) Instances > control whole blocks. The activation maps are presented on top of inflated cortex (right panels, P < 0.001, cluster corrected) and flattened cortex (left panels, pale-transparent colors show zero-thresholded data; outlined activations mark significant increases and decreases in activity, P < 0.001 cluster corrected). Abbreviations: paracingulate (PCG), precentral gyrus (PreCG), middle frontal gyrus (MFG), inferior frontal gyrus (IFG), anterior insula (AI), inferior temporal gyrus (ITG), intra-parietal sulcus (IPS), precuneus (PCUN), inferior parietal lobule (IPL).
As can be seen, all three tasks show increased BOLD activity during the voluntary blocks as compared with the control blocks in the left dorsal prefrontal cortex, specifically in the left inferior and middle frontal gyri (LH IFG and MFG), regions well-known to be involved in language production and processing. Increased activation was also apparent in the left precentral gyrus (PreCG), possibly related to the report of the generated verbal ideas by a button press. The anatomical distribution of these significant activations in the VF and INST tasks was relatively similar, while the extent of AU left prefrontal activation was more constrained. This could possibly be due to the significantly fewer ideas generated in AU blocks as compared with VF and INST blocks (see Behavioral Results section). Additionally, VF and INST blocks, as compared with control, showed significantly increased activations in the left anterior insula (AI), medial frontal regions including the PCG and ACC, the inferior temporal gyrus (ITG), and intraparietal sulcus (IPS). AU blocks displayed a positive trend of activations in these regions as well, though they were weaker and did not survive statistical corrections for multiple comparisons. These activations are in agreement with previous reports from VF and divergent and creative thinking imaging studies (e.g., Schlosser et al. 1998; Gonen-Yaacovi et al. 2013; Wagner et al. 2014; Wu et al. 2015).
In addition, the voluntary conditions, as compared with the control, showed significant decreases in DMN activity, specifically in the IPL in all three tasks, and in the precuneus (PCUN) and PCC in the VF and AU tasks (with an insignificant trend in the INST task). Supplementary Figure 2 depicts the average time-courses of the DMN responses to verbal events, for the IPL and PCUN ROIs. Both DMN ROIs revealed consistent gradual decreases in BOLD activity 2–4 s prior to the onset of voluntary verbal events, but not before the externally driven control events. As denoted by the yellow line, the DMN activity was significantly lower around the time of the voluntary event as compared with the control event, in all three tasks (two-tailed paired t-test, P < 0.05, cluster correction).
Anticipatory Buildup in BOLD Activity
In order to examine the dynamics of the fMRI-BOLD activations during the experimental tasks, we defined task-relevant ROIs for each participant individually, including the anterior and posterior LH IFG (or pars triangularis and pars opercularis, accordingly). We then extracted the BOLD ROI time-courses, locked to the time of the voluntary and control verbal generation events, excluding events that occurred in the first 30 s of each block, and events that were separated by less than 12 s from the previous event (see ROI definition and time-course extraction and ROI event-related response analysis in Materials and Methods).
Figure 5 depicts the mean BOLD event-related activations in the anterior LH IFG (pars triangularis) ROI, outlined in cyan in the inflated cortices, for the three tasks. It depicts the voluntary responses (left panels), the control responses (middle panels), and the difference between them, defined as the within subject voluntary minus control time-course differences, averaged across participants (right panels). Time points in which the response amplitude was significantly higher than baseline, defined as the mean amplitude across times −6 and −4 s before event onset, are denoted by the yellow line (P < 0.05, one-tailed paired t-test). Examining the VF response (Fig. 5A), it can be seen that the voluntary response showed a gradual build-up, reaching a significant increase above baseline at time 0, and preceding the control response, that reached significance only 4 s later. This trend was also seen in the differential averaged time-course, though reaching significance only later, at 2 s after event onset. Importantly, the fMRI time series were not shifted to compensate for the hemodynamic lag, so that the actual neural activity was increased ~2–4 s earlier.
Group mean voluntary and control word generation responses, and their averaged response differences in the anterior LH IFG (pars triangularis), for: (A) VF task; (B) AU task; and (C) instances (INST) task. Inflated left hemispheres display, for each task, the group contrast of the first 30 s of the voluntary blocks (VF/AU/INST) > baseline, thresholded at P < 0.0001 uncorrected, shown for ROI definition illustration purposes. The anterior LH IFG ROI is denoted by the cyan outline. The ROIs were defined individually for each participant, based on this contrast at the single subject level in conjunction with the anatomical mask of the pars triangularis (see ROI definition and time-course extraction in Materials and Methods for details). Note the presented ROI outlines here are defined based on the group maps, solely for visualization purposes. Mean percent signal changes in the BOLD signal, averaged across participants, are presented in red for the voluntary conditions, blue for the control conditions, and gray for the mean response difference, defined as the voluntary event time-course minus the control time-course. Transparent borders indicate the mean ± SE. Dashed vertical lines indicate the time of the button press, reporting a verbal generation event (voluntary and control). Yellow lines indicate a significant increase in response amplitude above baseline, defined as the average amplitude across times −6 and −4 s before event onsets (one-tailed paired t-test, P < 0.05).
Similar analyses were conducted for the AU (Fig. 5B) and INST (5C) tasks. As in the VF condition, the results revealed a significant anticipatory buildup of BOLD activation during the voluntary, but not during the control conditions. Specifically, the AU response reached a significant increase at time −2 s before the voluntary event onset, and the INST response at time 0. Increases in both voluntary responses preceded their corresponding control responses, evident also in the differential time-courses.
Similar dynamics were observed in the other ROIs studied. Supplementary Figures 3–5 show the voluntary, control, and difference averaged time-courses for the VF experiment (Fig. 3), the AU experiment (Fig. 4) and the INST experiment (Fig. 5), for all the ROIs defined, specifically the posterior LH IFG (pars opercularis), PCG cortex, LH precentral gyrus, and the LH anterior insula for the VF task. The preceding anticipatory buildup was evident in all the task-related ROIs that were examined, prior to the voluntary verbal events, but not before the externally driven control events, in the three experimental tasks (except for the PCG in the AU task, in which there was no difference between the voluntary and control conditions in the time of the significant increase in amplitude). This suggests that the anticipatory buildup was a common effect in all ROIs participating in the voluntary task.
Additionally, we normalized the responses of individual participants to their peak value, resulting in equal peak amplitudes of 1 in both the voluntary and control responses (see BOLD Amplitude Normalization Control in Supplementary Methods and Analyses). Even after this procedure, the gradual buildup during the voluntary condition preceded the control-related activations, demonstrating that this result was not related to possible amplitude differences between the conditions (see Supplementary Figs 6–8 for the VF, AU, and INST tasks, accordingly).
Thus, our results demonstrate a significant gradual buildup both in pupil dilation as well as in BOLD activity 1–2 s prior to the voluntary events, but not before the control, externally determined events. However, it is important to note that this gradual buildup was found after averaging across multiple individual events. Critically, it could be the case that the gradual buildup did not occur at the single trial level, but was rather artificially produced by inaccuracies in the timing of the participants’ reports of the occurrence of their freely generated verbal events. A simulation of this artificial buildup is presented in Figure 6A, demonstrating how averaging across abrupt, step-wise events, with jittered onset times, can appear as a gradual buildup. Importantly, the mean of such time-jittered step function trials is indistinguishable from the mean of true gradual buildup occurring on every single trial, as shown in Figure 6B.
(A) Step-function model simulation. The left panel displays the simulated neural activity of individual trials (n = 30), shown in different colors, with a jittered onset time of the amplitude step-increase. The mean time course is shown in a thicker black line. The middle panel shows the BOLD response estimates of the individual trials, as well as the averaged signal, obtained through convolution of the neural activity simulation estimates with the standard HRF (Boynton et al. 1996). The left panel shows the across-trials variance time-course of the simulated BOLD responses, with a clear significant increase above baseline between −2 and 0 s relative to “event onset” (P < 0.005, permutation test). (B) Gradual-slope model simulation. Similar to (A), only here, all trials have a gradual positive slope, which is constant with some added noise. The simulated BOLD variance time course in this case is flat, with no significant changes. (C) Variance time courses of experimental data from the anterior LH IFG ROI, in all three experiments, averaged across participants, for the voluntary conditions (red) and control conditions (blue). The time-courses are locked to event onset, as reported by button presses. In all tasks and conditions, the intertrial variance remained relatively flat at all times, and did not significantly increase above the baseline variance. Transparent borders indicate the mean ± SE in all plots. See variance control simulation and analysis in Supplementary Methods and Analyses for additional details on the simulation and analyses.
Fortunately, the across-trial variance of these two simulated possibilities can clearly differentiate between them: in the case of individual jittered step-functions, the variance time course shows a significant peak coinciding with the time of the anticipatory buildup, as shown in the right panel of Figure 6A. By contrast, single trial gradual-ramping stimulation displayed a flat across-trial variance time course, as seen in the right panel of Figure 6B (see Variance control simulation and analysis in Supplementary Methods and Analyses for additional details).
Figure 6C displays the across-trial variance time courses of the experimental data from the anterior LH IFG, for the voluntary and control conditions separately. All three tasks reveal relatively flat variance time-courses, with no significant increases during the anticipatory buildup duration, relative to baseline variance (P > 0.3 for all time points in the voluntary conditions; P > 0.15 for all time points in the control conditions; one-tailed paired t-test). The variance time courses of the additional ROIs, as well as the pupil, also did not reveal significant increases in variance during the duration of the gradual signal increase. Thus, our variance analysis clearly ruled out the possibility that the observed anticipatory buildup was due to a time jitter of step-function-like, or extremely rapid, activations.
Resting-State Fluctuations are Correlated to the Anticipatory Buildup
Finally, we examined the critical question: was there a link between the resting-state BOLD fluctuations, measured in a resting-state scan prior to the experimental task runs, and the anticipatory buildup preceding the free verbal and creative events? In order to examine this question, we employed the commonly used measure of the fALFF, a method that allows measuring the relative portion of slow frequency fluctuations from the entire detectable power spectrum (Zou et al. 2008; Zuo et al. 2010) (see fALFF analysis in Materials and Methods). Furthermore, as shown in Supplementary Figure 9, fALFF values calculated from simulated “event-related” responses and simulated “resting-state time courses” are correlated with the slopes of these responses and time-courses, suggesting that this measure holds information regarding the shape of these responses or fluctuations. Furthermore, the fALFF values of simulated “event-related responses” and “resting-state time courses” that share a common slope, also show a significant positive correlation, shown in Figure 9, thus further emphasizing the logic behind the usage of this common method for uncovering a link between the experimental resting-state fluctuations and event-related responses (for additional details, see fALFF simulation in Supplementary Methods and Analyses).
Therefore, we calculated the fALFF values of each individual participant’s resting-state activity fluctuations, as well as the fALFF values of the mean VF, AU, and INST voluntary event-related responses. The fALFF values of the matched average deterministic control responses from the three tasks were also obtained. To examine the possible link between resting-state fluctuations and the anticipatory buildups, we correlated the fALFF values of the resting-state fluctuations with the fALFF values of the voluntary condition responses, and compared them with the correlations between the resting state and the control response activations, across participants. The results of this analysis for all three tasks, from the anterior LH IFG (pars triangularis) ROI, are shown in Figure 7. Left-side scatter plots depict the correlation between individual participants’ voluntary event-related responses and the resting-state time courses, and the right-side plots show the correlations between the control activations and the resting-state fluctuations (VF task results shown in A; AU task shown in B; and INST task in C). As can be seen, significant correlations were found between the voluntary average buildups and the resting-state fALFF values, in all three tasks (VF task: Spearman’s R = 0.71, P < 0.001; AU task: Spearman’s R = 0.52 P = 0.017; INST task: Spearman’s R = 0.62 P = 0.009. All P-values derived from a subject-wise label-shuffling permutation test, across 10 000 permutations, and survived FDR correction for multiple comparisons at P < 0.05). Critically, no significant correlations were found between the fALFF values of the mean deterministic control responses and the resting-state time courses (VF task: Spearman’s R = 0.10, P = 0.32; AU task: Spearman’s R = −0.11 P = 0.65; INST task: Spearman’s R = −0.05 P = 0.56. All P-values derived from a subject-wise label-shuffling permutation test, across 10 000 permutations). To directly compare these correlations between the anticipatory buildup during voluntary events and the resting-state dynamics to the correlations between the control activations and resting state, separately for each of the three tasks, we used a percentile bootstrapping test for comparing robust dependent correlations (Wilcox 2016). This analysis revealed that all voluntary versus resting-state correlations were indeed significantly higher than their counterpart control versus resting-state correlations (VF task: P = 0.02; AU task: P = 0.032; INST task: P = 0.008; dependent correlations test, one-sided α = 0.05).
fALFF correlation plots, depicting the correlations between the fALFF values of participants’ mean voluntary responses versus the fALFF values of individual participants’ resting-state time-courses in the anterior LH IFG ROI/pars triangularis (shown in red markers, each marker specifies an individual participant); and the correlations between fALFF values of participants’ mean control activations versus the fALFFs of their resting-state time-courses (shown in blue markers), for the same ROI. The results are shown for the VF task (A), the AU task (B), and the INST task (C). The inflated left hemispheres presented here denote the anterior LH IFG ROI, and are identical to those in Figure 5 (see Fig. 5 legend for details). Spearman’s R correlation coefficients are presented in the plots, together with their P-values, derived from a subject-wise label-shuffling permutation test (10 000 permutations). Significant correlations are marked with an asterisk (P < 0.05, FDR corrected for multiple comparisons across the additional ROIs defined, presented in Supplementary Figs 10–12). All tasks showed significantly higher correlations between the voluntary buildups and resting-state fALFFs, than between the control activations and the resting-state fALFFs (P < 0.05, dependent correlation percentile bootstrapping test, Wilcox 2016).
The voluntary versus rest and control versus rest fALFF across-subject correlations were inspected in additional ROIs, aside from the anterior LH IFG, including the posterior LH IFG (or pars opercularis), PCG cortex, LH precentral gyrus, and LH anterior insula (for the VF task only). The correlation plots for these ROIs are presented in the Supplementary Figure 10 (for the VF task), Supplementary Figure 11 (AU task) and Supplementary Figure 12 (INST task). Significant correlations between the fALFF values of mean voluntary responses and resting-state fluctuations were found in the posterior LH IFG ROI in the VF and AU tasks (VF task: Spearman’s R = 0.58, P = 0.002; AU task: Spearman’s R = 0.57 P = 0.006; all P-values derived from a subject-wise label-shuffling permutation test, across 10 000 permutations, and survive FDR correction at P < 0.05), and these correlations were higher than the nonsignificant correlations between control and resting-state in this ROI (VF task: P = 0.012; AU task: P = 0.04; Wilcox dependent correlations test, one-sided α = 0.05, statistical significance marked by yellow frames in Supplementary Figs 10 and 11). Yet this effect in the posterior LH IFG was not evident in the INST task (Spearman’s R = 0.02, P = 0.47, permutation test). Additional positive correlations were found between the AU voluntary response and resting state in the PCG cortex (Spearman’s R = 0.68, P < 0.001, permutation test, survived FDR correction at P < 0.05) and between the INST voluntary response and rest in the LH precentral gyrus (Spearman’s R = 0.64, P = 0.008, permutation test, survived FDR correction at P < 0.05), though in both cases, these correlations were not statistically higher than the correlations between control and rest (P = 0.12 and P = 0.31, accordingly, Wilcox dependent correlations test, one-sided α = 0.05). Importantly, no significant correlations were found between the mean control responses and resting-state fluctuations, in all tasks and in all ROIs examined (see Supplementary Figs 10–12 for all correlation plots and R and P values).
Additionally, we run a control ROI definition analysis, as there might be differences in neuronal activity patterns between the initial and latter parts of the block, due to, for example, temporal order effects, influencing the ROI obtained. Therefore, we defined the anterior LH IFG ROI for the VF task, based on the latter, instead of the initial, period of the block (see Control ROI definition in the Supplementary Methods and Analyses for details). The correlation results did not change in a significant manner; see Supplementary Figure 13, further establishing this finding.
Together, our analyses reveal significant correlations between the dynamics of the anticipatory build-ups preceding voluntary verbal events, and the dynamics of resting-state fluctuations. Critically, this correlation does not exist between the externally driven control verbal events and resting state, in all three tasks.
Discussion
Our study supports the notion that a common neuronal mechanism underlies all types of free voluntary behaviors. In our study, we focused specifically on voluntary verbal behaviors, including the free generation of verbal exemplars, ideas, and creative thoughts. By using three different tasks—a very common language production and fluency task (verbal fluency or VF), a classic verbal creativity task (alternative uses or AU) and a verbal divergent thinking task (instances or INST), we were able to highlight the common neuronal mechanism for the internal, unpredictable, “free” generation of a verbal idea, that can be generalized across different tasks and contexts.
Our findings thus extend previous studies (reviewed in Moutard et al. 2015) indicating that a common neuronal “signature” of free behavior (as we operationally defined in the introduction) is a slow buildup of activity in the relevant task-related networks preceding the actual moment of free behavior. In the present paper, we have extended this common principle to the case of free verbal behaviors. Specifically, in all three tasks, we found a slow, gradual buildup of BOLD signal preceding the reported time of the freely generated verbal idea or creative event by ~1–2 s, evident in language and additional task-relevant brain regions (see Fig. 5). Importantly, this gradual anticipatory increase was not present before the control events: deterministic, externally driven word repetitions. Thus, the anticipatory buildup appears to be specifically linked to free, voluntary, internally generated events, and not to verbal responses in general.
Further support for the validity of the anticipatory buildup as a signature of free behavior is provided by our pupil dilation measures. These pupillary measurements were conducted concurrently with the task performance and manifested a slow buildup of similar dynamics prior to the voluntary “free” events, but not before the deterministic control events (as shown in Fig. 3). This buildup likely reflects the increased processing and cognitive demand in the participants prior to voluntary, but not externally driven events, as previous works, by ourselves and others, have proposed that such pupil dilations provide a reliable index to processing levels (e.g., Kahneman and Beatty 1966; Alnaes et al. 2014; de Gee et al. 2014; Yellin et al. 2015; Broday-Dvir et al. 2018). This observation was also nicely compatible with the suggestion that free behaviors are characterized by a slow anticipatory buildup of subliminal activity. Furthermore, the increased pupil diameter prior to voluntary, but not control events, nicely fits with previous results demonstrating increased pupil size during internally directed, as compared with externally directed attention (Benedek et al. 2017).
An important concern to be ruled out is that the observed gradual ramping of activity was in fact merely an artifact, caused by the lack of precision in participants’ ability to accurately report the timing of the emergence of the verbal idea or exemplar. Thus, it could be argued that in fact each free-behavior event is characterized by a rapid, step-wise, activation increase (see Fig. 6). Under such an interpretation, the gradual buildup that was observed might in fact be merely a “smeared” byproduct of averaging multiple step-functions with jittered timings. Indeed, in our simulation analysis (presented in Fig. 6), we were able to recreate a slow buildup by averaging a set of such temporally jittered step-function responses. However, a major discrepancy between the true, single trial, gradual buildup model we proposed, and the averaged, jittered step-function model, is revealed when considering the intertrial variance (right-side panels in Fig. 6A,B). Here, the jittered step-function model predicts a significant increase in the across-trial variance during the buildup period (due to the jittered event timing), while the single trial-ramping model predicts similar variance levels across the entire anticipatory period. Careful examination of the BOLD variance during our experiments revealed a flat variance time-course, with no increases (or any significant differences at all) compared with baseline variance levels, unequivocally supporting a real gradual buildup prior to every voluntary event (see Fig. 6C).
What could be the neural mechanism that accounts for the slow anticipatory buildup? Two aspects of this phenomenon are helpful in narrowing the range of possibilities. First, free behavior is an extremely ubiquitous phenomenon, occurring at diverse modalities and functions (Moutard et al. 2015), from the classically studied decisions to perform a simple movement (Libet et al. 1983; Schurger et al. 2012), to spontaneous music and dance improvisation, as well as idea generation and creative behaviors. Even tasks that are typically considered to be mainly stimulus driven, such as visual perception, can manifest free or voluntary aspects, for example in the spontaneous alterations during bistable perception (Gelbard-Sagiv et al. 2018) or spontaneous visual imagery (Norman et al. 2017, 2019). Consequently, if our hypothesis states that there is a shared neural mechanism underlying this diverse set of free behaviors, then it should be a ubiquitous neuronal process, that can be found essentially across all cortical networks.
Another key aspect of this neuronal process is its slow dynamics: examining both the BOLD activity changes and the pupillary dilations here, as well as EEG and ECoG signals from previous studies of free behavior (Gelbard-Sagiv et al. 2008; Schurger et al. 2012; Norman et al. 2019), reveals a process that has a time constant of 1–2 s. This is an order of magnitude slower than the typical 200-ms stimulus–response cortical dynamics (e.g., Bitterman et al. 2008; Fisch et al. 2009; Podvalny et al. 2017).
Examining the possible neuronal candidates manifesting both slow and ubiquitous cortical activity reveals a readily obvious candidate: the spontaneous (also termed resting state) fluctuations that have been observed and studied extensively across essentially every human cortical network (Biswal et al. 1995; Arieli et al. 1996; Nir et al. 2006; Fox and Raichle 2007). So how can these ultra-slow resting-state fluctuations account for the ramping buildup observed prior to free behaviors? Our hypothesis, extended from an earlier proposal by Schurger et al. (2012) for the case of volitional movements, is illustrated in Figure 1A. Essentially, we propose that free and creative behaviors are initiated by the slow resting-state fluctuations. These fluctuations drive the network that is relevant for the voluntary task across a decision threshold. Thus, prior to free behavior, the activity in the network manifests slow spontaneous fluctuations, and when such a fluctuation crosses the activation threshold, a free behavior can emerge. A strong prediction of this hypothesis is that prior to each and every free behavior event, we expect to see the slow uprising phase of a spontaneous fluctuation, hence the slow ramping activity evident prior to the free verbal and creative moments shown here.
An obvious counter-argument to this proposed mechanism could be that the observed similarity between the slow time constants of the resting state and the anticipatory slow buildup are a mere coincidence. Accordingly, it could be argued that the two phenomena are unrelated, and are derived from completely different mechanisms that simply happen to both exhibit slow time constants.
If this was indeed the case, the temporal dynamics of the resting-state fluctuations and the anticipatory buildup should also be independent of each other. However, measuring the response characteristics across individuals revealed significant positive correlations between the fALFF of the resting-state time-courses and the anticipatory buildup preceding the freely generated verbal responses (see Fig. 7). This correlation was evident across all three different tasks that were examined, including the VF task—a classic language production task (Troyer et al. 1997; Schlosser et al. 1998), the AU task—used commonly for verbal creativity assessment (Guilford 1967; Torrance 1988), and the instances task, a divergent thinking task (Silvia 2011). This replication across the three separate tasks suggests that the resting-state fluctuations constitute a common mechanism underlying a variety of diverse voluntary verbal behaviors. The correlation between the anticipatory buildup and the resting-state fluctuations was spatially specific to the LH IFG, the central region involved in language processing and generation (e.g., Petersen et al. 1988; Gabrieli et al. 1998), which are of course the key components of these tasks. Yet it should be noted that the prefrontal cortex often suffers from signal drop-out, which may have masked more frontal yet potentially relevant regions from our analysis. Importantly, the correlations were specific to the free-behaviors, and were not evident between the deterministic externally driven control events and resting-state dynamics (see Fig. 7). Thus, this effect could not be attributed to general individual differences in the BOLD responses, or to verbal-related or language-production responses. Rather, it indicates that the link between resting state and task was specific to voluntary, internally driven events. Together, these results strongly support our hypothesis (shown in Fig. 1A) that the resting-state fluctuations constitute a common neuronal mechanism that drives free behaviors, and that their rising phase constitutes the anticipatory buildup observed prior to the initiation of these behaviors.
Our study strongly supports the notion that different internally generated verbal behaviors, including creative verbal behaviors, rely on a similar neuronal mechanism. Three aspects of the results support this conclusion. First, our behavioral results show significant similarities in the participants’ performance across the three different tasks, including correlations in number of words and verbal ideas produced, as well as common temporal dynamics of generation, shown in Figure 2. Second, the slow anticipatory buildup, both in the BOLD signal as well as in the pupillary response, precedes all voluntary events, but not the control events. Finally, the link between resting-state fluctuations and the anticipatory buildup was evident in the three different tasks as well. Additionally, the proposed role of spontaneous fluctuations in the emergence of creative thoughts and verbal ideas also fits nicely with previous reports of links between resting-state activity patterns and creative abilities (Takeuchi et al. 2012; Beaty et al. 2014; Yin et al. 2015; Beaty et al. 2018; Shi et al. 2019; Sun et al. 2019), as well as changes in resting-state connectivity patterns following creative or divergent-thinking training (Wei et al. 2014; Fink et al. 2018).
According to the twofold model of creativity, the process of creative thinking includes a generation step and an evaluation step (Chrysikou 2018; Kleinmintz et al. 2019), the latter being specifically shown to involve the LH IFG (Kleinmintz et al. 2018). Here, we suggest that these processes involve accumulation of stochastic noise in the specific task-relevant regions, in our case the language-related ROI. We propose that the spontaneous, or resting-state fluctuations, constitute a stochastic exploration process, which may tilt the system towards a specific idea, exemplar or creative thought at a random moment in time. Within the context of the twofold model the decision component may map onto the evaluation system. The stochastic search, and the verbal exemplars or creative ideas that are generated, depend on previous experience and learning, the underlying brain structure and connectivity, etc.
Given the importance of creative behavior to human progress and achievements, our findings, pointing to resting-state fluctuations as a candidate driving mechanism for creativity, are particularly significant. It is perhaps no coincidence that many anecdotal reports of creative ideas and problem solving in day-to-day life, as well as stories of great ideas or inventions, occur while taking a walk (Nikola Tesla and the AC motor), in the shower or bath (Archimedes’ principles of density and buoyancy) or while daydreaming in public transport (J.K. Rowling and Harry Potter). These are all situations that reduce attention to specific tasks or external stimuli, and hence likely enhance resting-state fluctuations. As our results suggest that these intrinsic fluctuations are involved in driving idea generation, it is an intriguing question whether this type of behavior will be enhanced in contexts in which these fluctuations flourish. Thus, our present findings may open future informed directions for identifying the optimal conditions and even developing methods for enhancing human creativity.
Notes
The authors wish to thank Dr Edna Furman-Haran, Fanny Attar and Eiska Tegareh at the Weizmann MRI center for their assistance.
Conflict of Interest
The authors declare no conflict of interests.
Funding
Joy Ventures grant and CIFAR Tanenboum Fellowship (to R.M.).
