Human brain network control of creativity

Human brain network control of creativity | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results Human brain network control of creativity View ORCID Profile Rikki Rabinovich , View ORCID Profile Rhiannon L. Cowan , Mark Libowitz , View ORCID Profile Tyler S. Davis , View ORCID Profile Shervin Rahimpour , View ORCID Profile Roger E. Beaty , View ORCID Profile Eleonora Bartoli , View ORCID Profile Sameer A. Sheth , View ORCID Profile Elliot J. Smith , View ORCID Profile Ben Shofty doi: https://doi.org/10.1101/2025.06.03.657648 Rikki Rabinovich 1 Department of Neurosurgery, University of Utah , Salt Lake City, UT, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Rikki Rabinovich Rhiannon L. Cowan 1 Department of Neurosurgery, University of Utah , Salt Lake City, UT, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Rhiannon L. Cowan Mark Libowitz 1 Department of Neurosurgery, University of Utah , Salt Lake City, UT, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Tyler S. Davis 1 Department of Neurosurgery, University of Utah , Salt Lake City, UT, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Tyler S. Davis Shervin Rahimpour 1 Department of Neurosurgery, University of Utah , Salt Lake City, UT, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Shervin Rahimpour Roger E. Beaty 2 Department of Psychology, Pennsylvania State University , University Park, PA, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Roger E. Beaty Eleonora Bartoli 3 Department of Neurosurgery, Baylor College of Medicine , Houston, TX, USA 4 Department of Electrical and Computer Engineering, Rice University , Houston, TX, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Eleonora Bartoli Sameer A. Sheth 3 Department of Neurosurgery, Baylor College of Medicine , Houston, TX, USA 4 Department of Electrical and Computer Engineering, Rice University , Houston, TX, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Sameer A. Sheth Elliot J. Smith 1 Department of Neurosurgery, University of Utah , Salt Lake City, UT, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Elliot J. Smith Ben Shofty 1 Department of Neurosurgery, University of Utah , Salt Lake City, UT, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Ben Shofty For correspondence: ben.shofty{at}hsc.utah.edu Abstract Full Text Info/History Metrics Preview PDF Summary The ability to think creatively is fundamental to human cognition, shaping how we interact with the world and navigate complex problems. Creativity depends on the interaction of multiple large-scale brain networks, but how these networks represent creative thought and distinguish it from other cognitive states remains unclear. Here, we use invasive intracranial recordings in the human brain to explore the network-level representations underlying creative thinking. Our results demonstrate that cortical networks in the human brain differentially encode distinct cognitive states: we found unique brain states underlying creative thinking vs. arithmetic calculations, particularly in the default mode network. Further, in the dorsal attention network, we uncovered nonlinear, high-dimensional representations of moment-by-moment fluctuations in creative performance. These representations define a neural creativity axis shared between network-level and single-neuron computations. Our findings reveal widespread neural representations of cognitive state and suggest distinct roles of specific cortical networks in controlling creativity, with the default mode network gating access to creative states and the dorsal attention network regulating the quality of creative output. Introduction Consider your pet dog, or the squirrel it chases up a tree. There is no doubt they each exhibit behaviors that suggest sensation, responses to salient stimuli, and perhaps emotions such as excitement or fear. These “simple” behavioral and cognitive processes are easy to ascribe to many animals; yet, as we watch our pets, we may wonder about their thoughts and sensations that may not obviously manifest in observable behaviors. These questions are intriguing because it is this self-aware, internally-directed thought that seems so key to consciousness. Among forms of internal mentation, creativity is unique in that it involves a fluid, multimodal interaction between introspection and the external world (e.g., generating novel solutions to existing problems). While creativity describes a multifaceted set of processes, here, we focus on divergent thinking, or the mode of thought that spontaneously generates multiple diverse ideas 1 , 2 . Hereafter, we use the two terms (creativity and divergent thinking) interchangeably. As a vital component of human experience and essential to societal progress, creativity has received considerable attention from the scientific community; yet, despite numerous attempts to characterize this cognitive process, the neural underpinnings of creative thinking remain elusive. So far, prior research has led to a consensus that no single brain region is responsible for divergent thinking; rather, interactions within and between large-scale cortical networks linking widespread brain areas likely underlie creative ideation and related cognitive processes 3 – 7 . One such network, known as the default mode network (DMN), is thought to play a role in self-referential or internally-directed cognition. Originally conceptualized as the “default mode” of the brain, this network is active during periods of rest and is often suppressed by task engagement and external attention 8 . Since then, it has been implicated in a range of cognitive processes, from episodic memory retrieval to dreams and dream recall, as well as creative thinking 9 , 10 . Specifically, existing literature links DMN activity to the generation of novel ideas, though the contribution of DMN to creativity likely depends on interaction with cognitive control networks, which may regulate and filter creative output to ensure generated ideas are both original and useful 11 – 15 . However, conventional fMRI approaches lack the temporal resolution to reveal how creative thoughts unfold over time; as a result, most prior studies focus on individual differences in creativity or static measures of network engagement, without resolving the neural dynamics that define creative states within individuals 9 , 16 – 18 . Further, while one causal study found that disrupting the DMN did not affect the quality or originality of responses—only the number generated 11 —another found that stimulating the DMN did impact originality 19 . This discrepancy underscores the complexity of creative ideation and reveals the lack of a unifying, mechanistic model for divergent thinking. Thus, a fundamental question remains about what distinguishes creative thought in the brain and how neural dynamics support it. The present study employs invasive recordings in the human brain to examine neural representations of creativity at multiple scales, ranging from single-neuron activity to large-scale brain networks, with the goal of disentangling the neural basis of creative ideation. In line with prior work, we hypothesized that DMN activity patterns would contain distinct representations of the creative state as well as encode the quality of creative output. We predicted that other networks may become engaged at various points in the task but would only weakly encode creativity. Finally, we hypothesized that creativity representations could be identified at the single neuron level. While our data supported some of our hypotheses, our results painted a much more complex picture of the network-level encoding of creative thought. Results To probe the role of the DMN and other networks in divergent thinking, we leveraged intracranial electrodes implanted in the brains of 15 neurosurgical patients undergoing monitoring for drug-resistant epilepsy; from these electrodes, we directly recorded neural signals in the human brain while subjects either engaged in creative thinking or simple arithmetic. To distill creative ideation into a simple process that can be analyzed and manipulated, we apply a version of the Alternate Uses Task (AUT). This paradigm is a widely used method for probing divergent thinking and requires participants to invent creative uses for everyday objects 2 , 3 , 16 , 17 . A basic arithmetic task (BAT), in which participants solved simple addition or subtraction problems, served as a contrast ( Fig. 1a ). We chose BAT as our comparison task based on precedents in prior research where arithmetic calculations were used as control conditions for verbal tasks and because of shared semantic control processes across both verbal and numerical domains 20 – 23 . Since both tasks had identical structure (silent stimulus window followed by spoken verbal response), we were able to ensure that the only distinction between the tasks was the cognitive component. Download figure Open in new tab Figure 1. Experimental design. a) Left: task design. Trials occur in blocks of three, with instructions appearing on the screen at the start of each block. For each trial, the prompt (in a colored font) is present for a 10 second stimulus window, after which the font turns white, indicating the beginning of the 5 second response window. Right: stylized example responses. For the prompt “cup”, possible high-originality uses include “build a sandcastle” (semantic distance score of 0.909), “turn it into a planter” (0.930), or “use as a doorstop” (0.967). b) Patients’ behavioral responses across all trials (n=15 patients; 316 trials with valid responses). Each dot represents a single trial’s response; distance from center is the semantic distance score for that response. Each cluster of same-color dots corresponds to 1 patient. c) Signals are recorded from implanted electrodes during task execution. The model brain shows locations of all contacts. Contacts within the DMN are colored orange. In our version of the AUT ( Fig. 1a ), a cue word (e.g., “cup”) appeared on a screen: during the 10-second stimulus window, subjects silently thought of possible uses for the item. Then, a change of font color prompted subjects to choose a single “most creative” use for the object and verbalize their response. Responses were automatically scored for originality using natural language processing tools (semantic distance, Fig. 1b ). For the prompt “cup”, the response “drink tea” scores low on originality (semantic distance=0.504), while “build a sandcastle” scores high (0.943), as it is semantically far from the meaning of the cue word. Patients exhibited satisfactory task performance; each subject’s set of responses spanned a range of semantic distance scores, indicating trial-to-trial variability in creativity level ( Fig. 1b ; mean originality=0.716; standard deviation=0.123); meanwhile, the distribution of scores was similar across patients, with no outlier subjects. Among the datasets presented here, we never observed “nonsensical” responses, in which the answer was completely unrelated to the prompt word. During task execution, we recorded LFP and single-unit responses from intracranial depth electrodes implanted throughout various brain regions and cortical networks ( Fig. 1c ). Networks sampled were default mode network (DMN), frontoparietal network (FPN), somatomotor network (SMN), limbic network, visual network, dorsal attention network (DAN), and ventral attention networks (VAN). Network dynamics reflect cognitive state First, we examined the temporal dynamics of DMN engagement in the creativity task as well as the control task ( Fig. 2a , Supp. Fig. 1). We found that creativity trials displayed increased high-frequency and decreased low-frequency LFP power. Specifically, during AUT, high-frequency broadband activity (70-140Hz, also referred to as high gamma) was elevated, exhibiting the greatest increase during the response window. Meanwhile, no DMN high gamma modulation occurred in BAT (or, if anything, it slightly decreased). Theta power (4-8Hz), by contrast, decreased in DMN during the stimulus period of the creativity task (whereas BAT was characterized by increased theta power throughout the stimulus presentation). At the end of the stimulus, AUT theta returned to baseline, while BAT theta remained elevated. Upon entering the response window, theta power declined in both trial types, nearing baseline levels for BAT and falling below baseline for AUT. Since high gamma activity reflects local neuronal population firing 24 – 28 (while low-frequency LFP activity corresponds to network disengagement), these results suggest that DMN is recruited during divergent thinking, reproducing and expanding our previous findings on DMN power changes. Download figure Open in new tab Figure 2. Default mode network representations distinguish creativity from arithmetic trials. a) Percent change from baseline of LFP power in high gamma (70-140Hz, left) and theta (4-8Hz, right) frequency bands over the time course of an average AUT (orange) and BAT (blue) trial. Difference between AUT and BAT before stimulus onset (time 0) is due to bleed-over from the previous trial. Shading indicates standard error. n=15 patients. b) Linear SVM cognitive state decoding accuracy for AUT ( t =10.3; p =6.3 × 10 -8 ; two-sided 1-sample t -test) vs. BAT ( t =5.7; p =5.9 × 10 -5 ) trials. Each dot is the decoding performance for a single subject. c) Visualization of trial-type separation, based on low-dimensional embeddings from the training and testing set of 1 cross-validation fold. Left: embeddings generated by CEBRA models trained on ¾ of the trials for each patient to separate AUT (orange) and BAT (blue) trials. Right: Held-out data transformed into the same latent space. d) Trial type decoding performance using linear SVM for each network. e) Dimensionality reduction using CEBRA trained to separate activity of different networks. Left: latent embeddings reveal network clusters. Each dot is 1 recording contact at 1 timepoint in a trial (16 total points per contact, because 16-second trial duration). Colors correspond to different networks. Right: pairwise centroid distances quantify distance in latent space between each pair of networks. Lighter (yellow) colors indicate greater difference; darker (blue) colors signify similarity between networks. Note that DMN is on average most different from the other networks, while visual network, DAN, and VAN are very similar to each other. f) Latent embeddings generated by self-supervised CEBRA models for DMN (time-contrastive training; no behavior labels). Each dot is 1 timepoint. Embeddings are color-coded by trial type: AUT=orange; BAT=blue. g) KNN decoding of trial type from embeddings in (f) and Supp. Fig. 4, holding out 1 patient at a time when training CEBRA models, for cross-validation. Each dot is 1 patient. i) Projection of DMN embedding for two example patients onto a cognitive state coding axis defined by linear regression. Colored trace is the projection value; grey trace is the true trial type for each trial. Vertical dashed lines mark trial onsets. To further disentangle the neural patterns related to creative thinking, we applied machine learning methods to investigate how network activity encodes cognitive states (creative vs. arithmetic-related modes of thought, corresponding to AUT vs. BAT trial types). Using a linear support vector machine (SVM) classifier to decode cognitive state from the electrophysiology data, we were able to successfully distinguish trials involving divergent thinking vs. mathematical reasoning on a trial-by-trial basis from DMN ( Fig. 2b,c ) and several other cortical networks—specifically the frontoparietal, limbic, and ventral attention networks; somatomotor, visual, and dorsal attention networks were not significantly predictive of cognitive state ( Fig. 2d ; Supp. Fig. 2a; DMN: t (14) = 9.2, p =1.9×10 - 6 ; FPN: t (14) =6.8, p =2.8×10 -5 ; SMN: t (9) = 2.3, p =0.055; limbic: t (11) =5.3, p =0.00055; visual: t (8) =2.5, p =0.055; DAN: t (7) =0.97, p =0.36; VAN: t (8) =3.1, p =0.027; two-sided 1-sample t -tests; FDR corrected for multiple comparisons). To control for possible contamination from speech, we repeated this analysis by decoding only from the timepoints within stimulus window (during which subjects were silent). Decoding performance within the stimulus window was similar to that of the entire trial, except for the visual network, for which decoding performance crossed the significance threshold (DMN: t (14) =7.6, p =1.8×10 -5 ; FPN: t (14) =4.7, p =0.0013; SMN: t (9) =1.0, p =0.38; limbic: t (11) =4.7, p =0.0015; visual: t (8) =3.8, p =0.0091; DAN: t (7) =0.66, p =0.53; VAN: t (8) = 2.6, p =0.041). The accurate cognitive state decoding suggests that divergent thinking depends on a unique network-wide pattern of neural activity and that distinct “brain states” underlie various modes of thought. Although multiple brain networks were predictive of cognitive state, DMN was the most predictive: classifiers trained on DMN activity significantly outperformed those based on somatomotor, limbic, visual, dorsal attention, and ventral attention networks ( Fig. 2d ; Supp. Fig. 2a; compared to SMN: p =0.0085; limbic: p =0.012; visual: p =0.012; DAN: p =0.0083; VAN: p =0.0085; two-sided Mann Whitney U test). DMN-based decoding also trended higher than frontoparietal-based decoding, though the difference was not significant ( p =0.30). Note that variations in number of contacts between networks do not explain this difference: we observed no correlation between number of contacts and decoding performance in any given network (DMN: 8-29 contacts, p =0.85; FPN: 2-17 contacts, p =0.12; SMN: 1-7 contacts, p =0.80; limbic: 1-9 contacts, p =0.85; visual: 1-16 contacts, p =0.85; DAN: 2-9 contacts, p=0.85; VAN: 1-9 contacts, p =0.40; two-sided Pearson correlation). After decoding cognitive state from network LFP activity, we inspected the classifier’s coefficient weights in order to identify which DMN subregions were most predictive of cognitive state: the features with the largest weights contribute the most to the classification. In 7 of the 15 datasets, the most predictive region was middle temporal gyrus (5 left; 2 right hemisphere). Superior frontal gyrus was most predictive in 3 datasets (2 left; 1 right), and superior temporal gyrus (left), inferior temporal gyrus (right), angular gyrus (right), posterior cingulate gyrus (right), and hippocampus (left) were each most predictive in 1 dataset. To further interrogate the difference between cortical networks’ representations of cognitive state, we applied CEBRA, the recently developed nonlinear neural network encoder and dimensionality reduction tool (CEBRA 29 ). Clustering network activity during only the AUT trials revealed a separation between DMN and other networks ( Fig. 2e ; Supp. Fig. 3), supporting the hypothesis that DMN has a unique role in creative thinking. To corroborate the SVM decoding result and to visualize the separation between creative and arithmetic-related cognitive states, we employed CEBRA to extract the latent space that describes each dataset. We trained separate CEBRA models on the time series data for each patient, using trial type (AUT vs. BAT) as a supervisory signal (i.e., every timepoint for every trial was labeled as belonging to an AUT or BAT trial). We then used the model to transform the data, generating embeddings in a low-dimensional latent space. Importantly, to ensure proper cross-validation, we trained each model only on subsets of the trials (training set) and performed the same transformation on the training data and the held-out test data. When aligned between patients, the resulting manifolds clearly separated creativity from arithmetic states ( Fig. 2c , left); moreover, embeddings from a held-out validation dataset displayed a similar pattern ( Fig. 2c , right). We quantified this separation using a K-nearest neighbors classifier to predict cognitive state based on the embeddings corresponding to each test subset (the held-out data). Decoding performance was similar to that produced by the SVM, implying that cognitive state is linearly-separable at the network level (Supp. Fig. 2b). Our ability to decode trial-by-trial cognitive state from network activity, combined with the correlational analysis of LFP power that showed DMN engagement during AUT, reveals strong network-level representations of cognitive state, particularly in DMN. Cognitive state is an intrinsic component of DMN activity In our decoding approach above, the CEBRA models were able to distinguish creativity vs. arithmetic trials when specifically trained to do so using trial type labels (AUT vs. BAT) supplied to the model. However, if cognitive state is a fundamental property of brain network activity patterns, it might be possible to predict trial type without ever providing the labels to the model. Hence, we asked whether an unsupervised or self-supervised model could still discriminate between creativity and arithmetic cognitive states. To this end, we trained self-supervised CEBRA models using only time as a label, and no information about cognitive state, on the full dataset for each patient. To visualize the patterns related to cognitive state for each cortical network, we aligned across subjects and color-coded the latent space by trial type ( Fig 2f ; Supp. Fig. 4), revealing a separation between creativity and arithmetic trials that was qualitatively strongest in DMN. Subsequently, we attempted to predict cognitive state from these embeddings. Holding out one patient at a time, we again fit a K-nearest neighbors classifier to the remaining embeddings (training set) and asked it to classify trials for the held-out subject. Decoding performance was lower compared to that of the supervised models, but still significantly above chance for all networks except ventral attention network ( Fig. 2g ; median decoding performance for DMN: 67.29%, t =7.13, p =5.12×10 -6 ; FPN: 56.04%, t =4.61, p =0.0009; SMN: 53.95%, t =2.93, p =0.02; limbic: 54.17%, t =3.44, p =0.006; visual: 52.72%, t =3.77, p =0.006; DAN: 55.83%, t =3.95, p =0.006; VAN: 50.83%, t =0.73, p =0.76; two-sided 1-sample t -test; FDR corrected). Furthermore, decoding performance was strongest for DMN (compared to FPN: p =0.003; SMN: p =0.0009; limbic: p =0.0005; visual: p =0.001; DAN: p =0.003; VAN: p =0.0005; two-sided Mann Whitney U test; FDR corrected). Moreover, if we reduce the data even further to a single dimension along the coding axis (by projecting the data onto an axis defined by the weights of a linear regression model; see Methods), we find that its moment-by-moment trajectory tracks with the ground truth regarding creative vs. mathematical cognitive mode ( Fig. 2i ). We remind the reader that this predictive latent space was generated by a self-supervised model, which had no knowledge of trial type information or behavior. Consequently, we can conclude that the representations of cognitive state are embedded in the intrinsic geometry of network activity, especially in DMN. Nonlinear network representations encode creativity level We uncovered a relationship between neural activity patterns and creativity levels, as measured by the semantic distance between the prompt and the response for each trial. Specifically, we observed increased broadband high-frequency activity (high gamma; Fig. 3a ) across DMN nodes in high-creative compared to low-creative states during the stimulus period of the AUT ( t =1.78; p =0.038; one-sided paired t -test). The frontoparietal network also showed this pattern, but none of the other networks did (FPN: p =0.009; SMN: p =0.142; limbic: p =0.538; visual: p =0.296; DAN: p =0.356; VAN: p= 0.158; one-sided paired t -tests). DMN low frequency (theta) power was lower on average for highly creative responses (Supp. Fig. 6a), but the difference did not reach the significance threshold ( p =0.072). This finding of increased high gamma power in the most creative trials reveals state-specific DMN recruitment during highly creative states. Download figure Open in new tab Figure 3. Dorsal attention network activity predicts performance on creativity trials through nonlinear representations. a) Difference in DMN network states between high vs low creativity: average high gamma LFP power (quantified as % change from baseline) of the ⅓ least creative trials is subtracted from the power of the ⅓ most creative trials. Outward pointing bars indicate a positive difference (i.e., greater power for high-originality trials); inward pointing bars correspond to negative difference. Scale-bar indicates magnitude of the difference. n =268 recording sites. b) Example of low-dimensional embeddings from the training and testing set of 1 cross-validation fold. Left: embeddings generated by CEBRA models trained on ¾ of the data for each patient to distinguish creativity levels. Right: Held out data transformed into the same latent space. Higher originality scores are pink; lower scores are blue. Each dot is 1 timepoint. c) Creativity decoding performance for each network based on the latent embeddings (average across 4 folds). Originality was discretized into 10 bins, so chance performance is 10%. d) Confusion matrix showing decoding performance based on DMN embeddings. Color indicates the proportion of trials for which a creativity score on the y-axis is predicted to be the value on the x-axis. e) Latent embeddings for DMN based on a CEBRA model trained on all data from all subjects. Black line is the coding axis determined by linear regression. Coding axis goodness-of-fit: R 2 =0.86. f) Projection of all datapoints of the latent embedding onto the coding axis shown in (e) (x-axis) compared to true normalized originality scores (y-axis). Y-axis for (f) is the same as the colorbar axis for (e). Pearson correlation r =0.93. g) same as (d), but for dorsal attention network. h) same as (e), but for dorsal attention network. R 2 =0.73. i) same as (f), but for dorsal attention network. r =0.86. However, originality is not easily divided into two classes (high vs. low originality); rather, subjects’ responses comprise a distribution of semantic distance scores ( Fig. 1b ). Thus, creativity should be considered a continuous or multi-class metric. While a linear SVM was sufficient to decode creative vs. mathematical cognitive state with high accuracy, we were not able to decode levels of creativity with this methodology. Neither 2-class (high vs. low creativity) nor multi-class (creativity scores discretized into 10 bins) SVMs were able to accurately decode creativity scores, suggesting that any creativity representations, if present, are likely encoded nonlinearly. Indeed, using a nonlinear dimensionality reduction approach through CEBRA, we were able to identify creativity level-related patterns in latent embeddings of the neural activity. To use this tool for decoding analyses, we trained a CEBRA model on subsets of LFP data (filtered into discrete frequency bands) from each subject, using discretized (binned and normalized) creativity score as a supervisory signal, and then applied the same transformation to the training set and held-out data to generate low dimensional embeddings ( Fig. 3b ). We then used the embeddings from the training set to decode discretized creativity level. We observed above-chance decoding performance for all sampled networks, while models trained on data with shuffled labels performed at chance level ( Fig. 3c , Supp. Fig. 6b). However, unlike the cognitive state representations, for which DMN was the most predictive network, creativity level encoding was strongest in the dorsal attention network ( Fig. 3c,d ,g), with the dorsal attention network being significantly more predictive than DMN (p=0.02; two-sided Mann Whitney U test). The same pattern was observed when including timepoints from the entire trial as when only including the stimulus period; in fact, the stimulus-only embeddings yielded slightly higher decoding performance across all networks (Supp. Fig. 6b). Given our ability to decode creativity level from neural activity, we asked whether the representations could be mapped onto a single dimension serving as a creativity axis (akin to coding axes identified for other cognitive or behavioral properties 30 – 32 ) . We generated new latent embeddings ( Fig. 3e,h ; Supp. Fig. 5), this time including data from all participants in one model for each brain network (treated by CEBRA as a “pseudo-subject”). In this analysis, continuous originality scores served as the labels for supervised learning. We then defined a coding axis via linear regression across embeddings ( Fig. 3e,h ; DMN: R 2 =0.77; DAN: R 2 =0.66). To confirm that this axis robustly described creativity performance, we projected the latent manifold onto this axis, producing a “creativity projection” value of the neural activity at each timepoint in each session. This projection correlated well with true originality scores ( Fig. 3f,i ; DMN r =0.87; DAN r =0.79 Pearson correlation), corroborating the presence of a creativity axis in the human brain while providing an instantaneous readout of a patient’s ongoing level of creativity. Note that in this case, the relationship between DMN and creativity level appears stronger than that of dorsal attention network, an observation that holds even if we use normalized creativity values (see Methods for scores) and when we only compare data from the same subset of subjects (DMN subset: R 2 =0.75, r =0.87; DAN subset: R 2 =0.66, r =0.79). This dichotomy—better decoding from dorsal attention network but a more explanatory coding axis in DMN—could arise due the difference in the form of the supervisory signal: we used binned labels for classification analysis, but continuous labels to discern a coding axis. While both networks’ representations of creativity are nonlinear, having already gone through a nonlinear dimensionality reduction, we hypothesized that the dorsal attention network may be encoding creativity with a higher degree of nonlinearity, making it more suitable for discrete classification but less easily described by a linear regression model. To investigate the nonlinearity of creativity coding, we fit polynomial regression models on the embeddings and normalized continuous originality scores. Compared to a linear model, a quadratic polynomial model described both networks’ representations of creativity better than the linear models did, with DMN models showing a more pronounced improvement. As we further increased the degree of the polynomial of the DMN models, the goodness of fit plateaued (i.e., greatest increase was between linear and quadratic). In contrast, DAN models showed the opposite pattern, with only small improvements in goodness of fit at first, and then increasingly larger improvements as the polynomial degree grew, overtaking the DMN model after a polynomial degree of 8 (Supp. Fig. 7a). This finding confirms the higher nonlinearity of the DAN creativity representation. Finally, we considered the dimensionality of the representations. Simply performing extreme dimensionality reduction (using CEBRA to reduce the data down to a single latent dimension) could not identify a coding axis: cross-validated decoding performance based on a 1-dimensional latent space was at chance level for both networks (Supp. Fig. 7b). Thus, the coding axis we described above had to be embedded in a higher-dimensional latent space. Further, as the number of latent dimensions increases from 1 to 5, the difference between decoding based on DAN vs DMN widens (and then plateaus after 5 dimensions), suggesting that DAN representations of creativity are more high-dimensional. Together, these findings expose brain-wide network-level creativity representations that are encoded along a nonlinear coding axis in a high-dimensional latent space. Further, these results hint at a role of the dorsal attention network in modulating the quality of individual subjects’ creative output through a highly nonlinear representation of creativity. Cognitive state is encoded at single unit level Given the large-scale, network-level representations of cognitive state we uncovered, we next asked whether individual cells represent these distinct cognitive states. In a subset of patients (10), we recorded single-unit activity in addition to LFP responses. Indeed, we found cells in multiple brain regions that were selective for trial type (creativity vs. math cognitive states): some exhibited increased firing for AUT trials ( Fig. 4a,b ) while others preferred BAT trials ( Fig. 4c,d ). Of the 127 cells we recorded, 41% significantly encoded cognitive state ( Fig. 4e ; two-sided paired t -tests; FDR corrected). Of these selective cells, 43.9% were more responsive for AUT trials. Download figure Open in new tab Figure 4. Trial type and creativity level are encoded at a single unit level. a-d) Four example units, recorded from (a) medial frontal cortex, (b) gyrus rectus, (c) middle cingulate gyrus, and (d) anterior cingulate gyrus. For each panel, Left: All waveforms of the identified unit. Light grey lines are individual waveforms; thicker black line is average. Right: PSTH (top) and raster plot (bottom) showing each cell’s responses to AUT (orange) and BAT (blue). For raster plot, each row is a trial, and each mark is a spike. Stimulus onset is at time 0. e) Proportions of cells identified as having a trial type preference: 23% of cells are selective for BAT trials; 18% prefer AUT trials, based on paired t-test across time bins; FDR corrected. n=124 cells. f) Example of latent embeddings from the training and testing set of 1 cross-validation fold. Left: embeddings generated by CEBRA models trained on ¾ of the single unit data for each patient to distinguish trial type. Right: Held out data transformed into the same latent space. g) Decoding trial type based on latent embeddings, for AUT and BAT (n=10 patients). h) Confusion matrix showing creativity level decoding performance based on single unit embeddings. Color indicates the proportion of trials for which a creativity score on the y-axis is predicted to be the value on the x-axis. p=0.007. i) Projection of single unit (black) and DMN (orange) data onto the creativity coding axis, for 2 representative patients. Vertical dashed lines indicate start of each trial. Y-axis is the projected creativity value. j) CCA between latent embeddings based on DMN and single unit data. Plot shows the first CCA component for the two data types ( r =0.73). Color gradient indicates creativity value ( r =0.88 for DMN and r =0.70 for single units). Selective cells were identified in the following regions: medial temporal lobe (N=47, 29.29% selective), ventromedial prefrontal cortex (N=37; 40.54% selective), dorsomedial prefrontal cortex (N=15; 53.33%), middle cingulate cortex (N=20; 45%), and posterior cingulate cortex (N=8; 50%). Based on anatomical location, we classified recorded units as belonging to DMN (N=24, 54.17% selective), limbic network (N=80; 35%), frontoparietal network (N=14; 42.86%), and ventral attention network (N=9; 33.33%). Further, the neuronal population’s representations of different cognitive states were separable ( Fig. 4f ). We trained CEBRA models on the single-unit responses for each patient, and from the resulting embeddings, we were able to distinguish AUT vs BAT trials with ∼65% decoding accuracy ( Fig. 4g ). Thus, the single-neuron responses among the population of recorded cells reflect engagement in creative vs. mathematical thinking. Single unit and network activity encode creativity similarly Finally, we investigated the relationship between creativity representations in the network-level LFP activity and the single-unit spiking patterns. As with LFP data, we used CEBRA to identify the single unit representations of creativity for each patient (holding out data from CEBRA training with 5-fold cross-validation). We found that not only cognitive state but also creativity level was encoded in the responses of the recorded neuronal population. Classification performance on single unit responses was lower than on network-level LFP activity—which is to be expected given the low number of units (features on which the model is trained) in each patient. Nonetheless, creativity scores could still be decoded above chance from embeddings based on the single unit population ( Fig. 4h ; p =0.007). We then computed the creativity coding axis using linear regression on the single unit spiking data. Then we projected the data onto this coding axis and compared the result to the LFP creativity projection. We observed close alignment between the two projections ( Fig. 4i ), suggesting shared representations between these two data sources. To quantify this relationship further, we employed canonical correlation analysis (CCA) on the creativity representations (low-dimensional embeddings generated by CEBRA) corresponding to single unit and DMN LFP activity. We found a strong correlation between the LFP and single unit embeddings ( Fig. 4j ; first CCA component correlation between spiking embeddings and DMN embeddings: r =0.73; FPN: r =0.70; SMN: r =0.65; limbic: r =0.58; visual: r =0.68; DAN: r =0.69; and VAN: r =0.74), demonstrating that LFP and spiking activity share a dominant latent dimension in the creativity embedding space and thus revealing the close relationship between the single unit spiking and network-level representations of creativity. Both LFP and single unit embeddings displayed high correlation between their first CCA component and creativity scores ( r =0.88 for DMN and r =0.70 for single units). The higher correlation of LFP embeddings with creativity suggests that this representation describes more of the variance in creativity performance than do single unit embeddings, a difference that may arise from the more widespread coverage of DMN by LFP data compared to limited and local single unit recordings. Nevertheless, we show shared creativity representations between single neuron firing and large-scale brain networks. Discussion Despite growing interest in the neuroscience of creativity, the mechanisms through which neural dynamics give rise to spontaneous thoughts and novel ideas remain mysterious. Our investigation of a neural basis for creativity unearthed widespread encoding of cognitive state and originality of creative thought. Indeed, we found that unique brain states distinguished creative thinking and arithmetic calculations, allowing us to disambiguate these cognitive states from the activity of several brain networks, including DMN, frontoparietal, limbic, and ventral attention networks. Of these networks, we identified DMN as the most linearly predictive of cognitive state. Not only was DMN predictive of cognitive state (creativity vs. arithmetic trial type), but also its activity structure was intrinsically organized to reflect this feature: a self-supervised dimensionality reduction approach revealed patterns of activity that distinguished creativity from mathematical thinking. This separation of cognitive state was much clearer in DMN than in any other network. Importantly, this finding is not inevitable: just because the trial type information can be extracted from DMN activity does not imply that this variable is a fundamental property of DMN activity, and we did not initially expect to be able to distinguish modes of thought using a model with no supervisory signal or behavioral information. The striking observation that a self-supervised model spontaneously separated creativity and arithmetic states indicates that cognitive state is inherent to the structure of DMN activity (i.e., intrinsically embedded in its representational geometry). While the creative state (compared to math) was most strongly encoded by DMN, it was the dorsal attention network that contributed most to varying levels of creative output. This finding was especially surprising given that the dorsal attention network was the least predictive of cognitive state. We therefore speculate that the DMN plays a role in “ gating ” creative thought, while the dorsal attention network is more involved in “ regulating ” levels of originality (as assessed by the alternate uses task), albeit with a highly nonlinear representation. Creative ideation is thought to involve two parallel processes: idea generation and evaluation. Both processes likely occur during our divergent thinking task, though our behavioral design does not dissociate between them. Due to its involvement in episodic memory and self-referential thinking 10 , 33 , 34 , DMN is believed to play a stronger role in idea generation, during which thoughts arise spontaneously or through associative thinking 35 , while networks related to executive control and goal-directed thought contribute more to idea evaluation 15 . Accordingly, a gating DMN may help a participant think of ideas, but they may not necessarily be good ideas; involvement of the dorsal attention network may then introduce top-down modulatory constraints to regulate the ideas and select the optimal choice, thereby controlling response originality. Since we only have access to the final spoken response, the involvement of the dorsal attention network in selecting that response would explain why it most strongly predicts the quality of the final output. This framework, with DMN gating divergent thinking, is also supported by previous research in which disruption of lateral DMN nodes through high-frequency stimulation decreased the number of ideas generated (a metric known as “creative fluency”) but did not alter originality scores 11 . In line with this evidence, our more recent work 19 revealed that DMN recruitment during AUT (as measured by increases in gamma power) was driven by activation in lateral DMN nodes. Crucially, stimulation to medial DMN nodes interfered with originality scores, but not fluency, suggesting the existence of anatomo-functional sub-specializations of gating within the DMN, for quantity and quality of creative output. In our experimental cohort, we had coverage of both DMN subsystems, but the regions most predictive of cognitive state were predominantly localized in lateral DMN nodes: of the 15 patients, lateral DMN regions were most predictive of cognitive state in 10 subjects (7 middle temporal gyrus, 1 superior temporal gyrus, 1 inferior temporal gyrus, and 1 angular gyrus), while dorsal and medial regions were more predictive in only 5 subjects (3 superior frontal gyrus, 1 posterior cingulate gyrus, and 1 hippocampus). Consequently, if we account for the variations in anatomical targeting between previous studies, our discovery of a creativity-gating role of lateral DMN is consistent with previous findings. As an aside, variations in task design may also explain the diversity of results across these studies. Consider the task differences from the generation-evaluation framework: when given enough time to engage in mind-wandering and permission to verbalize each idea (as in our previous AUT experiments), idea generation is likely to be the dominant mode of thought 19 . In contrast, a more engaging task with a shorter stimulus window and instructions to only report a single response (as in the current task design) may increase reliance on the idea evaluation subprocess. Thus, while our current paradigm places participants in the same creativity-requiring context, the mechanisms driving divergent thought are likely distinct, with greater emphasis on executive control and top-down decision making. Besides the inter-network coding differences, we also discovered a brain-wide creativity-coding axis. The projection of creativity-related manifolds onto this axis (i.e., the moment-by-moment representation of creativity level) correlated well with true originality scores. Thus, despite the multidimensional geometry of the latent space (corresponding to even higher-dimensional neural data), a discrete coding axis can be isolated. Furthermore, this axis is preserved across scales, from single units to distributed networks, suggesting that creativity is a hierarchically-organized, multiscale process. When we projected all datapoints in each network’s latent space onto the creativity-coding axis, the correlation of the resulting projection with true creativity values was highest for DMN compared to the other networks, including dorsal attention network. Similarly, the goodness of fit of the linear regression model was higher for DMN than for dorsal attention network, meaning the creativity coding axis in DMN explained the data better. Yet, earlier, we discussed our observation that dorsal attention network, not DMN, was most predictive of creativity level. Though these analyses both delve into similar questions about network-level representations of creativity level, there is a key difference: the decoding analysis was based on embedding generated from binned creativity data, whereas the coding axis was inferred based on continuous originality scores. We speculate that the dorsal attention network’s representation of creativity is higher-dimensional and more nonlinear. This conjecture is supported by our observation of improved fit of curved coding axes generated by higher-polynomial regression models, and that increasing the degree of the polynomial has a stronger effect on the dorsal attention network than DMN. Thus, it was possible to recover a latent space from DMN that encodes creativity as a continuous spectrum along a near-linear coding axis spanning all creativity levels, meanwhile the same approach revealed that the dorsal attention network’s representation is highly nonlinear. This result is consistent with our proposed framework in which the dorsal attention network regulates creativity through goal-directed attentional control of thought processes: unlike DMN’s simpler and stable relationship with creativity, the dorsal attention network generates a highly complex representation of creative states, which may fluctuate with task demands and cognitive engagement. Finally, our examination of single-unit recordings points to a potential mechanism for creativity encoding at the single-cell level that may underlie network-wide creativity representations. The creativity coding axis we observed in network-wide LFP data was also found in single-unit spiking patterns, and the trajectories of the creativity-trained embeddings based on the LFP and single-unit data closely followed each other. Therefore, the way creativity is represented by the spiking of single units is largely conserved at the level of network activity. We found this relationship between spiking and network representations for each network sampled, likely because we considered all recorded units as a single population and did not split them according to anatomical or functional boundaries. Future work will collect a larger sample of single units and enable more fine-grained analysis. In addition, stimulation experiments to manipulate network activity and synchrony, combined with single unit recordings, will be important to clarify the role played by specific neuronal ensembles in generating creative thought. In the meantime, we conclude that the default mode network maintains a stable representation of cognitive state, while high-dimensional nonlinear brain states across networks flexibly encode creativity level. Although cognitive state and creativity level can be decoded from activity brain-wide, the default mode network likely plays a key role in gating creative thought, and its activity patterns are uniquely shaped by creative ideation. In contrast, our results suggest that the dorsal attention network generates a complex, high-dimensional representation of creativity level that imposes context-dependent constraints and flexibly regulates the quality of creative output. Together, our findings serve to illuminate the neural landscape underlying creative thought. Methods Subjects 16 neurosurgical patients who were undergoing clinical monitoring for epilepsy participated in the study. All subjects provided informed consent prior to participation. Separate informed consent was obtained for implantation of microelectrodes and for participation in cognitive tasks. All experimental procedures were performed according to protocols approved by the University of Utah Institutional Review Board. 1 patient was excluded from analysis due to experiment interruption for clinical care. No patients with major neurophysiological anomalies or cognitive deficits were included. Electrophysiological recordings Depth electrodes were implanted in patients’ brains as part of their epilepsy monitoring, and the location of the implants was dictated by clinical need. For the patients who consented, 2-3 electrodes used were Behnke-Fried hybrid electrodes, which contained both macrocontacts as well as 40um microwires at the end. These microwires splay out during implantation and provide recordings of single unit activity. All other electrode contacts recorded local field potentials (LFP). Recording sites were classified as belonging to a particular region and brain network using the LeGui software package 36 , based on a post-operative CT scan and co-registered structural MRI and normalized to MNI space. Regions were identified using the Neuromorphometrics (NMM) atlas and networks were identified with the Yeo atlas 37 . All patients had coverage in DMN and FPN networks; however, other networks were less frequently sampled. The number of patients with coverage in each networks was as follows: DMN: 15; FPN: 15; SMN: 10; limbic: 12; visual: 9; DAN: 8; VAN: 9. 12 sessions were conducted using a Blackrock Microsystems recording system (128 channels, 1KHz sampling rate); 3 sessions were recorded with a Ripple Grapevine system (128 channels, 2KHz sampling rate, downsampled to 1KHz for analysis). Single unit activity was recorded from 10 subjects. Behavior Behavioral paradigm : All subjects performed two tasks with identical structure: an alternate uses task (AUT) and a basic arithmetic task (BAT). Prior to initiating the task, the experimenter explained the instructions and task structure. Then, subjects were exposed to a “training round” to help them understand the task. The training round cycled between one AUT and one BAT block; once subjects were comfortable with task structure, the training ended and the real task began. The tasks were designed using psychtoolbox. Each trial consisted of a 10 second stimulus window followed by a 5 second response window. Each block of trials contained 3 trials of one type (either 3 AUT trials or 3 BAT trials). AUT and BAT trials were randomly interleaved. At the beginning of each block, instructions were shown for the corresponding trial type. For each trial within a block, a prompt (cue word or arithmetic equation) was presented on a screen for 10 seconds, during which subjects were instructed to silently think of their answer. For AUT, the instructions and cue word were displayed in orange. For BAT, the font is blue. After the 10 second stimulus window, the prompt text changed to white, signaling subjects to verbally respond by either stating a single “most creative” use for the object or providing the answer to the equation. Each session included 15-30 trials of each type. AUT prompts were randomly selected from a list of 100 common items (e.g., spoon, newspaper, hanger) without repetition. For each BAT trial, subjects were prompted to add or subtract two random numbers between 0 and 12 (e.g., 8+5 or 4-12). We chose simple arithmetic to avoid the engagement of creative problem-solving that might be required for more complex mathematic reasoning. Performance scoring : Participants’ responses were collected with a microphone, and transcribed offline. AUT performance was judged based on the originality of each response. Originality was quantified as the semantic distance between the prompt and the response, using the Open Creativity Scoring algorithm 38 , 39 . After removing stopwords (e.g., “the”, “and”, “to”), this natural language processing algorithm uses a semantic embedding model to convert both prompt and response to vectors in semantic space and calculate the cosine of the angle between the vectors. The larger the angle, the more dissimilar the response is from the prompt, and the greater the originality score. Semantic distance primarily reflects the novelty of responses and correlates to human creativity ratings 40 . Analysis of LFP power The Python MNE library was used to analyze LFP data. Data preprocessing included removing noisy channels, notch filtering to remove power line noise (60Hz and 120Hz harmonic), and common average referencing. Data was epoched to only include relevant time windows. Frequency bands were split as follows: theta (4-8Hz), alpha (8-13Hz), beta (13-30Hz), gamma (30-70Hz), and high gamma (70-140Hz). Power spectra were calculated using the Welch method. Baseline LFP power was calculated by taking intervals of 1 second duration prior to each trial block, beginning 2 seconds before the start of the block and averaging all these pre-trial baselines. LFP power was baseline corrected by dividing the signal by this averaged baseline. For our analysis evaluating network states for high vs low creativity, we split the behavior data into thirds, and compared the LFP activity in different frequencies for the highest-scored third vs the lowest-scored third of each patient’s responses. Decoding Linear classification : We used Python scikit-learn software 41 to train a support vector machine (SVM) with a linear kernel to classify trials as belonging to AUT or BAT types, based on LFP power at different frequency bands across the time course of the trial, or based on firing rates of single cells. In all decoding analyses, decoding performance was evaluated with 4-fold cross-validation, with the exception of the self-supervised training, since the models never had access to labels, so there was no need to hold out any data. Nonlinear dimensionality reduction and decoding : We used CEBRA to identify low-dimensional latent embeddings from the high dimensional data 29 . For LFP data, we generated 5-dimensional embeddings. We confirmed that this value was a reasonable choice by generating creativity-related embeddings using all numbers of latent dimensions from 0-10 and inspecting creativity decoding performance (cross-validated) based on these embeddings (Supp. Fig. 7b). Single unit data had lower dimensionality already, so we generated 3-dimensional embeddings instead. We fit independent CEBRA models on the data for each patient, then used the trained model to transform the data into the low dimensional embeddings. We then aligned the embeddings across all patients with Procrustes alignment. For decoding analyses, we used 4-fold cross validation: we randomly split the data into 4 folds, preserving temporal structure (timepoints within the same trial were not separated; only whole trials were held out). Train/test splits were chosen prior to training the models. We then trained the CEBRA models, holding out each fold (test set) while training on the remaining data (training set). For each fold, we then used a K-nearest neighbors classifier to predict trial labels from the embeddings (concatenated across patients). Since we aligned embeddings across patients and only ran the decoder once per validation fold, we first simply used a 1-sample t-test to compare the performance on all cross-validation folds to chance level. Traditional permutation tests with CEBRA models were not feasible, given the size of the data and the required computing power. Instead, to confirm that our model validation was correct (i.e., no “data leakage” occurred during decoding) and that the model was in fact learning real patterns in the data, we also trained a CEBRA model on data with shuffled labels for each validation fold and confirmed that these models could not fit the shuffled datasets and decoding failed. When decoding creativity score, we discretized and normalized the continuous score into 10 bins, so we defined chance level in this situation as 10%. For visualization (e.g., Fig. 3b ) and for the linear regression to define a coding axis, we trained the CEBRA models on the continuous score, normalized from 0 to 1 to improve consistency across patients. For the analysis that clustered network activity, we used network identity as a supervisory signal for a multi-subject CEBRA model, and only trained the model on AUT trials. Identifying creativity coding axis A CEBRA model was trained on all patients’ datasets, with continuous normalized creativity score as a supervisory signal, resulting in 5 dimensional embeddings for LFP data and 3 dimensional embeddings for single unit data. Then, we fit a linear regression model and used the model weights to compute the creativity coding axis as a 5-or 3-dimensional unit vector that best explains the variance in the embedding. We then projected the embedding onto this axis (dot product of embedding and coding axis vector), yielding a 1-dimensional representation of creativity values along the coding axis. The same process was used to project data onto the trial type coding axis. Note that using normalized continuous creativity values as labels (i.e., eliminating individual differences in behavior between patients) improves the model fit (DMN: R 2 =0.86; DAN: R 2 =0.73) and correlations (DMN r =0.93; DAN r =0.86) relative to raw originality scores (DMN: R 2 =0.77, r =0.87; DAN: R 2 =0.66, r =0.79), but the same pattern remained, with the most descriptive coding axis being within the DMN. Single unit analysis Spike sorting was performed using Offline Sorter (Plexon, Dallas, TX). Initially, the signal threshold was set to-3.5 RMS. Next, an automatic spike detection was applied with a T-Distribution Expectation Maximization sorting method (parameters: degrees of freedom = 4, # of expected units = 3). Then, the classification of waveforms was followed by manual curation with each microelectrode channel being visually inspected for additional units that went undetected during the automatic sorting step. Units came from the following regions: 47 in medial temporal lobe, 34 in ventromedial prefrontal cortex, 13 in dorsomedial prefrontal cortex, 20 in middle cingulate cortex, and 8 in posterior cingulate cortex. 24 units were anatomically localized to DMN, 14 to FPN, 80 to limbic network, and 9 to VAN. Statistics Nonparametric statistical tests were used when comparing across networks due to the variable coverage. We report t statistics with degrees of freedom in parentheses. False detection rate (FDR) correction for multiple comparisons was applied for tests considering the 7 cortical networks, as these tests always had 7 comparisons. For these comparisons, all reported p -values (but not t -values) are corrected for multiple comparisons. FDR was also applied when determining selectivity of single unit responses, since we had to do a comparison for each of the 127 cells. All t-tests and nonparametric null hypothesis tests were two-tailed unless otherwise noted. All reported correlation coefficients r come from Pearson correlation tests. The goodness of fit of the linear regression that identified the coding axes was quantified as the coefficient of determination ( R 2 ). When directly comparing correlations or goodness of fit of specific networks (e.g., DMN vs DAN), we also re-did the comparison including only subjects that had coverage in both networks, and saw no difference in the results. Author contributions B.S. conceived and supervised the project. B.S., R.R., E.B., E.H.S., R.B., and S.R. designed the experiments and critically reviewed the methods. B.S., S.S., R.R., E.H.S., R.B., and T.D. developed the overall research strategy. R.R., R.C., M.L., and T.D. collected and analyzed electrophysiological and behavioral data. E.S., R.R. and R.C. contributed to data preprocessing pipelines and software development. B.S. and R.B. provided conceptual input on cognitive paradigms and creativity metrics. R.R., and R.C performed the single-unit analyses and wrote the initial draft of the manuscript. E.S. provided critical guidance on electrophysiology and clinical implementation. All authors contributed to data interpretation and provided critical revisions to the manuscript. Download figure Open in new tab Supplementary Figure 1. Average spectrogram for AUT trials (top) and BAT trials (bottom). Spectrograms are averaged across all DMN channels. Increased LFP power appears red/orange. Download figure Open in new tab Supplementary Figure 2. Left: ROC curves to quantify trial type decoding performance of a linear SVM, across networks. Corresponds to result shown in Main Fig. 2d . Right: Trial type decoding performance based on each network’s latent embeddings, as generated by CEBRA models trained to separate AUT and BAT trials. Download figure Open in new tab Supplementary Figure 3. Clustered network activity for individual subjects. Dimensionality reduction using CEBRA trained to separate activity of different networks, as in Main Fig. 2e , but for individual patients (first 11 patients shown). Download figure Open in new tab Supplementary Figure 4. Default mode network activity patterns intrinsically represent cognitive state. Latent embeddings generated by self-supervised CEBRA models (only time-contrastive training; no behavior labels to supervise training) for each network. Each dot is 1 timepoint. Embeddings are color-coded by trial type. AUT trials are orange and BAT trials are blue. Download figure Open in new tab Supplementary Figure 5. Latent embeddings for DMN based on a CEBRA model trained on all data from all subjects; embeddings shown for individual subjects (first 12 patients). Download figure Open in new tab Supplementary Figure 6. a) As in Main Fig. 3a , but for theta LFP power, comparing most creative – least creative trials. b) Decoding creativity level based on the latent embeddings, as in Main Fig. 3c , but for only the stimulus window. Bars on the right show decoding based on models trained with shuffled labels. Download figure Open in new tab Supplementary Figure 7. a) Goodness of fit of polynomial regression for DMN (orange) and DAN (purple), across different degrees of the polynomial. b) Decoding performance (creativity level) based on DMN and dorsal attention network embedding, for different numbers of latent dimensions. Performance increases with increased number of dimensions but starts to plateau after 5 dimensions. When reducing to a single dimension, decoding performance falls to chance for both networks. Dorsal attention network yields higher decoding performance for most numbers of dimensions. Acknowledgements We thank the labs of Elliot Smith, Ben Shofty, Cory Inman, and Shervin Rahimpour for feedback and discussion; Alex Price for help with electrophysiology recordings; Justin Campbell for LFP preprocessing instruction; Veronica Zarr for help with patient consent; Jack Bowler for analysis advice and comments on the manuscript; and the patients for participating in experiments. References 1. ↵ Saretzki , J. , Forthmann , B. & Benedek , M . A systematic quantitative review of divergent thinking assessments. Psychology of Aesthetics , Creativity, and the Arts ( 2024 ) doi: 10.1037/aca0000691 . OpenUrl CrossRef 2. ↵ Hass , R. W . Semantic search during divergent thinking . Cognition 166 , 344 – 357 ( 2017 ). OpenUrl CrossRef PubMed 3. ↵ Shi , L. et al. Large-scale brain network connectivity underlying creativity in resting-state and task fMRI: Cooperation between default network and frontal-parietal network . Biological Psychology 135 , 102 – 111 ( 2018 ). OpenUrl CrossRef PubMed 4. Kucyi , A. et al. Intracranial Electrophysiology Reveals Reproducible Intrinsic Functional Connectivity within Human Brain Networks . J Neurosci 38 , 4230 – 4242 ( 2018 ). OpenUrl Abstract / FREE Full Text 5. Christoff , K . Undirected thought: neural determinants and correlates . Brain Res 1428, 51 – 59 ( 2012 ). 6. ↵ O’Callaghan , C. , Shine , J. M. , Lewis , S. J. G. , Andrews-Hanna , J. R. & Irish , M . Shaped by our thoughts--a new task to assess spontaneous cognition and its associated neural correlates in the default network . Brain Cogn 93 , 1 – 10 ( 2015 ). OpenUrl CrossRef PubMed 7. ↵ Dixon , M. L. et al. Interactions between the default network and dorsal attention network vary across default subsystems, time, and cognitive states . NeuroImage 147 , 632 – 649 ( 2017 ). OpenUrl CrossRef PubMed 8. ↵ Raichle , M. et al. A default mode of brain function . PNAS ( 2001 ) doi: 10.1073/pnas.98.2.67 . OpenUrl CrossRef 9. ↵ Vallat , R. , Türker , B. , Nicolas , A. & Ruby , P . High Dream Recall Frequency is Associated with Increased Creativity and Default Mode Network Connectivity . Nat Sci Sleep 14 , 265 – 275 ( 2022 ). OpenUrl CrossRef PubMed 10. ↵ Das , A. , de Los Angeles , C. & Menon , V . Electrophysiological foundations of the human default-mode network revealed by intracranial-EEG recordings during resting-state and cognition . Neuroimage 250 , 118927 ( 2022 ). 11. ↵ Shofty , B. et al. The default network is causally linked to creative thinking . Mol Psychiatry 27 , 1848 – 1854 ( 2022 ). OpenUrl CrossRef PubMed 12. Beaty , R. E. , Benedek , M. , Silvia , P. J. & Schacter , D. L. Creative Cognition and Brain Network Dynamics . Trends Cogn Sci 20 , 87 – 95 ( 2016 ). OpenUrl CrossRef PubMed 13. Marron , T. R. et al. Chain free association, creativity, and the default mode network . Neuropsychologia 118 , 40 – 58 ( 2018 ). OpenUrl CrossRef 14. Girn , M. , Mills , C. , Roseman , L. , Carhart-Harris , R. L. & Christoff , K . Updating the dynamic framework of thought: Creativity and psychedelics . NeuroImage 213 , 116726 ( 2020 ). 15. ↵ Rastelli , C. et al. Neural dynamics of semantic control underlying generative storytelling . Commun Biol 8 , 1 – 17 ( 2025 ). OpenUrl CrossRef PubMed 16. ↵ Beaty , R. E. et al. Robust prediction of individual creative ability from brain functional connectivity . Proceedings of the National Academy of Sciences 115 , 1087 – 1092 ( 2018 ). OpenUrl Abstract / FREE Full Text 17. ↵ Berkovich-Ohana , A. , Glicksohn , J. , Ben-Soussan , T. D. & Goldstein , A . Creativity Is Enhanced by Long-Term Mindfulness Training and Is Negatively Correlated with Trait Default-Mode-Related Low-Gamma Inter-Hemispheric Connectivity . Mindfulness 8 , 717 – 727 ( 2017 ). OpenUrl CrossRef 18. ↵ Beaty , R. E. , Seli , P. & Schacter , D. L . Network Neuroscience of Creative Cognition: Mapping Cognitive Mechanisms and Individual Differences in the Creative Brain . Curr Opin Behav Sci 27 , 22 – 30 ( 2019 ). OpenUrl CrossRef PubMed 19. ↵ Bartoli , E. et al. Default mode network electrophysiological dynamics and causal role in creative thinking . Brain awa e199 ( 2024 ) doi: 10.1093/brain/awae199 . OpenUrl CrossRef PubMed 20. ↵ Heidekum , A. E. , Grabner , R. H. , De Smedt , B. , De Visscher , A. & Vogel , S. E . Interference during the retrieval of arithmetic and lexico-semantic knowledge modulates similar brain regions: Evidence from functional magnetic resonance imaging (fMRI) . Cortex 120 , 375 – 393 ( 2019 ). OpenUrl CrossRef PubMed 21. Solomon , E. A. et al. Dynamic Theta Networks in the Human Medial Temporal Lobe Support Episodic Memory . Current Biology 29 , 1100 – 1111 .e4 ( 2019 ). OpenUrl CrossRef PubMed 22. Umbach , G. et al. Time cells in the human hippocampus and entorhinal cortex support episodic memory . PNAS 117 , 28463 – 28474 ( 2020 ). OpenUrl Abstract / FREE Full Text 23. ↵ Annerer-Walcher , S. et al. How Reliably Do Eye Parameters Indicate Internal Versus External Attentional Focus? Cognitive Science 45 , e12977 ( 2021 ). OpenUrl CrossRef 24. ↵ Ray , S. & Maunsell , J. H. R . Do gamma oscillations play a role in cerebral cortex? Trends Cogn Sci 19 , 78 – 85 ( 2015 ). OpenUrl CrossRef PubMed 25. Ray , S. & Maunsell , J. H. R . Different Origins of Gamma Rhythm and High-Gamma Activity in Macaque Visual Cortex . PLOS Biology 9 , e1000610 ( 2011 ). OpenUrl CrossRef PubMed 26. Miller , K. J . Broadband Spectral Change: Evidence for a Macroscale Correlate of Population Firing Rate? J. Neurosci . 30 , 6477 – 6479 ( 2010 ). OpenUrl FREE Full Text 27. Manning , J. R. , Jacobs , J. , Fried , I. & Kahana , M. J . Broadband Shifts in Local Field Potential Power Spectra Are Correlated with Single-Neuron Spiking in Humans . J. Neurosci . 29 , 13613 – 13620 ( 2009 ). OpenUrl Abstract / FREE Full Text 28. ↵ Smith , A. D. & Bolam , J. P . The neural network of the basal ganglia as revealed by the study of synaptic connections of identified neurones . Trends in Neurosciences 13 , 259 – 265 ( 1990 ). OpenUrl CrossRef PubMed Web of Science 29. ↵ Schneider , S. , Lee , J. H. & Mathis , M. W . Learnable latent embeddings for joint behavioural and neural analysis . Nature 617 , 360 – 368 ( 2023 ). OpenUrl CrossRef PubMed 30. ↵ Mante , V. , Sussillo , D. , Shenoy , K. V. & Newsome , W. T . Context-dependent computation by recurrent dynamics in prefrontal cortex . Nature 503 , 78 – 84 ( 2013 ). OpenUrl CrossRef PubMed Web of Science 31. Rabinovich , R. J. , Kato , D. D. & Bruno , R. M . Learning enhances encoding of time and temporal surprise in mouse primary sensory cortex . Nat Commun 13 , 5504 ( 2022 ). 32. ↵ Cunningham , J. P. & Yu , B. M . Dimensionality reduction for large-scale neural recordings . Nat Neurosci 17 , 1500 – 1509 ( 2014 ). OpenUrl CrossRef PubMed 33. ↵ Sheline , Y. I. et al. The default mode network and self-referential processes in depression . Proc Natl Acad Sci U S A 106 , 1942 – 1947 ( 2009 ). OpenUrl Abstract / FREE Full Text 34. ↵ Poerio , G. L. et al. The role of the default mode network in component processes underlying the wandering mind . Social Cognitive and Affective Neuroscience 12 , 1047 – 1062 ( 2017 ). OpenUrl CrossRef PubMed 35. ↵ Beaty , R. E. & Kenett , Y. N . Associative thinking at the core of creativity . Trends in Cognitive Sciences 27 , 671 – 683 ( 2023 ). OpenUrl CrossRef PubMed 36. ↵ Davis , T. S. et al. LeGUI: A Fast and Accurate Graphical User Interface for Automated Detection and Anatomical Localization of Intracranial Electrodes . Frontiers in Neuroscience 15 , ( 2021 ). 37. ↵ Yeo , B. T. T. et al. The organization of the human cerebral cortex estimated by intrinsic functional connectivity . J Neurophysiol 106 , 1125 – 1165 ( 2011 ). OpenUrl CrossRef PubMed Web of Science 38. ↵ Dumas , D. , Organisciak , P. & Doherty , M . Measuring divergent thinking originality with human raters and text-mining models: A psychometric comparison of methods. Psychology of Aesthetics , Creativity, and the Arts 15 , 645 – 663 ( 2021 ). OpenUrl 39. ↵ Organisciak , P. & Dumas , D . Open Creativity Scoring. ( 2020 ). 40. ↵ Beaty , R. E. & Johnson , D. R . Automating creativity assessment with SemDis: An open platform for computing semantic distance . Behav Res 53 , 757 – 780 ( 2021 ). OpenUrl CrossRef 41. ↵ Pedregosa , F. et al. Scikit-learn: Machine Learning in Python . Journal of Machine Learning Research 12 , 2825 – 2830 ( 2011 ). OpenUrl View the discussion thread. Back to top Previous Next Posted June 03, 2025. Download PDF Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Human brain network control of creativity Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Human brain network control of creativity Rikki Rabinovich , Rhiannon L. Cowan , Mark Libowitz , Tyler S. Davis , Shervin Rahimpour , Roger E. Beaty , Eleonora Bartoli , Sameer A. Sheth , Elliot J. Smith , Ben Shofty bioRxiv 2025.06.03.657648; doi: https://doi.org/10.1101/2025.06.03.657648 Share This Article: Copy Citation Tools Human brain network control of creativity Rikki Rabinovich , Rhiannon L. Cowan , Mark Libowitz , Tyler S. Davis , Shervin Rahimpour , Roger E. Beaty , Eleonora Bartoli , Sameer A. Sheth , Elliot J. Smith , Ben Shofty bioRxiv 2025.06.03.657648; doi: https://doi.org/10.1101/2025.06.03.657648 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Neuroscience Subject Areas All Articles Animal Behavior and Cognition (7629) Biochemistry (17660) Bioengineering (13881) Bioinformatics (41910) Biophysics (21436) Cancer Biology (18576) Cell Biology (25480) Clinical Trials (138) Developmental Biology (13368) Ecology (19887) Epidemiology (2067) Evolutionary Biology (24302) Genetics (15598) Genomics (22482) Immunology (17726) Microbiology (40360) Molecular Biology (17163) Neuroscience (88534) Paleontology (666) Pathology (2830) Pharmacology and Toxicology (4821) Physiology (7637) Plant Biology (15129) Scientific Communication and Education (2045) Synthetic Biology (4290) Systems Biology (9817) Zoology (2269)