The default-mode network represents aesthetic appealthat generalizes across visual domainsEdward A. Vessela,1, Ayse Ilkay Isika, Amy M. Belfib, Jonathan L. Stahlc, and G. Gabrielle Starrd

aDepartment of Neuroscience, Max Planck Institute for Empirical Aesthetics, Frankfurt am Main 60322, Germany; bDepartment of Psychological Science,Missouri University of Science and Technology, Rolla, MO 65409; cDepartment of Psychology, The Ohio State University, Columbus, OH 43210;and dDepartments of English and Neuroscience, Pomona College, Claremont, CA 91711

Edited by Moshe Bar, Bar-Ilan University, Ramat Gan, Israel, and accepted by Editorial Board Member Michael S. Gazzaniga July 15, 2019 (received for reviewFebruary 13, 2019)

Visual aesthetic evaluations, which impact decision-making andwell-being, recruit the ventral visual pathway, subcortical rewardcircuitry, and parts of the medial prefrontal cortex overlapping withthe default-mode network (DMN). However, it is unknownwhetherthese networks represent aesthetic appeal in a domain-generalfashion, independent of domain-specific representations of stimuluscontent (artworks versus architecture or natural landscapes). Usinga classification approach, we tested whether the DMN or ventraloccipitotemporal cortex (VOT) contains a domain-general represen-tation of aesthetic appeal. Classifiers were trained on multivoxelfunctional MRI response patterns collected while observers madeaesthetic judgments about images from one aesthetic domain.Classifier performance (high vs. low aesthetic appeal) was thentested on response patterns from held-out trials from the samedomain to derive a measure of domain-specific coding, or from adifferent domain to derive a measure of domain-general coding.Activity patterns in category-selective VOT contained a degree ofdomain-specific information about aesthetic appeal, but did notgeneralize across domains. Activity patterns from the DMN, how-ever, were predictive of aesthetic appeal across domains. Impor-tantly, the ability to predict aesthetic appeal varied systematically;predictions were better for observers who gave more extremeratings to images subsequently labeled as “high” or “low.” Thesefindings support a model of aesthetic appreciation wherebydomain-specific representations of the content of visual experi-ences in VOT feed in to a “core” domain-general representationof visual aesthetic appeal in the DMN. Whole-brain “searchlight”analyses identified additional prefrontal regions containing infor-mation relevant for appreciation of cultural artifacts (artwork andarchitecture) but not landscapes.

visual aesthetics | default-mode network | artwork | architecture |natural landscape

Aesthetic appeal is a fundamental aspect of human experi-ence that touches many aspects of life. Aesthetic consider-

ations affect purchasing decisions (1–3), choice of leisure timeactivities (4–6), stress levels (7–9), recovery from illness (10, 11),and well-being more generally (12–15). Aesthetics affect centralaspects of perception and cognition, including visual orienting(16), learning (17, 18), and valuation (19, 20), and also affect ourinteractions with other people, from evaluations of attractivenessto judgments of trustworthiness and competence (21).An aesthetic experience is, in general, a perceptual experience

that is evaluative and affectively absorbing and engages com-prehension (meaning) processes (22, 23). Aesthetic experiencesoften have a significant conceptual component, such as duringencounters with conceptual art or abstract mathematical problemsolving (24, 25), and may also emerge in response to imaginedobjects. Such experiences may include feelings of pleasure orbeauty from engaging with an object, and judgments of liking orattractiveness, but can also include more complex responses suchas being moved and feelings of awe and the sublime (26, 27).People have strong aesthetic experiences with images from a

variety of different visual aesthetic domains, such as landscapes,

faces, architecture, and artwork. Yet the perceptual features thatsupport such aesthetic experiences are domain-specific. For ex-ample, the features that support perception and aesthetic ap-preciation of a mountain vista are different from those used toperceive and aesthetically evaluate a gothic church (28–31). Re-cent behavioral studies support a distinction between aestheticappraisals of natural kinds, such as faces and landscapes, versusthose of cultural artifacts, such as artwork and architecture: Aes-thetic appraisals of cultural artifacts are highly individual, whereasassessments of faces and landscapes tend to contain a strong de-gree of “shared taste” (32, 33).Is there a domain-general brain system that represents aes-

thetic appeal regardless of aesthetic domain, or, alternately, dodifferent aesthetic domains engage domain-specific processesfor aesthetic valuation? A number of studies on decision-makingpoint to a region of the ventral or anterior medial prefrontalcortex (vMPFC/aMPFC) as containing a domain-general repre-sentation of value (using money, food, water, arousing pictures,social feedback) (34, 35), and this same brain region has beenshown to be important for aesthetic appeal (36–38). However,the relationship between aesthetic appeal and valuation is notstraightforward. Aesthetic considerations represent only a subset ofinputs to valuation and subsequent decision-making (e.g., ref. 19),

Significance

Despite being highly subjective, aesthetic experiences are pow-erful moments of interaction with one’s surroundings, shapingbehavior, mood, beliefs, and even a sense of self. The default-mode network (DMN), which sits atop the cortical hierarchy andhas been implicated in self-referential processing, is typicallysuppressed when a person engages with the external environ-ment. Yet not only is the DMN surprisingly engaged when onefinds a visual artwork aesthetically moving, here we presentevidence that the DMN also represents aesthetic appeal in amanner that generalizes across visual aesthetic domains, such asartworks, landscapes, or architecture. This stands in contrast toventral occipitotemporal cortex (VOT), which represents thecontent of what we see, but does not contain domain-generalinformation about aesthetic appeal.

Author contributions: E.A.V., J.L.S., and G.G.S. designed research; E.A.V., A.M.B., and J.L.S.performed research; E.A.V., A.I.I., and A.M.B. contributed new reagents/analytic tools;E.A.V. and A.I.I. analyzed data; and E.A.V., A.I.I., A.M.B., J.L.S., and G.G.S. wrote the paper.

The authors declare no conflict of interest.

This article is a PNAS Direct Submission. M.B. is a guest editor invited by theEditorial Board.

Published under the PNAS license.

Data deposition: MRI and behavioral data are posted in anonymized form on Edmond(https://edmond.mpdl.mpg.de), the Open Access Data Repository of the MaxPlanck Society, and can be accessed at https://dx.doi.org/10.17617/3.2r.1To whom correspondence may be addressed. Email: ed.vessel@ae.mpg.de.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1902650116/-/DCSupplemental.

Published online September 4, 2019.

www.pnas.org/cgi/doi/10.1073/pnas.1902650116 PNAS | September 17, 2019 | vol. 116 | no. 38 | 19155–19164

NEU

ROSC

IENCE

PSYC

HOLO

GICALAND

COGNITIVESC

IENCE

S

Dow

nloa

ded

by g

uest

on

Nov

embe

r 18

, 202

0

and aesthetic judgments can be made in the absence of approachmotivation, a signature of reward (39). In addition, aesthetic ex-periences need not be based on either primary rewards or simpleassociations with primary rewards, such as when imagery (40) orinformation foraging, sense making, and uncertainty reduction(41–43) lead to pleasing aesthetic experiences.Yet if aesthetic appreciation is derived from evolutionarily

older processes of object appraisal, there should be overlap inthe neural mechanisms for the evaluation of a variety of aestheticdomains, at least partially coinciding with a general system forrepresenting subjective value. Indeed, a functional MRI (fMRI)study of beauty judgments of music and visual art found overlapat the group average level in a single region in vMPFC/ACC(anterior cingulate cortex) (44). Conversely, a metaanalysis ofover 93 studies of pleasurable experiences across 4 sensory mo-dalities (vision, taste, hearing, and smell) did not find evidencefor overlap in either vMPFC/aMPFC or striatum, and insteadidentified the anterior insula as a region with the highest overlap(in 3 of 4 domains) (45). A pattern of adjacency, rather thanoverlap, was seen in orbitofrontal cortex (OFC) and vMPFC.Unfortunately, the methods used in both of these studies wereinadequate for resolving the question of whether there was truedomain-generality at the level of individual observers: Adjacentbut nonoverlapping activations at the level of individuals canlead to spurious overlap at the group or metaanalytic levels,given the difficulty in precisely aligning cortical surfaces acrossobservers and studies.A recent study using a more precise multivoxel pattern anal-

ysis (MVPA) method found evidence for coding of valence inOFC and vMPFC for both strongly valenced (but not “aes-thetic”) visual images and gustatory stimuli (46). In the aestheticdomain, Pegors et al. (47) found evidence for overlapping acti-vations for positive aesthetic judgments of landscapes and facesin regions of the vMPFC, as well as domain-specific regions inother anterior parts of the brain. Yet, given the distinction be-tween behavioral responses to natural vs. artifactual aestheticdomains (see above), it is notable that none of these studiesaddresses this distinction.In addition to their importance for valuation, the vMPFC and

aMPFC are also central nodes of the default-mode network(DMN), a set of interconnected brain regions that are engagedby tasks that require inwardly directed attention or an assessmentof self-relevance (48–51). Surprisingly, the DMN is responsive tostrongly moving aesthetic experiences with visual artwork (52, 53).

Several nodes of the DMN, which are typically suppressed whenattention is directed to an external object such as a visual image(54–56), were found to be released from suppression when anobserver viewed paintings rated as strongly aesthetically moving,but not for paintings rated as merely pleasing or as not aestheticallyappealing. This effect was particularly strong in the aMPFC, a“hub” region of the DMN (48), but was also present in other nodes.We tested the hypothesis that the DMN contains a represen-

tation of aesthetic appeal that generalizes across different visualdomains, using a strong multivariate test. We trained classifierson voxelwise patterns of fMRI blood-oxygen-level-dependent(BOLD) activity to distinguish trials of high versus low aestheticappeal for one domain (artworks, natural landscapes, or archi-tecture; Fig. 1), and tested the classifiers’ ability to predict ob-servers’ judgments on the other domains. Given the hypothesizedimportance of DMN regions for representing and processing self-relevant, internally directed information (51, 57), we predicted thatclassifiers trained on data from the DMN would perform well onacross-domain classifications, providing strong evidence for adomain-general representation. Alternatively, it is possible that theDMN represents information about a variety of aesthetic domains,but in a manner that is domain-specific. This would lead to accu-rate classification within domain, but poor across-domain gener-alization. Lastly, it is also possible that only specific subregions ofthe DMN such as vMPFC or aMPFC, but not the DMN as awhole, contain domain-general representations of aesthetic appeal.Additionally, we used a “searchlight” multivariate mapping

technique to identify regions of cortex that contained domain-specific representations of aesthetic appeal for artifacts of humanculture (artworks, architecture) or natural landscapes. Similarapproaches have been used to understand brain systems for val-uation of money, food, and consumer goods (20, 34, 35, 58). Whilethis approach has previously been extended to the aesthetic do-main (47), no direct comparison between natural and artifactualaesthetic domains has been made, nor has there been any attemptto topographically map domain-specific and domain-general aes-thetic processing across the cortex.We found evidence for a domain-general representation of

aesthetic appeal in the DMN, whereas higher-level visual regions(ventral occipitotemporal cortex, VOT) exhibited a domain-specific pattern. Importantly, our ability to classify trials as highor low in aesthetic appeal on the basis of fMRI signal wasstrongly related to the degree to which each observer behavior-ally distinguished those high versus low trials. The searchlight

Fig. 1. Stimuli and experimental design. Examples of images used in the experiment: (A) visual art, (B) interior and exterior architecture, and (C) naturallandscapes. (D) Each trial began with a fixation point (1 s), followed by an image of an artwork (4 s) and a rating period (4 s) during which the observer used atrackball to indicate their response on a visual slider. (E) Classifiers were trained on multivoxel patterns of trialwise (beta-series) estimates taken from in-dividual observer ROIs. Three sets of “within-domain” classification scores and 6 sets of “across-domain” classification scores were derived for each observer usingcross-validation. (Fig. 1 A, Left) Reprinted from ref. 87. (Fig. 1 A, Right) Reprinted from ref. 88. (Fig. 1 B, Left) Image courtesy of Alec Hartill (photographer).(Fig. 1 B, Right) Image courtesy of R. Hoekstra (photographer). (Fig. 1C) Images are examples from the SUN database, https://groups.csail.mit.edu/vision/SUN/.

19156 | www.pnas.org/cgi/doi/10.1073/pnas.1902650116 Vessel et al.

Dow

nloa

ded

by g

uest

on

Nov

embe

r 18

, 202

0

analysis revealed several additional regions in prefrontal cortexthat contained information about aesthetic appeal of art andarchitecture, whereas domain-specific information about aestheticappeal of landscapes was mostly confined to the ventral visualpathway.

ResultsClassifier Performance in the DMN. Patterns of multivoxel activityacross the entire DMN contained a strong signature of domain-general aesthetic processing (Fig. 2). The 3 within-domain clas-sifiers (Fig. 2A) achieved an average classification of high vs. lowpreferred trials at a rate of 63.8% (95% CI [61.4 66.4], averagearea under the curve [AUC] 0.68, d = 2.78; Fig. 2C, gray bars),which was highly significant compared to chance performance of50% (permutation testing of individual trial labels, P < 0.005).The 6 across-domain classifiers also performed well abovechance, with an average classification rate of 60.7% (95% CI[57.2 64.3], average AUC 0.64, d = 1.61, P < 0.005).Above-chance domain-general classification of aesthetic ap-

preciation was present in most of the DMN subregions tested,although the strength of both the domain-general and domain-

specific signals varied from region to region (Fig. 2 B and C, coolcolors). The strongest within- and across-domain classificationperformance was observed in the aMPFC (within-domain 60.1%,95% CI [57.5 62.8], average AUC 0.63, d = 2.03, P < 0.005;across-domain 58.3%, 95% CI [54.7 61.9], average AUC 0.61,d = 1.23, P < 0.005) and dorsomedial prefrontal cortex (dMPFC,within-domain 59.7%, 95% CI [57.4 61.9], average AUC 0.63,d = 2.30, P < 0.005; across-domain 58.5%, 95% CI [55.5 61.5],average AUC 0.61, d = 1.50, P < 0.005). Classifiers trained ondata from the inferior parietal lobule (IPL) also performed well,both within (58.8%, 95% CI [56.2 61.4], average AUC 0.61, d =1.81, P < 0.005) and across domains (55.8%, 95% CI [52.9 58.6],average AUC 0.58, d = 1.08, P < 0.005). Patterns of activation inposterior cingulate cortex (PCC, within-domain 56.8%, 95% CI[54.0 59.5], average AUC 0.59, d = 1.32, P < 0.005; across-domain 54.0%, 95% CI [51.6 56.4], average AUC 0.55, d =0.89, P < 0.005) and lateral temporal cortex (LTC, within-domain 54.5%, 95% CI [51.8 57.1], average AUC 0.55, d =0.90, P < 0.005; across-domain 53.5%, 95% CI [51.4 55.6], av-erage AUC 0.54, d = 0.89, P < 0.005) contained a degree

Fig. 2. Predicting aesthetic appeal from activity patterns in DMN and VOT. (A) For each region, a set of classifiers were trained on voxelwise activity patternsfrom trials of one stimulus domain and tested on separate trials of another domain to produce a 3 × 3 performance matrix. Classification performance(percent correct) was averaged across the 3 diagonal elements to produce a within-domain accuracy score, and the 6 off-diagonal elements were averaged toproduce an across-domain score. (B) Six bilateral DMN ROIs (Left) were identified in individual participants: aMPFC, vMPFC, dMPFC, PCC, IPL, and LTC. Theregions shown here are the “master” ROIs drawn on an average template brain—these regions were then used to mask individual DMN maps derived from aseparate “rest” scan. In addition, 3 category-selective, bilateral VOT regions (Right) were identified in individual participants using a separate object/face/place localizer scan (approximate location on the average template brain shown here): PPA, VOA, and FFA. An overall DMN mask was created by summing all6 DMN subregions together, and an overall VOT mask was created by projecting the larger VOT region shown here onto individual participant surfaces. (C)Within- and across- domain classification accuracy scores for each region. DMN regions are left of the dashed line in cool colors, and VOT regions are right ofthe dashed line in warm colors; n = 16. Error bars are 95% CIs; **P < 0.005, tested by comparison to a null distribution derived from 5,000 permutations ofindividual trial labels.

Vessel et al. PNAS | September 17, 2019 | vol. 116 | no. 38 | 19157

NEU

ROSC

IENCE

PSYC

HOLO

GICALAND

COGNITIVESC

IENCE

S

Dow

nloa

ded

by g

uest

on

Nov

embe

r 18

, 202

0

of domain-general decoding information that was less than theother regions.Surprisingly, the vMPFC did not show within-domain classi-

fication performance better than chance (52.1%, 95% CI [50.653.5], average AUC 0.52, d = 0.77, P = 0.37) and producedacross-domain classification that was higher than chance per-formance but with a lower effect size (52.6%, 95% CI [51.0 54.2],average AUC 0.54, d = 0.84, P < 0.005).

Classifier Performance in Category-Selective VOT Regions. Classifierstrained on signal from the category-selective regions of theVOT showed some evidence of domain selectivity, but little ev-idence of domain generality (Fig. 2 B and C, warm colors). All3 category-selective regions showed average within-domain per-formance that was better than chance (fusiform face area [FFA]55.6%, 95% CI [53.6 57.6], average AUC 0.57, d = 1.49, P <0.005; parahippocampal place area [PPA] 54.5%, 95% CI [51.257.8], average AUC 0.55, d = 0.73, P < 0.005; ventral object area[VOA] 53.5%, 95% CI [51.3 55.7], average AUC 0.53, d = 0.83,P < 0.005), but did not differ from chance for average across-domain classification (FFA 51.5%, 95% CI [49.2 53.7], averageAUC 0.52, d = 0.34, P = 0.45; PPA 50.9%, 95% CI [49.1 52.6],average AUC 0.51, d = 0.26, P = 1; VOA 50.7%, 95% CI [48.752.7], average AUC 0.51, d = 0.17, P = 1). A classifier trained ona large region of interest (ROI) covering the entire VOT performedbetter than the individual ROIs for both within-domain (57.5%,95% CI [55.6 59.4], average AUC 0.61, d = 2.12, P < 0.005) andacross-domain classifications (53.6%, 95% CI [51.3 55.9], averageAUC 0.55, d = 0.84, P < 0.005), although at a level that was stillinferior to individual regions of the DMN.

Comparison of Classifier Performance in DMN and VOT. A compari-son across all ROIs (Fig. 3) revealed that classifiers trained ondata from DMN ROIs, despite a range of overall performance,tended to have similar domain-general and domain-specificperformance, whereas classifiers trained on VOT ROIs tendedto show better performance within domain than across domain.When within- and across-domain performances were plottedagainst each other, DMN ROIs tended to lie on or just below thediagonal (Fig. 3, cool colors), whereas VOT ROIs tended to liesignificantly below the diagonal, closer to the line markingchance across-domain performance. These effects were not aconsequence of the independent method of ROI identification:A complementary analysis of regions activated by all 3 domainsversus a resting baseline also found domain-general behavior inMPFC but a more domain-specific signature in a large visualROI covering the entire ventral visual pathway (SI Appendix, Fig.S1 and Supplementary Results).This pattern—domain-general discriminability between high-

and low-rated trials in the DMN but poorer, nongeneralizedperformance in VOT—stands in contrast to standard activation-based measures for these same regions (SI Appendix, Fig. S2).VOT regions were activated by images of all 3 domains (all P <0.005, significance tested by group-level randomization test; seeSI Appendix, Supplementary Results) and were generally moreactivated for high vs. low trials (VOT and PPA P < 0.005 for alldomains; FFA P < 0.005 for architecture; VOA P < 0.005 forlandscape and P < 0.05 for architecture). DMN regions, on theother hand, were deactivated by images of all 3 domains [all P <0.005 except PCC landscape P < 0.05 and vMPFC all domainsnonsignificant (n.s.)], and generally did not show differences inaverage activation for trials labeled high vs. low (all P n.s. exceptvMPFC architecture, P < 0.005). Specific tests of differences acrossaesthetic domains were significant only for Image vs. Baselinecontrasts in VOT (all P < 0.005), but not for activation-basedcontrasts of high vs. low trials in VOT, nor for DMN.

Individual Variation in a Behavioral Distance Metric (d’) PredictedPerformance of Domain-General Classification from DMN. The abil-ity to classify single trials as high or low aesthetic appeal based ondomain-general activation patterns in the DMN varied from in-dividual to individual. We performed an analysis to test whetherthis variation in classifier performance was related to the strengthof observers’ aesthetic experiences. Behaviorally, we observed thatsome participants tended to use a much wider range of the ratingscale than did others; these participants also tended to have longerresponse times (SI Appendix, Supplementary Results). The selectionof 40 trials (10 from each run) as “high” and “low” for each do-main in each observer, regardless of the underlying distribution,resulted in high and low distributions with different distancesbetween them for different observers. This behavioral distance(expressed as a d’) strongly predicted how well classifiers trainedon data from that participant’s DMN performed across domains[r = 0.73, linear regression R2 = 0.53, F(1,14) = 15.9, P = 0.0013;Fig. 4]. This suggests that differences in scale use accuratelyreflected the distinctiveness of brain states subsequently labeled as“high” and “low” aesthetic appeal, and that variation in classifi-cation accuracy largely reflected this distinctiveness. This re-lationship was specific for signal from the DMN: VOT classifierperformance was not related to behavioral d’ [r = 0.27, R2 = 0.075,F(1,14) = 1.1, P = 0.31].

Cortical Topography of Domain-Specific and Domain-General Information.As a complement to the characterization of specific a priori ROIs,cortical maps of across-domain and within-domain performancewere generated using a “searchlight” technique (see Methods).Several clusters of better-than-chance across-domain classificationperformance were found on the medial surface of PFC in both

Fig. 3. Characterization of domain-general and domain-specific ROI sig-natures. Each dot represents one ROI (color key same as Fig. 2; gray, DMN;purple, aMPFC; maroon, vMPFC; indigo, dMPFC; blue, PCC; green, IPL; lightgreen, LTC; beige, VOT; yellow, PPA; orange, VOA; red, FFA). Average across-domain performance (y axis) is plotted against average within-domain per-formance (x axis). The DMN ROIs (cool colors) all tend to fall near the di-agonal, illustrating a “domain-general” signature (similar across-domainand within-domain performance levels). The VOT ROIs (warm colors), how-ever, tend to fall along the horizontal and are thus better characterized as“domain-specific” (significant within-domain performance but poor across-domain performance), with the exception of the overall VOT ROI (tan),which is above the horizontal line; n = 16. Error bars are 95% CIs.

19158 | www.pnas.org/cgi/doi/10.1073/pnas.1902650116 Vessel et al.

Dow

nloa

ded

by g

uest

on

Nov

embe

r 18

, 202

0

hemispheres (Fig. 5A and SI Appendix, Table S1; dMPFC, aMPFC/rostral paracingulate [rPaC], right ACC/paracingulate [PaC]),consistent with the DMN ROI findings. In addition, several ad-ditional prefrontal clusters were found in left inferior frontal sul-cus (IFS)/inferior frontal gyrus, pars opercularis (IFGop), rightdorsolateral prefrontal cortex (dlPFC), right frontopolar lateral(FPl), and left OFC. In posterior cortex, significant clusters wereobserved on the left hemisphere in higher-level visual regions(medial occipital gyrus [MOG], occipitotemporal sulcus [OTS]/inferior occipital gyrus [IOG]) and posterior cingulate, as well asin bilateral early visual cortex (see SI Appendix, Table S1 for acomplete listing of significant clusters). These results indicate thatdomain-general representations of aesthetic appeal exist not onlyacross large-scale patterns of activity but also in local regions ofcortex, including MPFC and OFC.The 3 maps of within-domain classification performance (one

each for art, architecture, and natural landscapes) containedseveral prefrontal clusters with domain-specific signatures (Fig.5B and SI Appendix, Table S1). For art (red), these clusters werelocated in the frontopolar regions (left FPl, left frontopolarmedial [FPm], bilateral dMPFC) and inferior frontal cortex(bilateral inferior frontal gyrus [IFGt], right IFS). For architec-ture (blue), clusters were observed in frontopolar regions (leftFPm, right FPl/frontomarginal gyrus [FM], bilateral dMPFC, leftsuperior frontal gyrus [SFG]), dorsolateral prefrontal (left MFG/IFS), lateral orbitofrontal, and medial prefrontal cortex (rostralanterior cingulate cortex, [rACC]/rPaC). A comparison to thedomain-general map (orange outlines in Fig. 5B) revealed sev-eral left hemisphere frontopolar clusters that are adjacent to, butnot overlapping with, the domain-general regions. Similarly, theclusters in left OFC/IFG are near, but not overlapping with, adomain-general cluster. The landscape map contained 2 pre-frontal domain-specific clusters on the lateral surface, in theprecentral sulcus and inferior frontal (IFS/IFGt).In addition to these prefrontal clusters, the landscape map

contained large domain-specific clusters bilaterally in VOT (Fig.

5B, green) and in some parts of early visual cortex that overlapwith those observed in the domain-general map. Domain-specificclusters were also observed in visual regions for artwork (IOG,early visual) and architecture (bilateral MOG, early visual) thatmostly, but not entirely, overlapped with domain-general clusters.

DiscussionDespite differences in how images of natural landscapes, archi-tecture, and visual art are encoded and evaluated, the DMNcontains a representation of aesthetic appeal that generalizesacross these visual aesthetic domains. A series of multivariateclassifiers were trained to distinguish trials rated as “high” vs.“low” aesthetic appeal using multivoxel patterns of BOLD signalfrom regions of the DMN. When the full network was used,robust performance was observed for classifiers trained andtested within the same domain as well as for classifiers trainedand tested across different domains. This is strong evidence thatthe DMN contains a representation of aesthetic appeal that isdomain-general. Additionally, domain-general classification perfor-mance in individual participants was strongly predicted by the dis-tance (d’) between each person’s behavioral ratings of exemplarslabeled “high” (top 27% of images in each domain) versus exem-plars labeled “low” (bottom 27% of images in each domain), sug-gesting that variability in classifier performance reflected thepsychological and neural span between the states evoked by the mostand least aesthetically appealing images in each observer, ratherthan measurement noise. Strong within- and across-domain perfor-mance was also observed in several subregions of the DMN, namelyaMPFC, dMPFC, and IPL. On the other hand, classifiers trained ondata from VOT did not perform as well and were strongly domain-specific: The voxel patterns that predicted aesthetic appeal for onedomain did not generalize to the other domains.Maps of cortical domain-general and domain-specific signals,

measured using a multivariate “searchlight” technique, confirmedthe presence of a domain-general representation in aMPFC anddMPFC, and also revealed adjacent domain-specific regions infrontal pole and inferior/orbital frontal cortex, primarily for ar-chitecture and art. Domain specificity for natural landscapes, incontrast, was primarily found in the ventral visual pathway. Finally,additional cortical fields containing domain-general informationabout aesthetic appeal were identified in the lateral and orbitalprefrontal cortex, as well as in posterior visual regions.Unlike metaanalyses or activation analyses, across-domain

classification based on voxelwise patterns is a very strong test ofdomain generality. A univariate activation analysis of the sameregions, which found sensitivity to ratings in VOT but not inDMN, might have led one to conclude that VOT possessedgreater sensitivity and domain generality for aesthetic appeal(despite discriminating domain identity). This discrepancy sug-gests that aesthetic appeal, like valence of nonaesthetic images(46), is encoded at a spatial scale that is smaller than brain re-gions, but still detectable at the voxel level, and may not have aconsistent local topography from one person to the next (see alsoSI Appendix, Supplementary Results). Although high multivariateclassification accuracy cannot prove that the identical neurons orcolumns are representing aesthetic appeal across different do-mains, it does show that individual voxels, in individual ob-servers, respond to high versus low exemplars in a consistentmanner regardless of aesthetic domain.

A Core System for Assessing Aesthetic Appeal. It is notable that theDMN representation of aesthetic appeal generalized across bothnatural kinds (landscapes) and cultural artifacts (architectureand artwork). Recent behavioral findings suggest that aestheticevaluations of natural kinds such as landscapes and faces rely onsimilar information across people, whereas evaluations of culturalartifacts are highly individual (32). Attractive faces and attractivelandscapes engage a region in MPFC (likely overlapping with the

Fig. 4. Variability in DMN classifier performance reflects the strength ofobservers’ aesthetic experiences. For each observer, the distance betweenthe top-rated trials (labeled “high”) and the bottom-rated trials (labeled“low”) was calculated based on behavioral ratings (d’). Therefore, observerswith greater variability in their ratings produced higher d’ values. Sepa-rately, classifiers were trained on BOLD signal patterns from each observers’DMN to distinguish trials labeled as “high” vs. “low.” Across observers,classifier performance (vertical axis) was strongly correlated with each ob-servers’ behavioral d’ distance measure (horizontal axis); n = 16.

Vessel et al. PNAS | September 17, 2019 | vol. 116 | no. 38 | 19159

NEU

ROSC

IENCE

PSYC

HOLO

GICALAND

COGNITIVESC

IENCE

S

Dow

nloa

ded

by g

uest

on

Nov

embe

r 18

, 202

0

DMN) in a similar manner (47). A potential interpretation of ourresults is that the DMN is part of a “core” system for assessingaesthetic appeal that is engaged by all domains.Aesthetic experiences are integrative in nature, drawing on

perception and imagery across multiple senses as well as onmemories, emotion, and associated meanings. The DMN’s ana-tomical position at one extreme of a cortical hierarchy (59)makes it well positioned to integrate information across multiplebrain systems. While it remains unclear whether DMN involve-ment is necessary for an aesthetic experience to be perceived asstrongly moving, it is the case that its activity reflects engagementwith artworks: DMN signal fluctuations “lock on” to aestheticallyappealing artworks, but are independent for nonappealing art-works (60). In addition, MPFC damage has been shown to re-duce the influence of certain types of affective information onaesthetic valuation (61). The current findings add to this un-derstanding by showing that the DMN encodes the aestheticappeal of visual images in a domain-general manner. Given theDMN’s strong link to assessments of self-relevance and self-referential processes (57, 62, 63), it is possible that the DMN’sengagement by and representation of aesthetically appealingevents reflects an assessment of self-relevance—in this case, thepotential self-relevance of an external object.Alternatively, the DMN’s engagement during aesthetic expe-

rience may reflect its theorized role in the construction of mentalscenes (48). Such mental imagery, involving an interplay of top-down information with bottom-up stimulus properties, is a key

aspect of many aesthetic experiences (40, 64). This balance ofactivation between higher-tier visual regions and DMN regionslikely depends on the degree to which an observer is able to rec-ognize familiar content (65). While this study was not designed totease apart responses to abstract versus representational artwork(out of 148 artworks, 11 were abstract), it is possible that aestheticexperiences with abstract artworks rely more on top-down pro-cesses of sense-making and imagery, consistent with the fact thataesthetic judgments for images of indeterminate content takelonger than for representational images (66). Yet the ability todecode aesthetic appeal across aesthetic domains, including pho-tographs of landscapes and architecture, suggests that even if theDMN activates differentially to representational or abstract con-tent, the multivoxel patterns for images experienced as high vs. lowappeal are similar.Additionally, the fact that DMN contains information about

aesthetic appeal for artworks and architecture suggests that theDMN is able to integrate information about nonperceptual as-pects of aesthetic experience, as the low degree of shared tastefor these domains indicates that visual features do not uniquelydetermine felt aesthetic appeal (32, 67, 68). It remains to be seenwhether the domain-general aesthetic appeal observed in theDMN for the visual domains studied here would also generalizeto nonvisual domains (music, poetry) or to highly conceptualexperiences such as the appreciation of mathematical beauty.The domain-specific regions identified in the searchlight anal-

ysis likely complement this core system. Interestingly, domain

Fig. 5. Topography of domain-general and domain-specific information. (A) Average of the 6 across-domain classifier maps (orange), rendered on an av-erage flattened cortical surface. (B) Three within-domain classifier maps and their overlap, as indicated by the color key. Outlines from the across-domainmaps are shown in orange. The right hemisphere is on the left side. Maps were cluster-corrected for multiple comparisons using Monte Carlo simulations (P <0.05); n = 16. Dark gray areas indicate regions of cortex where data from all participants were not available and were therefore not included in the map.

19160 | www.pnas.org/cgi/doi/10.1073/pnas.1902650116 Vessel et al.

Dow

nloa

ded

by g

uest

on

Nov

embe

r 18

, 202

0

specificity for architecture and artifacts was primarily observed inprefrontal regions. It is unlikely that these regions of cortex havefunctionality that is specifically relevant for artwork or archi-tecture—a more parsimonious explanation is that these regionssupport the idiosyncratic aspects of personal taste and experi-ence that support aesthetic assessments of cultural artifacts. Thefact that domain specificity for landscapes was primarily ob-served in the ventral visual pathway is in line with the fact thataesthetic appreciation of landscapes was more consistent acrosspeople (SI Appendix, Supplementary Results), and therefore moreclosely related to semantic and structural information present inspecific images.We found a strong brain−behavior correlation between how

strongly an individual participant discriminated between imageslabeled “high” and “low” aesthetic appeal (d’) and the ability topredict aesthetic appeal from the DMN. This finding is re-markable because the classifiers were trained solely with thelabels “high” or “low”—they received no information about theactual rating given by the participant. That the rated distancebetween these 2 sets of trials correlated with classifier perfor-mance indicates that 1) the different use of the rating scale bydifferent participants was not arbitrary, but actually reflected theirpsychological reality, and 2) this psychological difference wasreflected in the discriminability of the associated multivoxel pat-terns in the DMN. A potential consequence is that overall clas-sification accuracy was likely limited by the degree to whichobservers remained engaged with the images over the course ofthe entire experiment (444 images over 2 d). Selection of partic-ipants based on affinity for the domains selected, and presentationof fewer images in a session to fight potential “museum fatigue”(69), would likely result in higher classification accuracies.The poor ability of signal from vMPFC to predict aesthetic

appeal is surprising, given the number of studies in both aes-thetics (36, 44, 47) and decision-making (34, 35, 70) that reportvaluation signals here. There are several possible explanationsfor this discrepancy. The first is the fact that we only selectedvoxels in the vMPFC that were also part of the DMN, as definedby individual maps derived from resting-state fMRI. This likelyresulted in significantly smaller ROIs than a pure anatomicaldefinition of the vMPFC. Second, there is a potential in-consistency in naming conventions across studies. We have usedthe term vMPFC to describe the medial part of the gyrus rectus,lying below the superior rostral sulcus (SRS) and includingBrodmann areas 25 and 14, whereas others may include cortexabove the SRS in their definition of vMPFC. Finally, there is alsothe issue of distortion and dropout in this region of the brain andinconsistent processing of these distortions by different researchgroups. To combat these issues, we used a state-of-the-art mul-tiecho (ME) sequence to recover as much signal as possible fromOFC/vMPFC and correct for distortions in a manner that is lessprone to mixing and spreading of signals across this entire re-gion. Clarification of the exact topography in this region willrequire that all researchers consistently report relevant imagingparameters, such as sequence type, phase-encoding direction,echo time (TE), and the method used for distortion correction.

Aesthetic Appeal vs. Visual Features in the Ventral Visual Pathway.Despite the lack of domain generality in VOT, it is noteworthythat patterns of activation in this region were informative for pre-dicting within-domain judgments of aesthetic appeal. While theventral visual pathway is primarily viewed as important for extractionof visual characteristics of objects and scenes, a number of previousstudies report correlations between signal in the ventral visualpathway and aesthetic appeal (52, 71–73). One possible explanationis that certain visual features are correlated with aesthetic appeal (atleast on average), and it is these visual features that drive the ob-served correlations between brain activity and appeal. Yet activa-tions within VOT (52) and VOT response patterns (this paper)

appear to correlate with aesthetic appeal even for visual artworksand architecture, categories that produce very low interrater agree-ment (e.g., “shared taste”) (32). One possibility is that these regionsdo extract specific visual features, but that observers are differen-tially sensitive to these features. This would maintain a relationshipbetween attention to the feature and aesthetic appeal but also allowfor different observers to express divergent tastes. Alternatively,these regions may not represent stable visual attributes in all ob-servers, but may instead extract more subjective properties of visualexperience that have a positive relationship to aesthetic appeal.In addition to regions in prefrontal cortex and the ventral vi-

sual pathway, the searchlight analysis also identified early visualcortex as containing better-than-chance domain-general classi-fication performance. It is unclear whether this decoding abilitywas a result of bottom-up stimulus-driven differences in activa-tion patterns for high vs. low appeal images or of differential top-down modulation of early visual activity, such as by attention(74) or imagery (75). While the low across-observer agreementfor individuals’ aesthetic ratings (SI Appendix, SupplementaryResults) means that there was substantial overlap at the grouplevel for images shown on “high” trials and “low” trials, thisoverlap was not 100%, leaving room for potential residual dif-ferences in low-level features. As both the natural landscape andarchitecture image sets were contrast-equalized, image contrastis unlikely to be a major factor.

ConclusionsMoving aesthetic experiences are highly integrative. Situated atthe top of the cortical hierarchy, the DMN is in an ideal networkposition to integrate information across many sources. Usinga strong test, we found evidence that the DMN represents aes-thetic appeal in a domain-general manner. In contrast, higher-level visual regions were found to contain only weak and domain-specific information about aesthetic appeal. A searchlightanalysis confirmed the MPFC as part of a putative “core” domain-general system for assessing visual aesthetic appeal, and identifiedadditional domain-specific regions near the frontal poles thatcontained information relevant for aesthetic judgments of artifactsof human culture (architecture, artwork) but not for naturallandscapes. While the exact role of the DMN in aesthetic appealremains unclear, this work confirms that the DMN has access todetailed information about aesthetic appeal, and that this in-formation is not confined to a single node of the DMN.

MethodsParticipants. Eighteen participants were recruited at New York University(NYU) and paid for their participation. Two participants were excluded due toexcessive head motion that led to visible signal distortions and difficultieswith registration, leaving a final group of 16 participants (10 female, 16 right-handed; 25.7 ± 6.3 y). All had normal or corrected to normal vision and nohistory of neurological disorders. Informed consent was obtained from allparticipants. All experimental procedures, including informed consent, wereapproved by the NYU Committee on Activities Involving Human Subjects.

Stimuli. Imageswere presented using back-projection (Eiki LC-XG250 projector)onto a screen mounted in the scanner and viewed through a mirror on thehead coil. Stimulus presentation was controlled by a Macintosh Pro runningOS 10.6 and MATLAB R2011b (Mathworks) with Psychophysics Toolbox-3extensions (http://psychtoolbox.org) (76, 77).Visual art. The set consisted of 148 photographs of visual artworks (paintings,collages, woven silks, excluding sculpture) sourced from the Catalog of ArtMuseum Images Online database (Fig. 1A). A subset of these (109) were usedin a previous study (52). The set covered a variety of periods (fifteenthcentury to the twentieth), styles, and genres (landscape, portrait, abstract,still life), and diversely represented cultures of Europe, the Americas, andAsia. While all of the images were taken from museum collections, specialcare was used to ensure that only lesser-known artworks were included.When necessary, stimuli were cropped to remove the artist’s signature. Dueto the large differences in size and color content across different artworks,contrast equalization was not possible.

Vessel et al. PNAS | September 17, 2019 | vol. 116 | no. 38 | 19161

NEU

ROSC

IENCE

PSYC

HOLO

GICALAND

COGNITIVESC

IENCE

S

Dow

nloa

ded

by g

uest

on

Nov

embe

r 18

, 202

0

Architecture. For architecture, 148 photographs (74 exterior, 74 interior) wereselected, with the majority collected from ArtStor, a database of high-quality images representing multiple cultures and periods (Fig. 1B). Imagescontaining people were excluded. Interior images were chosen to highlightarchitectural detail, not interior décor. An effort was also made to utilizeexterior images that gave an impression of building detail as well as its placein a given setting, while excluding images that gave primary emphasis tofeatures of the landscape. Images represented a variety of structures (e.g.,skyscraper vs. single residence), styles (e.g., Gothic vs. classical), materials,and time periods. Images were cropped to a 4:3 (landscape) or 3:4 (portrait)aspect ratio and presented at a size of 13° of visual angle for the longerdimension. The images were contrast-equalized (SI Appendix, Supplemen-tary Methods) and displayed using a linearized color look-up table.Natural landscapes. For natural landscapes, 148 photographs were obtainedfrom a variety of sources, including the SUN Database (78), IMSI MasterClips,and MasterPhotos Premium Image Collection, and also from images publiclyavailable on the internet (Fig. 1C). Images were cropped to a 4:3 aspect ratioand presented at 13° of visual angle for the horizontal dimension. The im-ages were contrast-equalized (SI Appendix, Supplementary Methods) anddisplayed using a linearized color look-up table.

Procedure. The experiment took place over 2 sessions. Participants wereinstructed in the task and given a short practice (10 trials) to familiarize themwith the types of images they would be seeing, the task timing, and theresponse method. Participants were told that they would be viewing andevaluating images from 3 different aesthetic domains: natural landscapes,architecture, and visual art (Fig. 1 A–C). They were asked to rate how aes-thetically “moving” they found each depicted scene/structure/artwork (SIAppendix, Supplementary Methods).

In each session, participants completed 6 experimental scans composed of37 trials each. Each scan contained images of only a single domain (artwork,natural landscape, interior or exterior architecture) in order to allow participantsto fully engage with this domain over a several-minute period. One sessiontherefore contained 2 artwork scans, 2 landscape scans, 1 interior architecturescan, and 1 exterior architecture scan. Scan order was counterbalanced acrossparticipants, and image order within each scan was pseudorandomized and alsocounterbalanced by assigning alternating participants the reverse order of an-other participant. Each participant saw each image only once. Across the 2 ses-sions, this resulted in a total of 148 trials for each of the 3 domains.

Each trial began with a blinking 1-s fixation cross followed by imagepresentation for 4 s (Fig. 1D). After the image disappeared, a visual “slider”bar appeared on the screen, and participants had up to 4 s to indicate theirresponse on a continuous interval (marked with “L” and “H” at the ends)using a trackball. The mapping between observer movement (up/down) andmovement of the slider (left/right) was counterbalanced across participantsto remove any confounds associated with direction of hand movement. Thiswas followed by a variable intertrial interval drawn from a discrete ap-proximation of an exponential function with a mean of 2.6 s. Each scan alsoincluded 20 s of blank screen at the beginning and 10 s at the end to allowfor better baseline signal estimation and removal of T1 saturation effects.fMRI acquisition and reconstruction. All fMRI scans took place at the NYU Centerfor Brain Imaging (CBI) using a 3T Siemens Allegra scanner with a NovaMedical head coil (NM011 head transmit coil). Whole-brain BOLD signalwas measured from thirty-four 3-mm slices using a custom ME echo-planarimaging (EPI) sequence (2 s repetition time [TR], 80 × 64 3 mm voxels,right-to-left phase encoding, flip angle = 75°). The ME EPI sequence and atilted slice proscription (15° to 20° tilt relative to the anterior commissure–posterior commissure [AC–PC] line) were used to minimize dropout near theorbital sinuses. Cardiac and respiration signals were collected using Biopachardware and AcqKnowledge software (Biopac). We collected a customcalibration scan to aid in ME reconstruction, unwarping, and alignment. MEEPI images were reconstructed using a custom algorithm designed by theNYU CBI to minimize dropout and distortion, and were tested for dataquality (e.g., spikes, changes in signal-to-noise) using custom scripts.

Following the session 1 experimental runs, we collected a high-resolution(1 mm3) anatomical volume (T1 magnetization-prepared rapid gradientecho [MPRage]) for registration and segmentation using FreeSurfer (http://surfer.nmr.mgh.harvard.edu). Following the session 2 experimental runs,participants completed a 360-s eyes-open “rest” scan plus a 320-s visuallocalizer scan that contained blocks of objects, places, faces, and scrambledobjects. The visual localizer scan was fully described in Vessel et al. (52).

Analysis.Identification of individual DMN maps and DMN sub-ROIs. Participant-specificmaps of the DMN were obtained using the rest scan. Motion correction,

high-pass filtering at 0.005 Hz, and spatial smoothing with 6-mm FWHMGaussian filter were applied using the FMRIB Software Library (FSL, https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/). Independent component analysis (ICA) wasthen performed on individual subjects’ scans using FSL’s Multivariate ExploratoryLinear Optimized Decomposition into Independent Components (MELODIC) tool.MELODIC determines the appropriate size of the lower-dimensional space usingthe Laplace approximation to the Bayesian evidence of the model order (79, 80).This process resulted in an average of 21.88 spatial components (SD = 3.70) foreach subject. These ICA components were spatially thresholded at a z-score cutoffof 2.3, moved into MNI standard space, and compared to a set of predefinednetwork maps (81) using Pearson correlation. The component with the highestcorrelation to the Smith et al. (81) DMN map was then visually inspected toensure that its spatial distribution appeared similar to the canonical DMN.For 3 participants, the DMN was split between 2 ICA components similar tothose seen in ref. 48, which were combined to form a single map. The finalDMN ROI for each subject was then defined as the voxels from this com-ponent that also belonged to gray matter (as defined by the FreeSurfer graymatter segmentation).

Volumetric DMN maps were transformed to participant-specific surfacespace. The following 6 subregions of the DMN were then identified in bothhemispheres from the DMN ICA maps of individual subjects by combininganatomically defined boundaries with the participant-specific DMN maps:aMPFC, dMPFC, vMPFC, PCC, IPL, and LTC (see SI Appendix, SupplementaryMethods for additional information).Identification of VOT ROIs. For comparison to the DMN, 3 category-selectiveregions of the VOT were identified using standard methods (SI Appendix,Supplementary Methods).

Finally, an overall VOT ROIwas created by identifying a larger region of theVOT that included all 3 of these category-selective ROIs plus the voxels be-tween them (82). This ROI was first drawn on the FreeSurfer fsaveragecortical surface bilaterally using category-selective ROIs from 3 representa-tive participants as a guide and then projected onto individual hemispheres.The borders of the resulting ROI (tan-colored region in Fig. 2B) extendedfrom the depth of the occipitotemporal sulcus laterally to the center of thelingual/parahippocampal gyrus medially, and from the posterior collateraltransverse sulcus to the anterior collateral transverse sulcus (approximateMNI coordinates y = −72 to y = −31). All ROIs were combined across left andright hemispheres to form bilateral ROIs.ROI classification analysis of aesthetic appreciation. A beta-series general linearmodel (83) implemented using custom MATLAB code was used to extract anestimate of response amplitude for each trial, at each voxel in gray matter.Experimental scans were first preprocessed using FSL to correct for motion,align data across scans, and apply a high-pass filter (0.01-Hz cutoff). Nuisancesignals were then removed from the BOLD time courses of all 12 scans byprojecting out motion estimates and nuisance time series derived from asecond-order Taylor series expansion of cardiac and respiration measure-ments (RetroIcor method) (84). The cleaned time courses were then con-verted to percent signal change. Individual trials were modeled using a 4-s“on” period convolved with a canonical hemodynamic response functionfrom the SPM12 Toolbox (Wellcome Trust Centre for Neuroimaging, UniversityCollege), and the resulting trial-wise amplitude estimates were z-scoredseparately for each scan.

A series of participant-specific classifiers were then trained to distinguish“high” versus “low” aesthetic appreciation trials using the beta-series am-plitude estimates from all voxels in an individual ROI (Fig. 1E). This analysisused the 10 trials with the highest rating (out of 37; top 27%) and the 10 trialswith the lowest rating from each scan (bottom 27%). “Within-domain” classi-fication performance was evaluated by training logistic regression classifiers(logreg.m from PrincetonMVPA Toolbox, https://github.com/princetonuniversity/princeton-mvpa-toolbox) with high and low trials from 3 scans of one domainand then measuring classification accuracy on the high and low trials of 1 held-out scan of the same domain (4-fold cross-validation). The classifier penalty pa-rameter was set to equal 5% of the total number of voxels in an ROI. “Across-domain” classification performance was evaluated by training on high and lowtrials from 3 scans of one domain and then measuring classification accuracy onhigh and low trials in 1 scan from a different domain (again with 4-fold cross-validation). This resulted in a 3-by-3 matrix of classification scores with within-domain scores along the diagonal and across-domain scores off the diagonal.

The 3 within-domain scores along the diagonal were averaged for eachparticipant to produce an overall within-domain score, and the 6 off-diagonalacross-domain scoreswere also averaged to produce an overall across-domainscore. Maximum likelihood estimationwas then used to compute the averageand 95% confidence interval for both of these scores, across all 11 ROIs. AnAUC measure was also computed for each ROI and participant by averagingprobabilistic classifier predictions. Significance of classification performance

19162 | www.pnas.org/cgi/doi/10.1073/pnas.1902650116 Vessel et al.

Dow

nloa

ded

by g

uest

on

Nov

embe

r 18

, 202

0

across all participants was assessed through permutation testing. For eachsubject, 5,000 permutations of the high/low trialwise labels were computedand used to generate a null distribution of the 2 summary statistics in eachROI.We tested the hypothesis that classification performancewas better thanchance (one-tailed) at 2 critical alpha values, 0.05 and 0.005, Bonferroni-corrected for 22 total tests (11 ROIs by 2 statistics).Searchlight analysis.Maps of classification performance were computed acrossthe cortical surface using a “searchlight” approach. For each participant,linear support-vector machine classifiers were trained and tested in a seriesof 5-mm-radius spheres (thirty-three 3-mm voxels) centered on each voxelwhere data were collected. Three within-domain maps and 6 across-domainmaps were created from the average of a 4-fold cross-validation procedure(see above) for each participant. Maps of above-chance performance (>0.5)for the average of the across-domain classifiers (e.g., train on art, test onlandscape) and all 3 within-domain classifiers (art, architecture, landscape)were created for each participant, and then averaged across participants onthe FreeSurfer “fsaverage” surface. The resulting 4 maps were corrected forcomparisons using clusterwise thresholding derived from Monte Carlo sim-ulations (mri_glmfit-sim; ref. 85) with a voxelwise threshold of P < 0.001 anda cluster threshold of 0.05 across the entire cortical surface (2 hemispheres).

Correlation with behavioral discrimination. The distribution of rating responsesdiffered from observer to observer; some generated largely bimodal distri-butions, some uniform, and some strongly unimodal near the neutral point.Therefore, the ratings of the distributions of images subsequently labeled“high” (top 40 of 148 in each domain) and those labeled “low” (bottom40 of 148 in each domain) also differed across observers, with some showinga greater separation than others. In order to quantify this individual vari-ability, a distance measure (d’; e.g., distance between the mean of the2 distributions, rescaled in units of SD) was computed between the distri-butions of raw ratings for “high” and “low” labeled trials for each partici-pant, in each aesthetic domain. The average d’ scores across the 3 domainswere then used in a linear regression analysis as predictors for DMN domain-general classification performance. Data are available at https://dx.doi.org/10.17617/3.2r (86).

ACKNOWLEDGMENTS. We thank Tyra Lindstrom, Lucy Owen, OrianaNeidecker, and Melanie Wen for help with stimulus and data collection;Jean-Remi King for analysis advice; and David Poeppel for feedback on aprevious version of this manuscript. This work was supported by an NYUResearch Challenge Fund Grant to G.G.S. and E.A.V.

1. M. Reimann, J. Zaichkowsky, C. Neuhaus, T. Bender, B. Weber, Aesthetic package design: Abehavioral, neural, and psychological investigation. J. Consum. Psychol. 20, 431–441 (2010).

2. P. Silayoi, M. Speece, The importance of packaging attributes: A conjoint analysisapproach 41, 1495–1517 (2007).

3. U. Ritterfeld, G. C. Cupchik, Environmental psychology perceptions of interior spaces.J. Environ. Psychol. 16, 349–360 (1996).

4. J. H. Falk, Identity and the Museum Visitor Experience (Left Coast Press, Walnut Creek,CA, 2009).

5. P. P. L. Tinio, J. K. Smith, L. F. Smith, “The walls do speak: Psychological aesthetics andthe museum experience” in The Cambridge Handbook of the Psychology of Aes-thetics and the Arts, P. P. L. Tinio, Ed. (Cambridge University Press, 2015), pp. 195–218.

6. I. C. McManus, A. Furnham, Aesthetic activities and aesthetic attitudes: Influences ofeducation, background and personality on interest and involvement in the arts. Br. J.Psychol. 97, 555–587 (2006).

7. D. Haluza, R. Schönbauer, R. Cervinka, Green perspectives for public health: A nar-rative review on the physiological effects of experiencing outdoor nature. Int. J.Environ. Res. Public Health 11, 5445–5461 (2014).

8. A. Clow, C. Fredhoi, Normalisation of salivary cortisol levels and self-report stress by a brieflunchtime visit to an art gallery by London City workers. J. Holist. Healthcare 3, 29–32 (2006).

9. M. L. Chanda, D. J. Levitin, The neurochemistry of music. Trends Cogn. Sci. 17, 179–193 (2013).10. R. S. Ulrich, View through a window may influence recovery from surgery. Science

224, 420–421 (1984).11. R. S. Ulrich et al., A review of the research literature on evidence-based healthcare

design. HERD 1, 61–125 (2008).12. K. Cuypers et al., Patterns of receptive and creative cultural activities and their as-

sociation with perceived health, anxiety, depression and satisfaction with life amongadults: The HUNT study, Norway. J. Epidemiol. Community Health 66, 698–703 (2012).

13. R. Kaplan, The nature of the view from home. Environ. Behav. 33, 507–542 (2001).14. C. I. Seresinhe, T. Preis, H. S. Moat, Quantifying the impact of scenic environments on

health. Sci. Rep. 5, 16899 (2015).15. P. Leather, M. Pyrgas, D. Beale, C. Lawrence, Windows in the workplace: Sunlight,

view, and occupational stress. Environ. Behav. 30, 739–762 (1998).16. S. Shimojo, C. Simion, E. Shimojo, C. Scheier, Gaze bias both reflects and influences

preference. Nat. Neurosci. 6, 1317–1322 (2003).17. J. L. Plass, S. Heidig, E. O. Hayward, B. D. Homer, E. Um, Emotional design in multi-

media learning: Effects of shape and color on affect and learning. Learn. Instr. 29,128–140 (2014).

18. P. J. Silvia, Interest–The curious emotion. Curr. Dir. Psychol. Sci. 17, 57–60 (2008).19. S.-L. Lim, J. P. O’Doherty, A. Rangel, Stimulus value signals in ventromedial PFC reflect

the integration of attribute value signals computed in fusiform gyrus and posteriorsuperior temporal gyrus. J. Neurosci. 33, 8729–8741 (2013).

20. I. Levy, S. C. Lazzaro, R. B. Rutledge, P. W. Glimcher, Choice from non-choice: Pre-dicting consumer preferences from blood oxygenation level-dependent signals ob-tained during passive viewing. J. Neurosci. 31, 118–125 (2011).

21. A. Todorov, C. P. Said, A. D. Engell, N. N. Oosterhof, Understanding evaluation offaces on social dimensions. Trends Cogn. Sci. 12, 455–460 (2008).

22. M. T. Pearce et al., Neuroaesthetics: The cognitive neuroscience of aesthetic experi-ence. Perspect. Psychol. Sci. 11, 265–279 (2016).

23. A. Chatterjee, O. Vartanian, Neuroaesthetics. Trends Cogn. Sci. 18, 370–375 (2014).24. S. Zeki, J. P. Romaya, D. M. T. Benincasa, M. F. Atiyah, The experience of mathematical

beauty and its neural correlates. Front. Hum. Neurosci. 8, 68 (2014).25. S. Zeki, O. Y. Chén, J. P. Romaya, The biological basis of mathematical beauty. Front.

Hum. Neurosci. 12, 467 (2018).26. W. Menninghaus et al., What are aesthetic emotions? Psychol. Rev. 126, 171–195

(2019).27. W. Menninghaus et al., Towards a psychological construct of being moved. PLoS One

10, e0128451 (2015).28. A. Oostendorp, D. E. Berlyne, Dimensions in the perception of architecture: I. Identifi-

cation and interpretatoin of dimensions of similarity. Scand. J. Psychol. 19, 73–82 (1978).

29. J. Meyers‐Levy, R. Zhu, The influence of ceiling height: The effect of priming on thetype of processing that people use. J. Consum. Res. 34, 174–186 (2007).

30. M. R. Greene, A. Oliva, Recognition of natural scenes from global properties: Seeingthe forest without representing the trees. Cognit. Psychol. 58, 137–176 (2009).

31. A. Oliva, P. G. Schyns, Coarse blobs or fine edges? Evidence that information diagnosticitychanges the perception of complex visual stimuli. Cognit. Psychol. 34, 72–107 (1997).

32. E. A. Vessel, N. Maurer, A. H. Denker, G. G. Starr, Stronger shared taste for naturalaesthetic domains than for artifacts of human culture. Cognition 179, 121–131 (2018).

33. H. Leder, J. Goller, T. Rigotti, M. Forster, Private and shared taste in art and faceappreciation. Front. Hum. Neurosci. 10, 155 (2016).

34. O. Bartra, J. T. McGuire, J. W. Kable, The valuation system: A coordinate-based meta-analysis of BOLD fMRI experiments examining neural correlates of subjective value.Neuroimage 76, 412–427 (2013).

35. D. J. Levy, P. W. Glimcher, Comparing apples and oranges: Using reward-specific andreward-general subjective value representation in the brain. J. Neurosci. 31, 14693–14707 (2011).

36. H. Kawabata, S. Zeki, Neural correlates of beauty. J. Neurophysiol. 91, 1699–1705(2004).

37. T. Jacobsen, R. I. Schubotz, L. Höfel, D. Y. Cramon, Brain correlates of aestheticjudgment of beauty. Neuroimage 29, 276–285 (2006).

38. H. Kim, R. Adolphs, J. P. O’Doherty, S. Shimojo, Temporal isolation of neural processesunderlying face preference decisions. Proc. Natl. Acad. Sci. U.S.A. 104, 18253–18258(2007).

39. I. Aharon et al., Beautiful faces have variable reward value: fMRI and behavioralevidence. Neuron 32, 537–551 (2001).

40. A. M. Belfi, E. A. Vessel, G. Gabrielle Starr, Individual ratings of vividness predictaesthetic appeal in poetry. Psychol. Aesthetics Creat. Arts 12, 341–350 (2017).

41. C. Muth, V. M. Hesslinger, C.-C. Carbon, The appeal of challenge in the perception ofart: How ambiguity, solvability of ambiguity, and the opportunity for insight affectappreciation. Psychol. Aesthetics Creativity Arts 9, 206–216 (2015).

42. C. Muth, C. C. Carbon, The aesthetic aha: On the pleasure of having insights intogestalt. Acta Psychol. (Amst.) 144, 25–30 (2013).

43. I. Biederman, E. A. Vessel, Perceptual pleasure and the brain: A novel theory explainswhy the brain craves information and seeks it through the senses. Am. Sci. 94, 247–253 (2006).

44. T. Ishizu, S. Zeki, Toward a brain-based theory of beauty. PLoS One 6, e21852 (2011).45. S. Brown, X. Gao, L. Tisdelle, S. B. Eickhoff, M. Liotti, Naturalizing aesthetics: Brain

areas for aesthetic appraisal across sensory modalities. Neuroimage 58, 250–258(2011).

46. J. Chikazoe, D. H. Lee, N. Kriegeskorte, A. K. Anderson, Population coding of affectacross stimuli, modalities and individuals. Nat. Neurosci. 17, 1114–1122 (2014).

47. T. K. Pegors, J. W. Kable, A. Chatterjee, R. A. Epstein, Common and unique repre-sentations in pFC for face and place attractiveness. J. Cogn. Neurosci. 27, 959–973(2015).

48. J. R. Andrews-Hanna, J. S. Reidler, J. Sepulcre, R. Poulin, R. L. Buckner, Functional-anatomic fractionation of the brain’s default network. Neuron 65, 550–562 (2010).

49. G. Northoff et al., Self-referential processing in our brain–A meta-analysis of imagingstudies on the self. Neuroimage 31, 440–457 (2006).

50. J. M. Moran, T. F. Heatherton, W. M. Kelley, Modulation of cortical midline structuresby implicit and explicit self-relevance evaluation. Soc. Neurosci. 4, 197–211 (2009).

51. A. D’Argembeau et al., The neural basis of personal goal processing when envisioningfuture events. J. Cogn. Neurosci. 22, 1701–1713 (2010).

52. E. A. Vessel, G. G. Starr, N. Rubin, The brain on art: Intense aesthetic experience ac-tivates the default mode network. Front. Hum. Neurosci. 6, 66 (2012).

53. E. A. Vessel, G. G. Starr, N. Rubin, Art reaches within: Aesthetic experience, the selfand the default mode network. Front. Neurosci. 7, 258 (2013).

54. M. D. Fox et al., The human brain is intrinsically organized into dynamic, anti-correlated functional networks. Proc. Natl. Acad. Sci. U.S.A. 102, 9673–9678 (2005).

55. G. L. Shulman et al., Common blood flow changes across visual tasks. 2. Decreases incerebral cortex. J. Cogn. Neurosci. 9, 648–663 (1997).

Vessel et al. PNAS | September 17, 2019 | vol. 116 | no. 38 | 19163

NEU

ROSC

IENCE

PSYC

HOLO

GICALAND

COGNITIVESC

IENCE

S

Dow

nloa

ded

by g

uest

on

Nov

embe

r 18

, 202

0

56. J. R. Simpson , Jr, A. Z. Snyder, D. A. Gusnard, M. E. Raichle, Emotion-induced changesin human medial prefrontal cortex: I. During cognitive task performance. Proc. Natl.Acad. Sci. U.S.A. 98, 683–687 (2001).

57. P. Qin, G. Northoff, How is our self related to midline regions and the default-modenetwork? Neuroimage 57, 1221–1233 (2011).

58. F.-X. Neubert, R. B. Mars, J. Sallet, M. F. S. Rushworth, Connectivity reveals relation-ship of brain areas for reward-guided learning and decision making in human andmonkey frontal cortex. Proc. Natl. Acad. Sci. U.S.A. 112, E2695–E2704 (2015).

59. D. S. Margulies et al., Situating the default-mode network along a principal gradientof macroscale cortical organization. Proc. Natl. Acad. Sci. U.S.A. 113, 12574–12579(2016).

60. A. M. Belfi et al., Dynamics of aesthetic experience are reflected in the default-modenetwork. Neuroimage 188, 584–597 (2019).

61. A. R. Vaidya, M. Sefranek, L. K. Fellows, Ventromedial frontal lobe damage alters howspecific attributes are weighed in subjective valuation. Cereb. Cortex 28, 3857–3867(2018).

62. A. Abraham, The world according to me: Personal relevance and the medial pre-frontal cortex. Front. Hum. Neurosci. 7, 341 (2013).

63. R. N. Spreng, C. L. Grady, Patterns of brain activity supporting autobiographicalmemory, prospection, and theory of mind, and their relationship to the default modenetwork. J. Cogn. Neurosci. 22, 1112–1123 (2010).

64. G. G. Starr, “Multisensory imagery” in Introduction to Cognitive Cultural Studies, L.Zunshine, Ed. (The Johns Hopkins University Press, Baltimore, MD, 2009), pp. 1–29.

65. S. L. Fairhall, A. Ishai, Neural correlates of object indeterminacy in art compositions.Conscious. Cogn. 17, 923–932 (2008).

66. A. Ishai, S. L. Fairhall, R. Pepperell, Perception, memory and aesthetics of in-determinate art. Brain Res. Bull. 73, 319–324 (2007).

67. E. A. Vessel, N. Rubin, Beauty and the beholder: Highly individual taste for abstract,but not real-world images. J. Vis. 10, 1–14 (2010).

68. A. Schepman, P. Rodway, S. J. Pullen, J. Kirkham, Shared liking and association va-lence for representational art but not abstract art. J. Vis. 15, 11 (2015).

69. B. I. Gilman, Museum fatigue. Sci. Mon. 2, 62–74 (1916).70. F. Grabenhorst, E. T. Rolls, Value, pleasure and choice in the ventral prefrontal cortex.

Trends Cogn. Sci. 15, 56–67 (2011).71. O. Vartanian, V. Goel, Neuroanatomical correlates of aesthetic preference for

paintings. Neuroreport 15, 893–897 (2004).

72. A. Chatterjee, A. Thomas, S. E. Smith, G. K. Aguirre, The neural response to facialattractiveness. Neuropsychology 23, 135–143 (2009).

73. X. Yue, E. A. Vessel, I. Biederman, The neural basis of scene preferences. Neuroreport18, 525–529 (2007).

74. D. C. Somers, A. M. Dale, A. E. Seiffert, R. B. Tootell, Functional MRI reveals spatiallyspecific attentional modulation in human primary visual cortex. Proc. Natl. Acad. Sci.U.S.A. 96, 1663–1668 (1999).

75. S. H. Lee, D. J. Kravitz, C. I. Baker, Disentangling visual imagery and perception ofreal-world objects. Neuroimage 59, 4064–4073 (2012).

76. D. H. Brainard, The psychophysics toolbox. Spat. Vis. 10, 433–436 (1997).77. D. G. Pelli, The VideoToolbox software for visual psychophysics: Transforming num-

bers into movies. Spat. Vis. 10, 437–442 (1997).78. J. Xiao et al., Basic level scene understanding: Categories, attributes and structures.

Front. Psychol. 4, 506 (2013).79. C. F. Beckmann, S. M. Smith, Probabilistic independent component analysis for func-

tional magnetic resonance imaging. IEEE Trans. Med. Imaging 23, 137–152 (2004).80. T. P. Minka, Automatic choice of dimensionality for PCA. Adv. Neural Inf. Process Syst.

13, 598–604 (2000).81. S. M. Smith et al., Correspondence of the brain’s functional architecture during ac-

tivation and rest. Proc. Natl. Acad. Sci. U.S.A. 106, 13040–13045 (2009).82. U. Hasson, M. Harel, I. Levy, R. Malach, Large-scale mirror-symmetry organization of

human occipito-temporal object areas. Neuron 37, 1027–1041 (2003).83. J. Rissman, A. Gazzaley, M. D’Esposito, Measuring functional connectivity during

distinct stages of a cognitive task. Neuroimage 23, 752–763 (2004).84. G. H. Glover, T. Q. Li, D. Ress, Image-based method for retrospective correction of physi-

ological motion effects in fMRI: RETROICOR. Magn. Reson. Med. 44, 162–167 (2000).85. D. J. Hagler , Jr, A. P. Saygin, M. I. Sereno, Smoothing and cluster thresholding for

cortical surface-based group analysis of fMRI data. Neuroimage 33, 1093–1103 (2006).86. E. A. Vessel, A. I. Isik, A. M. Belfi, J. L. Stahl, G. G. Starr, The default-mode network

represents aesthetic appeal that generalizes across visual domains. Max Planck Soci-ety. https://dx.doi.org/10.17617/3.2r. Deposited 20 August 2019.

87. H. Regnauld, Seated African Woman, 1860s, oil on fabric, bequest of Noah L. Butkin,1980.280, The Cleveland Museum of Art, Cleveland.

88. J. Wright of Derby, Cottage on Fire, ca. 1786-1787, oil on canvas, The Putnam DanaMcMillan Fund and bequest of Lillian Malcolm Larkin, by exchange, 84.53, Minne-apolis Institute of Arts, Minneapolis.

19164 | www.pnas.org/cgi/doi/10.1073/pnas.1902650116 Vessel et al.

Dow

nloa

ded

by g

uest

on

Nov

embe

r 18

, 202

0