Localizing Brain Function Based on Full Multivariate Activity Patterns: The Case of Visual Perception and Emotion Decoding

Multivariate statistics and machine learning methods have become a common tool to extract information represented in the brain. What is less recognized is that, in the process, it has become more difficult to perform data-driven discovery and functional localization. This is because multivariate pattern analysis (MVPA) studies tend to restrict themselves to a subset of the available data, or because sound inference to map model parameters back to brain anatomy is lacking. Here, we present a high-dimensional (including brain-wide) multivariate classification pipeline for the detection and localization of brain functions during tasks. In particular, we probe it at visual and socio-affective states in a task-oriented functional magnetic resonance imaging (fMRI) experiment. Classification models for a group of human participants and existing rigorous cluster inference methods are used to construct group anatomical-statistical parametric maps, which correspond to the most likely neural correlates of each psychological state. This led to the discovery of a multidimensional pattern of macroscale brain activity which reliably encodes for the perception of happiness in the visual cortex, lingual gyri and the posterior perivermian cerebellum. We failed to find similar evidence for sadness and anger. Anatomical consistency of discriminating features across subjects and contrasts despite the high number of dimensions suggests MVPA is a viable tool for a complete functional mapping pipeline, and not just the prediction of psychological states.


Introduction
The mapping of segregated brain functions is far from a settled methodology. Take for instance the choice of statistical model in task-oriented functional neuroimaging for modalities like fMRI and PET: while the venerable mass-univariate analysis fits separate models to encode each brain time series based on experimental variables, multivariate pattern analysis (MVPA) reverses this -particularly multivariate pattern learning -fitting a single model to decode experimental conditions out of the joint activity of several brain signals. The former is excellent at uncovering simple correlations between loci and functions, whereas the latter can potentially provide increased sensitivity due to emergent informational dependencies (Wang et al, 2007;Huettel et al., 2009;Jolly and Chang, 2021), at the expense of computational complexity (De Martino et al., 2008).
In this vein, popular wisdom often dictates that multivariate searches should be limited in one way or another, via dimensionality reduction or restriction to regions of interest (ROI) (Cox and Savoy, 2003;Haynes and Rees, 2005;Kamitani and Tong, 2005) or moving searchlights (Kriegeskorte et al., 2006;Björnsdotter et al., 2011). Furthermore, unlike with the classical mass-univariate approach, the ability to map model parameters onto anatomy may not always be possible, depending on the transparency of the algorithm and its nonlinearities (Schmah et al., 2010).
Paradoxically, there is an enormous interest in studying how the holistic interactions of large-scale brain networks might subserve complex behaviors and mental processes, not just at rest but during tasks (Prinz, 2006;Chialvo, 2010;Avena-Koenigsberger et al, 2018). In fact, most of the recent breakthroughs in statistical learning have been achieved by embracing the high-dimensional nature of problems (Sejnowski, 2020), and safeguards against exaggerated (i.e., over-fitted) findings, such as stringent observation of cross-validation and ROC-curve standards, are now commonplace (Mahmoudi et al., 2012). Moreover, a brain-wide, mappable and statistically sound methodology for functional localization would complete the spectrum of available complementary analytical perspectives purported since the introduction of MVPA (Raizada et al., 2010, Gaonkar andDavatzikos, 2013;Jolly and Chang, 2021). Just as allowing for multivariate dependencies in the first place might provide a whole new picture (Wang et 49 50 al., 2007;Huettel et al., 2009;Mahmoudi et al., 2012;Lewis-Peacock and Norman, 2013), so could multivariate data drawn from different spatial (or temporal) scales (Hoel et al., 2013;Jolly and Chang, 2021). An auxiliary motivation simply comes from trying to avoid selection bias (Kriegeskorte et al., 2009) and the caveats of dimensionality reduction (Kherif et al., 2002;Wang et al., 2007;Palo et al., 2021).
Yet the merits of such relatively straightforward analysis have seldom -if ever -been fully put to test.
Pieces are scattered here and there in the literature: one year after introducing the use of general linear models (GLM) for per-voxel analysis (Friston et al. 1994), Friston et al. (1995 applied multivariate analysis of covariance on whole-brain data, reduced to a space of 35 eigenimages. Ever since, the existing whole-brain experiments either keep constructing a low-dimensional state-space from such multivariate methods like principal components analysis (McIntosh et al., 1996;Mørch et al., 1997;McKeown et al., 1998;Kjems et al., 2002;Kherif et al., 2002;Carlson et al., 2003;LaConte et al., 2003;Wang et al, 2007) or brain atlases and parcellations (Zhang et al., 2020). Another approach is to select individual voxels through (admittedly self-defeating) univariate (Polyn et al. 2005;De Martino et al., 2008;Mwangi et al., 2014) or multivariate means (Craddock et al., 2009). A few works have performed brain-wide multivariate searches without dimensionality reduction, while also failing to translate machine learning models into proper statistical parametric maps (Mourão-Miranda et al., 2005;LaConte et al., 2005;Polyn et al. 2005;Schmah et al., 2010;Raizada et al., 2010). Raizada et al. (2010) correctly identified and tried to fill the gap, by using GLM in a truly multivariate way. The approach proved promising keeping track of behaviorally-separable linguistic groups, although correction for multiple comparisons was absent in their final statistical-anatomical maps. Gaonkar and Davatzikos (2013) explicitly identified the gap again, and showed how to analytically approximate null models so that statistical significance maps can be derived from model parameters in the case of support vector machine (SVM) classifiers, which they applied to a lie-detection fMRI dataset. To the advantage of the multivariate approach, the False Discovery Rate of statistical maps was shown to be much lower than in the univariate methodology, but this post-hoc discovery was not incorporated to correct brain maps for multiple comparisons.
Here we tested pattern classification analysis as a methodology for anatomical localization in brain-wide, high-dimensional fMRI data. We employed linear SVM classifiers -a relatively interpretable multivariate algorithm, from which a spatial discriminative map can be extracted and directly related to the effect of different experimental conditions on brain activation (Mourão-Miranda et al., 2005;LaConte et al., 2005;Wang et al, 2007;Gaonkar and Davatzikos, 2013). This was tested on a sample of 16 human participants who performed a face perception task. Our goal is to provide anatomical statistical significance maps which are properly corrected for multiple comparisons, and for this we employ a stateof-the-art cluster-informed statistic well-known to the neuroimaging community. Moreover, our task is designed to probe the analysis against variable levels of decoding difficulty: from simple visual stimulation to face detection, to the more ethereal perception of three different emotions. We evaluated whether individual classifiers can learn to predict task state above empirically-estimated chance performance, and whether individual models converge on what the most relevant neural correlates of each cognitive ability are at the group level. We also provide comparisons to the classical mass-univariate (GLM) analysis.
The analysis is expected to reveal well-known early visual cortical areas in contrasts intended to capture the effect of visual stimulation, and components of the so-called face processing network in the ventral stream during face perception (Haxby et al., 2000;Haist and Anzures, 2017). If successful, this would provide greater confidence when pitching the method against a genuinely harder and open problem; namely, segregating and identifying emotions in the brain.
Although affective neuroscience has been fruitful in identifying the anatomical components of the emotional peripheral and central nervous systems; ample disagreement still exists on how to physiologically characterize particular emotional experiences, even among meta-analytical reviews (Vytal and Hamann, 2010;Lindquist et al., 2012;Hamann, 2012;Kragel and LaBar, 2014;Guillory and Bujarski, 2014;Kragel and LaBar, 2016;Celeghin et al., 2017). It is unclear how the distributed activity of many limbic and other mid-line structures, from the posterior perivermian cerebellum to the medial

Image acquisition
Images were obtained from a 3-Tesla General Electric Discovery MR750 scanner at the Magnetic Resonance Unit at UNAM campus Juriquilla, during a single session per participant. The protocol included five echo-planar imaging (EPI) blood-oxygenation-level-dependent (BOLD) sequences for fMRI, 185 volumes each. A T1-weighted (T1w) scan of head anatomy was also acquired. Sequence parameters are described in Table 3. Electromagnetic responses were recorded using a head-mounted 32channel coil.  8 Figure 1: Raw samples of both image modalities for a single subject in our dataset (in the same order as the columns in Table 3).

Stimuli and task
Each of the five fMRI sequences was temporally coupled to a psychological block-based task implemented in PsychoPy 3.0.1 (Peirce, 2007). All five tasks were identical, save for the pseudo-random order in which their 30 s blocks were administered. A total of six block classes were used: happy faces, sad faces, angry faces, neutral faces, pseudo (scrambled) faces and low-stimulation. Neutral/inexpressive faces might provide an extra control when contrasting among emotions. Pseudo-faces and dim blocks were introduced so as to buttress and diagnose the analysis pipeline, by way of more trivial contrasts (like pseudo-faces vs low-stimulation and faces vs pseudo-faces).
Each block, in turn, comprises 10 randomly-presented images belonging to that class, each one shown for about 3 s and without possibility of re-instantiation during the same block. Each block occurs twice per sequence, yielding a total of 12 of them (360 s = 6 min). After their presentation, participants had to wait for 10 s before concluding the sequence, in order to capture the hemodynamic response (HR) elicited by the last stimuli. A selection of 10 gray-scale photographs per category of frontal human faces (male and female) served as stimuli. These were chosen from the "Pictures of Facial Affect" database (Ekman, 1976). As for the low-stimulation (a.k.a. "dim") blocks, a small but visible fixation cross was made to fluctuate from quadrant to quadrant at random every 3 s. The whole task is summarized in Figure 2.
Additionally, behavioral responses were recorded throughout the task in order to measure performance and thus evaluate the suitability of physiological data for further analysis. Participants were given an orthogonal task: to indicate whether faces belonged to a man or a woman as soon as they were perceived.
The response was submitted with the press of a button, one in each hand. Analogously, for scrambled and dim blocks (when no faces should have been perceived), the instruction was to simply report image change, alternating between buttons. In this fashion, motor activity remained rather homogeneous for all blocks, minimizing a possible confounding effect when contrasting among faces and pseudo-faces.
Statistical analysis of behavioral data was conducted using the R programming language.

Code and data accessibility
All the software necessary for reproducing, analyzing or adapting the present study is made available as free software and may be downloaded at https://github.com/isacdaavid/np-mvpa . Original data and final group activation maps in standard space may be downloaded respectively from OpenNeuro (

Image preprocessing
T1w images were submitted to the volBrain tissue segmentation and volumetry web-based service (Manjón and Coupé, 2016). The resulting brain and gray/white matter masks we used for deskulling the field-bias-corrected T1w images and, later on, for selection of fMRI voxels. fMRI sequences were concatenated by temporal order in one long sequence per subject, then the result underwent the following preprocessing pipeline due to the FSL 6.0 utilities (Jenkinson et al., 2012): high-pass frequency filter (>50 s) and interpolation for slice-time correction (interleaved acquisition) (Woolrich et al., 2001), affine movement correction and co-registration (Jenkinson and Smith, 2001;Jenkinson et al, 2002) with the respective T1w anatomical reference and the standard MNI-152 T1w template (Fonov et al., 2009;Fonov et al., 2011) at 1 mm of resolution. After registration, the corresponding resulting matrices were applied to the volBrain masks, so as to transform them to the low-resolution subject-space of fMRI images. Gray matter time series were extracted afterwards (about 10000 time series, depending on the subject), and linear trends were subtracted by preserving residuals from a simple linear regression performed on each of the five sequences that comprise the long concatenated time series. Finally, and seeking not to bias classification models in any dimension, the composite time series at each voxel was normalized to zscores.

Univariate analysis
To assess the feasibility and performance of the brain-wide multivariate approach against the golden standard in functional brain mapping, we investigated the same two-way contrasts included in multivariate analysis using FSL 6.0 in a classical mass-univariate analysis. These contrasts are grouped into visual stimulation ("dim vs scrambled", "dim vs neutral", "dim vs angry", "dim vs sad", "dim vs happy"), face perception ("scrambled vs neutral", "scrambled vs happy", "scrambled vs sad", "scrambled vs angry") and emotion perception ("angry vs happy", "sad vs happy", "sad vs angry", "angry vs neutral", "happy vs neutral", "sad vs neutral"). Preprocessed data (up until subtracting linear trends and normalizing) were spatially-smoothed with a Gaussian convolution kernel of 5 mm FWHM. Each of the six block classes described for the task was considered as a column-vector regressor in the design matrix , after convolving them with a zero-lag, double-gamma HR curve. was augmented with the time derivatives of each convolved regressor, but no motion covariates were added. GLMs were fitted afterwards. GLM is a matrix-form extension to multiple linear regression, which models each physiological time series (column ) as a linear combination of plus some Gaussian error . The model reads (Mahmoudi et al., 2012): Assuming trial independence and homocedasticity, maximum likelihood estimation or ordinary leastsquares estimation may be followed to obtain the so-called normal equation, which optimizes parameters according to:^= After estimation, contrasts of parameter estimates are subjected to the same procedure as the multivariate model parameters for purposes of group-level inference. This is described in more detail in the 2.4.5 Group inference subsection.

Multivariate pattern analysis
Once preprocessed as described in subsection 2.4.2, multivariate decoding of fMRI patterns was conducted using the pyMVPA 2.5.0 Python library (Hanke et al., 2009). We trained a linear SVM classifier per subject and block contrast combination using all available brain volumes (120 volumes per class for the training phase, 30 for testing). In addition to the contrasts described in the previous section, emotion-related ones were augmented with the ( 4 3 ) possible three-way classification problems and the single four-way contrast. The supervised SVM algorithm learns a hyperplane for binary classification in high-dimensional state space (Vapnik and Chervonenkis, 1974;Boser et al., 1992). Given a vector orthogonal to the hyperplane, the SVM decision rule is equivalent to the sign of the projection of unseen data vectors i on , adding or subtracting the necessary constant b so as to make the result exactly 0 at the hyperplane: Out of all possible hyperplanes, SVM's key insight is to estimate the one that maximizes separation margin to the most difficult training data: the support vectors right above opposite margin lines. Since margin width can be calculated from pairs of positive-class and negative-class support vectors according to: by constraining the decision rule to satisfy | ⋅ i +b | ≥ 1 or similar criteria and substituting on equation (4), one can show that maximizing the margin -and therefore obtaining an optimal model -is equivalent to a quadratic programming problem with ‖ ‖ as the cost function to be minimized We also explored the effect of different delays between stimulus onset and volume labeling (from 0 s to 10 s, every 2 s), instead of assuming a single optimal HR peak (Lewis-Peacock and Norman, 2013); although, based on common practice and prior knowledge about typical HR, a labeling delay of 4 s was fixed a priori when conducting all group-level statistical inferences.

Group inference
To assess whether individual models reveal generalizable anatomical structures at the group level we employed yet another non-parametric test, one per contrast. In particular, we used FSL 4.0's randomise (Winkler et al., 2014) with 5000 rounds of the Threshold-Free Cluster Enhancement (TFCE) (Smith and Nichols, 2009) operating on the group of SVM weight vectors to generate null data and then perform a two-tailed test (models were spatially smoothed in advance with a 5 mm FWHM Gaussian kernel, and to make them comparable, they were transformed to the standard 1 mm MNI-152 space, as well as normalized so that all weight vectors become unitary). In the case of the univariate models, the only difference is that TFCE receives a group of GLM contrasts of parameters as input data (i.e., also in a twotailed test with both positive and negative effects, prior spatial smoothing and transformation to the 1 mm MNI-152 space).
Given an anatomical image h(v) with scalar values representing model parameters (or contrasts among them), the TFCE statistic at some voxel v is defined as the integral (in the Lebesgue sense) of cluster size s(v, h) times the cluster-defining "height" h: In practice, equation (6) is modified for fMRI and EEG data, where the default is to favor h, squaring it, while taking the square root of s (Smith and Nichols, 2009). Since all possible cluster-forming thresholds are considered by the integral, TFCE is generally regarded as a more principled alternative to other nonparametric cluster-informed inference methods, while still providing strong family-wise error (FWE) control (Roiser et al., 2016), as expected of nonparametric approaches. For the sake of visualization, we set a cluster cutoff value of 0.01 in the corrected p-value brain maps to render voxels that are unlikely to correlate with the task by pure chance.

Behavior
See the attached supplementary material.

Visual stimulation and face perception
All contrasts meant to distinguish between high and low visual stimulation and between face and pseudo-  Both univariate and multivariate approaches agree on two prominent, bilaterally-symmetric occipital clusters whose activity correlates with the presentation of visual stimuli (see Figure 4 and Table 4): one showed suprathreshold evidence at FFA (also see Figure 4 and Table 4).   It is suggested from Figure 5  On the other hand, classification models that excluded this emotion did not (sad vs neutral: p=0.38 ,  Group-level inference of SVM parameters resulted in suprathreshold clusters with high co-localizationnot just among subjects within the same contrast, but even among contrasts -with two possible characteristic anatomical spans, depending on whether neutral faces had been included in the classification problem. In their absence, detection of the happy faces consistently depended on activity at the occipital pole and its midline and ventral surroundings (posterior V1, posterior V2, ventral V3, V4/ V8) as well as anterior V1 and V2 (both dorsal and ventral). This is shown in yellow clusters in Figure 7.

Emotion perception
When also faced with the neutral control stimuli (cyan clusters in Figure 7), SVM was forced to extend the search to lower-order and higher-order structures: the pulvinar thalamic bodies, the anterior lingual gyrus (LG) -close to parahippocampal tissue -and the perivermian posterior cerebellar lobe (PPCL).
See table 4 for a summary of clusters and minima, together with their stereotactic coordinates. Although not shown, the most prominent subthreshold evidence ( p F W E <.05) was found at the left amygdala and LOC. Subject-level SVM models also gave prominent weighting to the orbitofrontal/ventromedial prefrontal cortex, however, the amount of clusters and their parameter signs in that region were too heterogeneous for evidence to accumulate at the group-level.
By contrast, discrimination of sadness and/or anger (whose decodability was considerably worse, as shown in Figures 5 and 6) resulted in virtually nonexistent anatomical clusters: the left pulvinar body in the thalamus (angry vs neutral) and posterior V1/V2 and PPCL (sad vs neutral). These are rendered in the three bottom rows of Figure 7. Remarkably, no suprathreshold cluster or voxel, correlated or anticorrelated, was found for any of the emotion-related contrasts using classical mass-univariate analysis.

Discussion
Results validate the feasibility of looking for multivariate correlations between functional neuroimaging and perceptual phenomena of varying complexity, and of turning learned data patterns into statisticalanatomical maps for localization of relevant brain structures. This was shown for an extrinsically highdimensional state space using all the available gray-matter information, which differs substantially from the routes taken by both univariate analysis and other multivariate studies. It is remarkable that an algorithm as modest as the kernel-less SVM can characterize many psychological states of scientific interest in a purely data-driven approach, to the point of surpassing the classical methodology for its brain mapping features.
Results for the simple visual stimulation subexperiment are almost identical according to both approaches ( Figure 4). However, remaining contrasts decidedly favored the brain-wide multivariate analysis.
Univariate analysis notoriously failed to consistently discover the right FFA for face perception contrasts, and in terms of the emotions carried by faces, it could not find a single correlated voxel. This suggests that the increased sensitivity of multivariate algorithms is also preserved at the brain-wide level, even in the absence of dimensionality reduction.
That said, it might still seem tempting to disparage results for the emotional subtask, due to major reliance on visual, rather than emotion-related areas other than the perivermian posterior cerebellar cortex (Schmahmann and Sherman, 1997) and the lingual gyrus (which has been involved both in complex visual processing and in emotional and face processing: Isenberg et al, 1999;Kehoe et al, 2013); although as mentioned in the Results section, SVM also relied on the dynamics of the amygdala and the ventromedial prefrontal cortex to drive a decision, albeit less prominently. However, the fact that at least one emotion class (happiness) reliably elicits a distributed activity fingerprint -which was invisible to GLM in the first place -speaks of the advantages of moving beyond simple mass-univariate modeling when it comes to not only neural decoding but also brain mapping. Whether this particular "happy interlocutor state" is truly a non-collateral biological feature of social significance is hard to answer with our data. On one hand, the connectivity and modulation of core affect regions upon the visual cortex has been well-attested by independent studies (Amaral and Price, 1984;Damaraju et al., 2009) and metaanalytic reviews of emotion (Vytal and Hamann, 2010;Lindquist et al., 2012). Moreover, recent experiments using electrophysiological and calcium-imaging techniques on rodents have emphasized the presence of notorious motor and arousal-related information in areas traditionally thought of as sensory (Vinck et al., 2015;Stringer et al., 2019). On the other hand, it may be argued from a conservative take on systems neurophysiology that primary visual areas are not particularly concerned with constructing face or affect percepts. Nonetheless, the classifier could have picked up and leveraged lower-order information in those areas to construct a statistical model about facial expression; similar to how artificial vision systems emulate cortical computations starting from nothing but raw pixels. That would certainly pose a methodological challenge to our approach (sensitivity in excess can be detrimental), which we showed was alleviated to some extent by the diligent use of control stimuli (neutral faces).
This result is of great interest, in light of the incipient works on emotion as seen through the MVPA prism. For instance some of the literature from table 1 also included anger and sadness-loaded stimuli in their experiments, reportedly with better results than ours (Ethofer et al., 2009;Said et al., 2010;Kotz et al., 2012). However, reasons exist to be skeptical of them upon closer inspection. The ROI-based, auditory study by Ethofer et al., for example, reported average classification accuracies (n=22) of 30% and >35% for sadness and anger respectively; among 5 emotions. Nonetheless, models were trained only pair-wisely: that is, contrasting target emotion against an "everything else" metaclass, which implies that performance was actually below the true chance level (50%). This one-vs-all scheme without nonparametric testing was repeated by Kotz et al., yet, here anger showed the poorest results. Other issues in the literature include comparing against a null model built from a scarce number of permutations, for instance, like in the study conducted by Said et al (2010).
Perhaps our results for these two basic emotions could have improved, had a more localized search been performed. Those affective-perceptual states might be genuinely underrepresented in the coarse fMRI data, or they may not be linearly separable, or the system dynamics may not be sufficiently stationary in the relevant dimensions. As argued during the Introduction, this study aimed at testing the limits of linear SVM as a data-driven anatomical mapping tool, at the expense of maximum decoding performance. In that sense, and joined by our modest sample size, being able to retrieve just some emotional states out of BOLD activity emanating from well-defined structures already cements the accomplishment of our goals.
The present work also suffers from limitations and future opportunity areas. Further analysis is required to characterize the intrinsic dimensionality of each cluster system. Similarly, system dynamics could be studied and modeled to provide further understanding of each successfully decoded state, as well as the encompassing attractor set. It would also be interesting to extend the task to other emotions, modalities and theoretical models of emotion; for instance to identify whether we have a sufficient characterization of happiness (as opposed to appetitive hedonic valence more generally, as posited by dimensional theories of emotion). A second strand of further studies could explore these findings using more direct causal interventions in the brain, so as to assess the relevance of the multivariate correlations we report. rates turn out to be similar to those reported in independent validations of other datasets (Tottenham et al., 2009;Conley et al., 2018). Similarly, per-participant χ 2 tests (with Bonferroni correction for FWE) revealed that even the worst-performing participant had a probability of less than 5 ⋅10 − 8 of being involved in guesswork. With regard to instantaneous responses during the task, we ran binomial tests to quantify success probability detecting face gender or image change, assuming statistical independence and a chance level of 50%. Figure 2 shows the aggregate of hits through time. Errors, in red, are comparatively low. A probability of 1.95 ⋅10 −60 (Holm-corrected) of finding such hits/misses ratio by chance was found for the worst participant, and the probability for the worst block type (including pseudo-faces and dim-stimulation) is even smaller. Participant's reaction times (RTs) were analyzed as well, as a measure of attention to the task. Each curve in Figure 2 corresponds to the RTs of some subject. The superimposed dotted black line projects the relevant part of a general linear mixed-effects model (GLMM). GLMM is a generalization of GLM regression which uses two -as opposed to one -design matrices to account for random effects. This is specially suitable to hierarchical factorial designs; since the variance of measurements at some time t could come from intrinsic differences among participants, whose personal variance is captured by the random effects. The model was fitted using the participant factor as a random effect, and block lapse and block class as fixed effects. The aim is to find the effect of time upon RTs, because big changes would signify disengagement from the task. On the contrary, we observed a negligible, downward slope (0.11 ms faster RTs every 30 s block).
Moreover, a post-hoc Tukey test for a one-way ANOVA of RTs was inspected, using block types as factor levels. The only emotion to elicit considerably different reaction times was anger (angry vs neutral p=.014, angry vs happy p=.03). No significant difference was found between reacting to pseudo-faces vs to fixation crosses. However, we measured extremely large differences between reacting to any type of visuofacial stimulus and any type of non-visuofacial stimuli, for which not only stimulus complexity is lower, but task complexity is also lower (telling gender vs noticing any change at all).
All these lines of behavioral evidence converge towards the conclusion that participants understood the task and that stimuli were correctly observed in general. Accordingly, no participant or block type was discarded for analysis of the fMRI data after this screening.