Awe is represented as an ambivalent experience in the human behavior and cortex

754 

Abstract Submission 

Jinwoo Yi¹, Danny Han², Jiook Cha^1,2,3,4

¹Department of Brain and Cognitive Sciences, Seoul National University, Seoul, Korea, Republic of, ²Interdisciplinary Program of Artificial Intelligence, Seoul National University, Seoul, Korea, Republic of, ³Department of Psychology, Seoul National University, Seoul, Korea, Republic of, ⁴Graduate School of Data Science, Seoul National University, Seoul, Korea, Republic of

Jinwoo Yi  
Department of Brain and Cognitive Sciences, Seoul National University
Seoul, Korea, Republic of

Danny Han  
Interdisciplinary Program of Artificial Intelligence, Seoul National University
Seoul, Korea, Republic of

Jiook Cha  
Department of Brain and Cognitive Sciences, Seoul National University|Interdisciplinary Program of Artificial Intelligence, Seoul National University|Department of Psychology, Seoul National University|Graduate School of Data Science, Seoul National University
Seoul, Korea, Republic of|Seoul, Korea, Republic of|Seoul, Korea, Republic of|Seoul, Korea, Republic of

Awe encompasses a complex emotional experience, transitioning from the initial negative feelings elicited by immense objects to subsequent positive ones (Keltner & Haidt, 2003). Nevertheless, the ambivalent nature of awe was not fully investigated due to its incompatibility with the prevailing framework of affective science (e.g., the constructionist approach). To address this knowledge gap, we delved into the following questions: (1) Does self-reported ambivalent states during awe experience better explain variances in awe intensity than simple positive/negative ratings? (2) Is the ambivalent state of awe also represented in the neural space as a distinctive valence state? We hypothesized that the self-reported ambivalence would predict awe intensity more accurately than positive/negative ratings. We also posited an existence of neural representation specific to ambivalent state during awe experiences.

We recorded a 19-channel electroencephalogram (EEG) while participants (N = 43) watched four 360° video stimuli: three awe-conditioned and one control. In each trial, they watched a clip and continuously rated subjective valence (i.e., positive, negative, neutral, and mixed) via key press. After the trial, they reported Awe Experience Scale (AWE-S) (Yaden et al., 2019), Evaluative Space Grid for valence (Larsen et al., 2009), and Likert-scaled arousal and motion sickness items. First, we examined which ratings significantly predicted AWE-S and how precisely they do so through a linear mixed model ('behavioral analysis'). Second, by applying contrastive learning specialized for time series data (CEBRA)(Schneider, Lee & Mathis, 2023), we constructed latent neural spaces that embed key-pressed valence labels in a within-subject design. To test whether each valence state could be distinctively clustered in the neural space, we calculated silhouette scores for subject's each trial and performed 1,000 permutation tests to assess their significance ('clustering analysis'). Third, we decoded the keypress valence labels using EEG time series to evaluate idiosyncrasy of neural representations linked to each valence state. We fitted a kNN classifier, the simplest model that does not require any training, with the out-of-bag samples from clustering analysis to predict key-pressed valence label in within-individual approach. We calculated the weighted F1 score using each sample's last 15% time-series as test set. To compare the performance with random chance, we also computed a weighted F1 score with a permuted dataset ('decoding analysis').

In behavioral analysis, we found that only the intensity of positivity (β = .094, p = .035) and ambivalence (β = .220, p = .001) rated in Evaluative Space Grid and "mixed" keypress duration (β = .565, p = .039) were significantly associated with the AWE-S scores when controlling for random effects of participants and clips. A linear mixed model with these three predictors explained 69.7% of the variance of AWE-S scores. From the clustering analysis, we discovered that each valence state was clearly congregated except for one participant (averaged silhouette score = .871). In the decoding analysis, we observed that participants' self-reported valence label was predicted with the neural embeddings at each time point above the random chance (averaged test F1 score = .332).

Our results imply that, at the behavioral level, awe is encoded as an ambivalent experience rather than a positive or negative one. We also found that ambivalent states during awe experience exhibit distinct neural representations, just as positive/negative states do. Although some unsolved questions remain (e.g., identifying the brain regions mainly engaging in ambivalent states, building EEG-valence decoder with better performance), these findings support ambivalent properties of awe and suggest reconsidering ambivalent state as a meaningful unit of neuronal representation of valence.

Emotional Perception ¹

EEG/MEG Modeling and Analysis ²

Other - awe; ambivalence; electroencephalography; machine learning