Predicting Activity in Brain Areas Associated with Emotion Processing Using Multimodal Behavioral Signals
Résumé
Artificial agents are expected to increasingly interact with humans and to demonstrate multimodal adaptive emotional responses. Such social integration requires both perception and production mechanisms, thus enabling a more realistic approach to emotional alignment than existing systems. Indeed, existing emotion recognition methods rely on behavioral signals, predominantly facial expressions, as well as non-invasive brain recordings, such as Electroencephalograms (EEGs) and functional Magnetic Resonance Imaging (fMRI), to identify humans' emotions, but accurate labeling remains a challenge. This paper introduces a novel approach examining how behavioral and physiological signals can be used to predict activity in emotion-related regions of the brain. To this end, we propose a multimodal deep learning network that processes two categories of signals recorded alongside brain activity during conversations: two behavioral signals (video and audio) and one physiological signal (blood pulse). Our network enables (1) the prediction of brain activity from these multimodal inputs, and (2) the assessment of our model's performance depending on the nature of interlocutor (human or robot) and the brain region of interest. Results demonstrate that the proposed architecture outperforms existing models in anterior insula and hypothalamus regions, for interactions with a human or a robot. An ablation study evaluating subsets of input modalities indicates that local brain activity prediction was reduced when one or two modalities are omitted. However, they also revealed that that the physiological data (blood pulse) achieve similar levels of predictions alone compared to the full model, further underscoring the importance of somatic markers in the central nervous system's processing of social emotions.
Domaines
| Origine | Fichiers éditeurs autorisés sur une archive ouverte |
|---|---|
| Licence |