A PROBABILISTIC APPROACH TO THE CONSTRUCTION OF A MULTIMODAL AFFECT SPACE

Understanding affective signals from others is crucial for both human-human and human-agent interaction. The automatic analysis of emotion is by and large addressed as a pattern recognition problem which grounds in early psychological theories of emotion. Suitable features are first extracted and then used as input to classification (discrete emotion recognition) or regression (continuous affect detection). In this thesis, differently from many computational models in the literature, we draw on a simulationist approach to the analysis of facially displayed emotions - e.g., in the course of a face-to-face interaction between an expresser and an observer. At the heart of such perspective lies the enactment of the perceived emotion in the observer. We propose a probabilistic framework based on a deep latent representation of a continuous affect space, which can be exploited for both the estimation and the enactment of affective states in a multimodal space. Namely, we consider the observed facial expression together with physiological activations driven by internal autonomic activity. The rationale behind the approach lies in the large body of evidence from affective neuroscience showing that when we observe emotional facial expressions, we react with congruent facial mimicry. Further, in more complex situations, affect understanding is likely to rely on a comprehensive representation grounding the reconstruction of the state of the body associated with the displayed emotion. We show that our approach can address such problems in a unified and principled perspective, thus avoiding ad hoc heuristics while minimising learning efforts. Moreover, our model improves the inferred belief through the adoption of an inner loop of measurements and predictions within the central affect state-space, that realise the dynamics of the affect enactment. Results so far achieved have been obtained by adopting two publicly available multimodal corpora.