• representation learning
• affective computing

(AC) is a field of biomedical research that builds an "affect model" by analyzing physiological signals. These models can be used to support therapeutic treatments for psychological disorders.
Most approaches involving affective computing are based on the analysis of EEG signals.
The multimodal nature of emotional expressions motivates the use of continuous representations in affective computing.
The representation of features extracted in latent space contains all the important information needed to capture the behavior of the original data; if the latent space is ordered, it is more interpretable.
This thesis focuses on designing and testing a new loss function that maps label ordering to the feature space of representation learning in order to understand whether feature ordering within the latent space is related to the ordering of the original data space.
In this thesis, we use EEG signals from the well-known DEAP and MAHNOB datasets to perform multiclass emotion classification.