Exploring Transfer Learning Focused on Physiological Signals for Emotion Recognition

Abstract

Recent work in the area of automatic emotion recognition has leveraged a large amount of publicly available data with transfer learning techniques to detect emotion on low-resource data. Previous work demonstrated that the use of maximum independence domain adaptation and transfer component analysis show promise in generalizing on unseen domains. While the accuracy increases are significant, they remain below within-dataset models. Other research concluded that a subspace alignment auto-encoder (SAAE) is useful for domain adaptation and is more effective than current techniques. Despite the encouraging results of these studies, more work needs to be done to extend this to real world brain computer interaction (BCI) applications. The primary goal of this thesis is to develop transfer learning techniques that leverage existing data and attempt to generalize them on unseen domains accurately enough for real-world applications. If the proposed endeavor is successful, emotion recognition for real-life applications will not need to include large amounts of data from the target domain since transfer learning techniques will be able to accurately generalize on unseen domains.