Analyzing the Potential Contribution of a Meta-Learning Approach to Robust and Effective Subject-Independent, Emotion-Related Time Series Analysis of Bio-signals

Emotion recognition has recently attracted much attention due to current progress in the fields of affective computing and human-computer interaction (HCI), leading to the development of various deep learning models tailored to emotion recognition. Applied to subject-dependent emotion recognition scenarios, these models have achieved great performance. However, in subject-independent scenarios, these models perform poorly. The challenge of subject-independent emotion recognition is primarily due to the high variability in how emotions are expressed among different subjects. In this work, we analyze and present the potential contribution of Meta-Learning to achieving robust and effective subject-independent emotion recognition. Our suggestion is predicated on the idea behind Meta-Learning, which is about developing models that can learn new tasks or adapt to new environments rapidly with minimal training data, drawing on knowledge from previous tasks. To achieve highly adaptive subject-independent emotion recognition, this work outlines a conceptual design of a novel Meta-Learning architecture that combines Message Passing Neural Networks (MPNN) and Few-Shot Learning (FSL), implemented as First-Order Model-Agnostic Meta-Learning (FOMAML). Used on the DEAP dataset, a multimodal benchmark emotion recognition dataset, our architecture outperforms the Bi-LSTM and the non-FSL MPNN baseline models by 17 and 9%, respectively, in the prediction of valence and arousal scores. These results highlight the fact that the analysis of EEG signals, the main signals of the DEAP dataset, can be improved by encoding the adjacency relation between electrodes by leveraging Meta-Learning techniques.