From gram to attention matrices: a monotonicity constrained method for eeg-based emotion classification

In this work, a parameter efficient attention module is developed for the task of emotion classification as well as improved model interpretability based on EEG source data. Inspired by the self-attention mechanism used in transformers, we propose a Monotonicity Constrained Attention Module (MCAM) that can help incorporate different priors easily on the monotonicity when converting Gram matrices from deep features into attention matrices for better feature refinement. In the subject-dependent classification task, MCAM achieves 95.0% mean prediction accuracy on four classification task with DEAP and 91.1% mean prediction accuracy on three classification task with SEED. On both datasets, MCAM is shown comparable to state-of-the-art attention modules in terms of boosting the backbone network’s predictive performance while requiring significantly fewer parameters. A thorough analysis is also performed on tracking the different effects inserted modules have on the backbone model’s behavior. For example, visualization and analysis techniques are presented to examine changes in spatial attention patterns reflected via kernel weights, change in prediction performance when different frequency information is filtered out, or changes that occur when different amplitude information is suppressed; as well as how different models change their predictions along linear morphisms between two samples belonging to different emotion categories. The results help to reveal what different modules learn and use during prediction, and can also provide guidance when applying them to specific applications. Graphical Abstract