Paradox learning: Contradictory labels to disentangle symmetric and asymmetric EEG spatial patterns

Emotions induce specific brain activity patterns, which may be symmetric or asymmetric between the left and right hemispheres. Deep Electroencephalography (EEG) classification networks have been applied to utilize these spatial patterns for emotion recognition. However, most research has neglected the symmetric spatial patterns and solely focuses on exploiting the asymmetric ones, leading to limited performance improvement. To rigorously compare the importance of asymmetric and symmetric patterns in emotion recognition and simultaneously improve their application efficiency, we propose a Paradox Learning scheme. In this scheme, the original samples are flipped left to right to get mirror samples which are labeled with the true or opposite emotion labels as the clue to introduce contradictory supervision. Two sub-networks are constructed to extract the symmetric and asymmetric patterns respectively with the contradictorily labeled samples. The same labels on the original and the mirror samples guide the sub-network to learn the symmetric patterns which are not influenced by the left-to-right hemisphere switch, while the contradictory labels of the mirror samples lead to the extraction of the asymmetric patterns. The two sub-networks share an identical structure and identical training set. The learned symmetric and asymmetric patterns are ensembled to fulfill the emotion classification task. In this way, a paradox learning scheme is developed to extract multi-view features from the augmented samples with contradictory supervision. Extensive experiments and visualization results confirmed that symmetric and asymmetric patterns were indeed disentangled and independently utilized by the two sub-networks. Symmetric patterns are more beneficial for emotion recognition than asymmetric ones. The optimal network structures for effectively utilizing symmetric and asymmetric spatial patterns differ from each other. Furthermore, the multi-view network that integrates the two sub-networks outperformed each individual sub-network and achieved approximately 2% higher accuracy compared to state-of-the-art methods, indicating a promising research direction for paradox learning.