PDAML: A Pseudo Domain Adaptation Paradigm for Subject-independent EEG-based Emotion Recognition

Domain adaptation (DA) and domain generalization (DG) methods have been successfully adopted to alleviate the domain shift problem caused by the subject variability of EEG signals in subject-independent affective brain-computer interfaces (aBCIs). Usually, the DA methods give relatively promising results than the DG methods but require additional computation resources each time a new subject comes. In this paper, we first propose a new paradigm called Pseudo Domain Adaptation (PDA), which is more suitable for subject-independent aBCIs. Then we propose the pseudo domain adaptation via meta-learning (PDAML) based on PDA. The PDAML consists of a feature extractor, a classifier, and a sum-decomposable structure called domain shift governor. We prove that a network with a sum-decomposable structure can compute the divergence between different domains effectively in theory. By taking advantage of the adversarial learning and meta-learning, the governor helps PDAML quickly generalize to a new domain using the target data through a few self-adaptation steps in the test phase. Experimental results on the public aBICs dataset demonstrate that our proposed method not only avoids the additional computation resources of the DA methods but also reaches a similar generalization performance of the state-of-the-art DA methods.