Robustness and Generalization Capability Validation of Convolutional Neural Network on a Chinese EEG Auditory Attention Decoding Dataset

Various solutions have been proposed to correlate electroencephalogram (EEG) signals with audio stimuli, with DNN-based methods significantly outperforming rule-based approaches in auditory selective attention decoding (ASAD). However, DNN models often overfit due to their capability to capture intricate patterns in training data, necessitating careful preprocessing and cross-validation to ensure robustness and generalization. Existing EEG datasets mainly focus on single speakers or two speakers' scenarios in fixed positions, neglecting more realistic scenarios. Our prior work introduced an EEG dataset with fifteen alternative competing speaker directions, but it was only verified on seen subjects and trials. In this paper, we extend our dataset validation using various leave-one-out procedures involving unseen subjects, trials, and attended audios. Results show our EEG dataset contains informative features for successful ASAD in binary classification under leave-one-subject-out, leave-one-trial-out, leave-one-audio-out, and leave-one-moment-out paradigms. Additionally, successful decoding of attended directions motivates further exploration of multiclass ASAD, crucial for developing neuro-steered hearing aids.