ViT-Based EEG Analysis Method for Auditory Attention Detection

This paper presents our EEG analysis method for auditory attention detection (AAD) as utilized in the first Chinese Auditory Attention Decoding Challenge, ISCSLP 2024. Given the impressive performance of transformers across various domains and the inherent correlation between EEG data and images, we selected the vision transformer (ViT) as the backbone of our auditory attention detection network. In this work, we first present our data transformation and augmentation method, which converts EEG signals into images. Considering the limited training data available in the challenge, we explore the evaluation method for the model's generalization performance in cross-domain scenes. Furthermore, we examine the impact of different model sizes on the detection results. Our submitted result achieves fifth in Track1 with an average accuracy of 91.73% and third in Track2 with an average accuracy of 51.36% in the challenge ranking. The code and network weights are available at https://github.com/xfy0032/add.