DBGMS: A Dual-Branch Generative Adversarial Network with Multi-Task Self-Supervised Enhancement for Robust Auditory Attention Decoding

Shuai Huang, yongxiong wang, Chendong Qin

Keywords: electroencephalogram(EEG), Auditory Attention Decoding(AAD), Dual-branch, generative adversarial networks(GANs)

Abstract:

Detecting auditory attention from brain signals has been a significant challenge in neuroscience and brain-computer interface research. While progress has been made in EEG-based auditory attention detection, existing methods often struggle with limited data and short decision windows, particularly in complex auditory environments. In this paper, we propose DBGMS (Dual-Branch Generative Adversarial Network with Multi-Task Self-Supervised Enhancement), a novel framework for robust auditory attention decoding from electroencephalogram (EEG) signals. There are three key innovations in our approach:
(1) A dual-branch architecture is developed that combines temporal attention and frequency residual learning, enabling more comprehensive feature extraction to be achieved from EEG signals;
(2) Branch-specific generative adversarial networks (GANs) are designed to generate high-quality augmented samples in both temporal and frequency domains, effectively addressing the data scarcity issue in auditory attention decoding;
(3) Attention mechanisms and graph convolution operations are incorporated in both temporal and frequency domains.
(4) A multi-task self-supervised learning strategy is introduced, incorporating several complementary tasks such as temporal order prediction, frequency band reconstruction, and time-frequency consistency. This approach leverages unlabeled data to enhance the model's ability to capture subtle attention-related features from multiple perspectives, thereby improving generalization across subjects and listening conditions.
In contrast to state-of-the-art methods, DBGMS presents significant improvements in detection accuracy and robustness, particularly for short decision windows. Our framework is evaluated on two public EEG datasets, including KUL and DTU, demonstrating its effectiveness across various experimental settings.

Primary Area: other topics in machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4590