Dynamic learning of auditory attention patterns under combinatorial space

Abstract

Humans exhibit auditory attentional selectivity in noisy environments, which provides biological inspiration for addressing the auditory attention decoding (AAD) problem. In recent years, geometric learning (GL) has demonstrated impressive performance in decoding noisy electroencephalography (EEG) signals. However, current techniques can be further improved while confronting low signal-to-noise ratio and spatiotemporal complexity. Here, this paper proposes a multi-space dynamic learning strategy (MEC-MSL) under manifold–Euclidean combinatorics, which embeds knowledge spectra in different spaces to achieve robust feature learning in multi-dimensional spaces thereby decoding efficient auditory attention representations. Specifically, MEC-MSL enhances the representation of knowledge spectra in Euclidean space from multiple perspectives and incorporates an attention mechanism to calibrate contextual information, and then capturing more discriminative patterns. Simultaneously, a manifold self-attention module is developed to explore high-dimensional prior topological representations. Furthermore, a dynamic topological pattern learning module is constructed to fully characterize the local and global functional linkages of EEGs for efficient feature extraction in hybrid space. Experimental results on the KUL, DTU and AVED auditory attention datasets demonstrate that the proposed MEC-MSL achieves state-of-the-art decoding performance and computational efficiency, providing a novel solution for robust and efficient auditory attention representation. The source code is publicly available at https://github.com/dddyuu/MEC-MSL.