Proceedings of the Annual Meeting of the Cognitive Science Society

In recent years, EEG-based emotion recognition has emerged as a research hotspot in affective computing, yet mainstream deep learning approaches demand considerable time and substantial computational resources. To address these challenges, this paper proposes the HTK-BLS model, which accelerates both training and testing while maintaining competitive accuracy and reducing resource consumption. First, high-order singular value decomposition (SVD) is employed to enhance the network's capacity to capture tensor-level information from EEG frequency domain features. To further improve BLS's feature extraction capabilities, a novel stacked architecture employing approximate kernel functions is introduced with an optimized objective function to eliminate redundant nodes. In addition, weight-matrix computations are refined through truncated SVD, improving the robustness of the network. Experimental results on the DEAP dataset demonstrate that HTK-BLS achieves competitive recognition accuracy while enabling efficient training without backpropagation.