GZSL-Lite: A Lightweight Generalized Zero-Shot Learning Network for SSVEP-Based BCIs.

Abstract

Generalized zero-shot learning (GZSL) networks offer promising avenues for the development of user-friendly steady-state visual evoked potential (SSVEP) based brain-computer interfaces (BCIs), aiming to alleviate the training burden on users. These networks only require the user to provide training data from partial classes during training, yet they demonstrate the capability to classify all classes during testing. However, these GZSL networks have a large number of trainable parameters, resulting in long training times and difficulty to practicalize. In this study, we proposed a dual-attention structure to construct a lightweight GZSL network, termed GZSL-Lite. We first embedded the input training data-constructed class templates, manually constructed sine templates, and electroencephalogram (EEG) signals using convolution-based networks. The embedding part uses the same network weights to embed the features across different stimulus frequencies while reducing the depth of the network. After embedding, two branches of the dual-attention use class and sine templates to guide the feature extraction of the EEG signal with the attention mechanism, respectively. Compared to the networks that extract all features equally, dual-attention focuses only on EEG features relative to templates, which helps classification with fewer parameters. Finally, we used depthwise convolutional blocks to output classification results. Experimental evaluations conducted on two publicly available datasets demonstrate the efficacy of the proposed network. Comparative analysis reveals a remarkable reduction in trainable parameters to less than 1% of the SOTA counterpart, concurrently showing significant performance improvement. The code is available for reproducibility at https://github.com/xtwong111/GZSL-Lite-for-SSVEP-Based-BCIs.