Go to ICLR 2025 Conference homepageFSL-MIC: An Attentional Few-Shot Learning Framework for EEG Motor Imagery Classification
Electroencephalography (EEG) is a key non-invasive technique used to investigate brain activity, particularly in motor imagery (MI) research. Traditional methods for classifying EEG signals often rely on handcrafted features and heuristic parameters, which can limit generalization across tasks and subjects. Recent advances in deep learning, particularly few-shot learning (FSL), offer promising alternatives to improve classification accuracy in scenarios with limited training data. This study explores the effectiveness of FSL algorithms, including Relation Networks, to enhance MI classification. It also examines how transfer learning and data augmentation techniques contribute to improving classification performance.
We propose a novel framework with three core modules—feature embedding, attention, and relation—that facilitates the classification of unseen subject categories using only a few labeled samples. The attention mechanism identifies key features related to the query data, while the relation module predicts query labels by modeling relationships between support and query data across subjects. Our experimental results demonstrate the effectiveness of our approach on two benchmark datasets, BCI 2a and BCI 2b, as well as our experimental dataset. The proposed FSL framework significantly outperforms traditional methods, offering promising applications in real-time Brain-Computer Interface (BCI) systems across various EEG setups. This research advances the understanding of machine learning in EEG applications and highlights the potential of FSL techniques in overcoming the challenges of limited training data in MI classification.