A Multi-Source Few-Shot Learning Framework for Emergency Braking Intensity Prediction with Eeg

Accurate prediction of braking intensity enables a control strategy balancing comfort and safety. Electroencephalogram (EEG) contributes to earlier prediction of braking intensity but necessitates large-scale EEG dataset for each subject. To address the problem, this paper proposes a multi-source fewshot learning framework, integrating knowledge from multisubject datasets (source domains) to enhance the ability to adapt to an unseen subject (target domain) with a few samples. The framework consists of two sequential phases: (1) Pretraining via Domain AggRegation Network (DARN) adaptively allocating domain-specific weights to enhance effective sample size and avoid negative effects; (2) Fine-tuning, where pretrained model parameters are further optimized according to few labels in target domain. Experiments demonstrate that the proposed method can predict braking intensity accurately with minimal data in target domain.