Tensor Decomposition-Based Feature Transfer Learning for BCI Applications

Event-Related Potential (ERP) detection and enhancement remain significant challenges in EEG signal processing due to the complex and multidimensional nature of ERP signals. This paper addresses these challenges by proposing a novel framework that combines low-rank tensor representation with heuristic feature transfer learning for efficient ERP processing. By leveraging tensor decomposition, we capture intricate multidimensional patterns in ERP data, facilitating the extraction of informative features. These features greatly reduce processing time, making the method particularly suitable for online Brain-Computer Interface (BCI) applications. We also present a feature transfer learning mechanism utilizing tensor decomposition, which facilitates the efficient adaptation of features trained on multiple subjects to new subjects, requiring only a small amount of calibration data. This approach leverages shared classification structures across subjects, allowing for rapid deployment and minimizing the need for extensive retraining. Our results demonstrate that the proposed method achieves reasonable ERP detection accuracy while drastically reducing computational overhead, presenting a robust solution for real-time BCI systems.