EEG 기반 SPD-Net에서 리만 프로크루스테스 분석에 대한 연구
Research of Riemannian Procrustes Analysis on EEG Based SPD-Net
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This paper investigates the impact of Riemannian Procrustes Analysis (RPA) on enhancing the classification performance of SPD-Net when applied to EEG signals across different sessions and subjects. EEG signals, known for their inherent individual variability, are initially transformed into Symmetric Positive Definite (SPD) matrices, which are natu- rally represented on a Riemannian manifold. To mitigate the variability between sessions and subjects, we employ RPA, a method that geometrically aligns the statistical distributions of these matrices on the manifold. This alignment is designed to reduce individual differences and improve the accuracy of EEG signal classification. SPD-Net, a deep learning archi- tecture that maintains the Riemannian structure of the data, is then used for classification. We compare its performance with the Minimum Distance to Mean (MDM) classifier, a conventional method rooted in Riemannian geometry. The ex- perimental results demonstrate that incorporating RPA as a preprocessing step enhances the classification accuracy of SPD-Net, validating that the alignment of statistical distributions on the Riemannian manifold is an effective strategy for improving EEG-based BCI systems. These findings suggest that RPA can play a role in addressing individual variability, thereby increasing the robustness and generalization capability of EEG signal classification in practical BCI applications.
Abstract
This paper investigates the impact of Riemannian Procrustes Analysis (RPA) on enhancing the classification performance of SPD-Net when applied to EEG signals across different sessions and subjects. EEG signals, known for their inherent individual variability, are initially transformed into Symmetric Positive Definite (SPD) matrices, which are natu- rally represented on a Riemannian manifold. To mitigate the variability between sessions and subjects, we employ RPA, a method that geometrically aligns the statistical distributions of these matrices on the manifold. This alignment is designed to reduce individual differences and improve the accuracy of EEG signal classification. SPD-Net, a deep learning archi- tecture that maintains the Riemannian structure of the data, is then used for classification. We compare its performance with the Minimum Distance to Mean (MDM) classifier, a conventional method rooted in Riemannian geometry. The ex- perimental results demonstrate that incorporating RPA as a preprocessing step enhances the classification accuracy of SPD-Net, validating that the alignment of statistical distributions on the Riemannian manifold is an effective strategy for improving EEG-based BCI systems. These findings suggest that RPA can play a role in addressing individual variability, thereby increasing the robustness and generalization capability of EEG signal classification in practical BCI applications.
참고문헌 (16)
- Kothe CA, "BCILAB : a platform for brain–computer interface development", Journal of Neural Engineering, 10(5), pp.056014~ (2013)
- Yger F, "Riemannian approaches in brain-computer interfaces : a review", IEEE Transactions on Neural Systems and Rehabilitation Engineering, 25(10), pp.1753~1762 (2016)
- Congedo M, "Riemannian geometry for EEG-based brain-computer interfaces; a primer and a review", Brain-Computer Interfaces, 4(3), pp.155~174 (2017)
- Rodrigues PLC, "Riemannian Procrustes analysis : transfer learning for brain–computer interfaces", IEEE Transactions on Biomedical Engineering, 66(8), pp.2390~2401 (2018)
- Huang Z, "A Riemannian network for SPD matrix learning", Proceedings of the AAAI Conference on Artificial Intelligence, 31(1), pp.1~10 (2017)
- Bhatia R, "Positive definite matrices", pp.1~200 (2009), Princeton University Press
- Gower JC, "Procrustes problems", pp.1~10 (2004), Oxford University Press
- Duan RN, "Differential entropy feature for EEG-based emotion classification", Proceedings of the 6th International IEEE EMBS Conference on Neural Engineering, pp.81~84 (2013)
- Koelstra S, "DEAP : A database for emo-tion analysis using physiological signals", IEEE Transactions on Affective Computing, 3(1), pp.18~31 (2011)
- Cho H, "EEG datasets for motor imagery brain-computer interface", GigaScience, 6(7), pp.gix034~ (2017)
- Graf AB, "Prototype classification : insights from machine learning", Neural Computation, 21(1), pp.272~300 (2009)
- Rodrigues PLC, ""When does it work?" : An exploratory analysis of transfer learning for BCI", Proceedings of the BCI, pp.1~6 (2019)
- Gwon D, "A Comparative Analysis of Motor Imagery, Execution, and Observation for Motor Imagery-based Brain-Computer Interface", Journal of Biomedical Engineering Research, 43(6), pp.375~381 (2022)
- Kim BH, "A discriminative SPD feature learning approach on Riemannian manifolds for EEG classification", Pattern Recognition, 143, pp.109751~ (2023)
- Brooks D, "Riemannian batch normalization for SPD neural networks", Advances in Neural Information Processing Systems, 32, pp.1~10 (2019)
- Flamary R, "Optimal transport for domain adaptation", IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9), pp.1853~1865 (2016)