Transfer Learning for
Cerebral interfaces for assisting people
Controlling machines directly from the thoughts is the main objective of cerebral interfaces (Wolpaw et al. 2002) or Brain-Computer Interfaces (BCI). The LISV lab is conducting signal processing research on cerebral signals and is internationally known on some of the most efficient techniques. The main idea is to rely on a representation space relevant for the spatial covariance matrices, that are endowed with adequate metrics (Chevallier et al. 2018).
FIGURE 1 – An evoked potential in EEG
Nonetheless, BCI is still very experimental due to several scientific and technological difficulties:
- Signal processing algorithms have very weak results for a large number of subjects, circa 20%, with almost no robust explanations (Vidaurre and Blankertz 2010, 10).
- EEG signals have very small amplitudes, hence difficult to record and largely afflicted by noise.
- Interacting with a BCI requires a focused mind and could be quickly exhausting.
Transfer learning for BCI
A previous research work (Kalunga, Chevallier, and Barthélemy 2018) introduced a new method to alleviate the first and second problem exposed above. This transfer learning approach allows to train the algorithms on existing datasets and to generalize directly on a new user, without any calibration phase. To do so, the proposed algorithm try to find users in the existing dataset with similar resting brain waves than a newly seen user.
This previous work was focused on one specific brain signal. The objective of this work is to apply this method on various kinds of BCI signals and to propose a systematic validation. The library MOABB (Jayaram and Barachant 2018) is a benchmarking tool to ease the scientific reproducibility by implementing automated tests. The objective of this work is to conduct experiments with MOABB to obtain publishable results that could be easily reproduced.
You could read the work of (Kalunga, Chevallier, and Barthélemy 2018) that introduces a good bibliographical synthesis on the existing approaches. You could find state-of-the-art methods from those bibliographical pointers. The part regarding the adequate space for handling covariance matrices is described in two complete and complementary reviews (Yger, Berar, and Lotte 2017; Congedo, Barachant, and Bhatia 2017).
The desired output of this part is a bibliographical synthesis to demonstrate the pros and cons of Riemannian geometry applied to the space of symmetric positive-definite matrices in the context of brain-computer interfaces.
To process the covariance matrices with tools borrowed from algebraic geometry is freely available from in Python. Regarding MOABB, you will find all the code and documentation online. To implement this part, some modification should be integrated in MOABB, but it could be handled in LISV if this task is too complex.
Applications and benchmarking
The objective to apply the transfer learning of (Kalunga, Chevallier, and Barthélemy 2018) on different context and datasets will demonstrate the strength of the proposed approach. You could test and evaluate the obtained results on this dataset to extend or modify the proposed method.
This small project could help you to gain a strong international visibility and could be a remarkable asset on your CV. This project has also a large potential to open on a research work and scientific publication.
For any question, you could reach me by mail or on Skype:
Sylvain Chevallier, LISV, Vélizy, France
Mail : email@example.com
Skype : sylvainchevallier-lisv
Chevallier, S., E. Kalunga, Q. Barthélemy, and F Yger. 2018. “Brain Computer Interfaces Handbook: Technological and Theoretical Advances.” In . CRC Press.
Congedo, M., A. Barachant, and R. Bhatia. 2017. “Riemannian Geometry for EEG-Based Brain-Computer Interfaces; a Primer and a Review.” Brain-Computer Interfaces 4: 1–20. https://doi.org/10.1080/2326263X.2017.1297192.
Jayaram, Vinay, and Alexandre Barachant. 2018. “MOABB: Trustworthy Algorithm Benchmarking for BCIs.” Journal of Neural Engineering 15 (6): 066011.
Kalunga, Emmanuel K, Sylvain Chevallier, and Quentin Barthélemy. 2018. “Transfer Learning for SSVEP-Based BCI Using Riemannian Similarities between Users.” In 2018 26th European Signal Processing Conference (EUSIPCO), 1685–1689. IEEE.
Vidaurre, Carmen, and Benjamin Blankertz. 2010. “Towards a Cure for BCI Illiteracy.” Brain Topography 23 (2): 194–198. https://doi.org/10.1007/s10548-009-0121-6.
Wolpaw, Jonathan R., Niels Birbaumer, Dennis J. McFarland, Gert Pfurtscheller, and Theresa M. Vaughan. 2002. “Brain–Computer Interfaces for Communication and Control.” Clinical Neurophysiology 113 (6): 767–791.
Yger, F., M. Berar, and F. Lotte.
2017. “Riemannian Approaches in Brain-Computer Interfaces: A Review.” IEEE
Trans Neural Syst Rehabil Eng 25 (10): 1753–1762.