Adaptive CCA for c-VEP BCI

Context

The brain-computer interface (BCI) that uses the code-modulated visual evoked potential (c-VEP) is one of the fastest non-implanted BCIs wordwide. A much used machine learning technique in c-VEP BCI is canonical correlation analysis (CCA). CCA allows to learn spatial filters, and when using the reconvolution method also temporal filters.

The original CCA method as applied in the domain of BCI is a supervised method, requiring labeled training data. The amount of training data may depend on the current BCI user and quality of the EEG, which can be formulated as an optimization problem.

Additionally, CCA has been extended to a calibration-free method, and recently been compared to another sophisticated calibration-free method, called unsupervised mean-difference maximization (UMM). These two calibration-free methods allow decoding without the use of training data, effectively decoding every trial as if it were the first, so-called instantaneous decoding. However, both methods can also cumulatively learn from previously classified trials, making them adaptive.

Learning from previous trials comes with certain challenges. One could assume previous trials were classified correctly, though potentially some were more or less reliable. This calls for both a way to express classification certainty, as well as a way to incorporate certainty in the classification update.

Further optimizations, or regularizations, can be applied to (calibration-free) CCA, that may substantially reduce again the amount of required (training) data, as well as further improve decoding accuracy and speed.

Image credit: Thielen, J., Marsman, P., Farquhar, J., & Desain, P. (2021). From full calibration to zero training for a code-modulated visual evoked potentials for brain–computer interface. Journal of Neural Engineering, 18(5), 056007.

Research question

As part of this project, you will implement and investigate one or some of the following ideas:

• Can we regularize CCA in a similar way as UMM is regularized?
• Can we define a criterion signaling whether sufficient training data is recorded to train CCA?
• Can we define a classification certainty signaling whether sufficient testing data is recorded to emit the classification (i.e., dynamic stopping)?
• Can we define a classification certainy and incorporate this into the adaptive update of CCA over trials?

You will focus predominantly on implementing the specific methodology and evaluating the results on an existing c-VEP dataset, associated to the original publications.

Literature

Original reconvolution CCA:

• Thielen, J., Van Den Broek, P., Farquhar, J., & Desain, P. (2015). Broad-band visually evoked potentials: re (con) volution in brain-computer interfacing. PLOS One, 10(7), e0133797. DOI: https://doi.org/10.1371/journal.pone.0133797

Original calibration-free reconvolution CCA:

• Thielen, J., Marsman, P., Farquhar, J., & Desain, P. (2021). From full calibration to zero training for a code-modulated visual evoked potentials for brain–computer interface. Journal of Neural Engineering, 18(5), 056007. DOI: https://doi.org/10.1088/1741-2552/abecef

Comparison of calibration-free reconvolution CCA with UMM:

• Thielen, J., Sosulski, J., & Tangermann, M. (2024). Exploring new territory: Calibration-free decoding for c-VEP BCI. Proceedings of the 9th Graz Brain-Computer Interface Conference 2024. DOI: https://doi.org/10.3217/978-3-99161-014-4-057

Skills / background required

• Knowledge and an interest in brain-computer interfacing
• Very proficient in Python
• Strong mathematical background
• Strong machine learning background