BCI Fundamental Didactic Sessions

BCI Fundamental Didactic Sessions are intended for students.  Sessions will take place on Monday, June 8 from 3:00pm to 5:00pm.

Session 1: Machine Learning, 3:00pm – 3:30pm

Presented by Christian Herff, Maastricht University

Machine Learning for BCI


Recorded brain activity displays complex non-linear dynamics and is highly participant and even session dependent. Furthermore, the signals often display many non-stationarities. To still make sense of these challenging data, specific consideration for the machine learning approach must be made.  In this didactic session, common challenges for machine learning in BCI research will be highlighted and some key considerations of how to address them will be discussed. Briefly, both a regression and a classification problem will be discussed. Quick examples, using both invasive and non-invasive data, will be used to illustrate problems and common fallacies. As in many time series data, specific attention must be paid to the evaluation in BCI research, as the strong autocorrelations of the data can otherwise induce exaggerated results. Finally, some key considerations for the translation from offline results to viable BCI solutions will be presented.

Session 2: EEG Analysis, 3:30 pm – 4:00pm

Presented by Jason Palmer, Osaks University

Title and abstract to come

Session 3: Adaptive BCI non-invasive, 4:00 pm – 4:30pm

Presented by Serafeim Perdikis, University of Essex

Adaptation and learning in non-invasive brain-computer interaction


Non-invasive brain-computer interface (BCI) has entered an era of relative technological maturity, where BCI prototypes are increasingly deployed as assistive devices and rehabilitation interventions, or towards able-bodied user applications like entertainment and driving assistance. The main obstacles hindering the translation and industrialization of non-invasive BCI are the intense performance fluctuations that often impede brain-actuated device operation, and the inability of large portions of prospective users to exhibit adequate BCI control after conventional user training. Adaptive BCI algorithms and the co-adaptive (human and machine) symbiotic regimes they give rise to, have been early proposed as a remedy to both these issues. In this session, we will first identify the machine learning and other technical user-training challenges that need to be addressed towards effective BCI adaptation, with references to possible solutions that have been proposed in the recent non-invasive BCI literature, and their limitations. Second, we will shift the focus to the–often, overlooked–topic of subject learning during co-adaptation, taking a critical viewpoint of the state-of-the-art and leveraging the evidence of recent works in order to pinpoint the currently missing links towards a truly mutual learning framework. Ultimately, this session aspires to survey and conceptualize the main theoretical and practical caveats of non-invasive, adaptive BCI, thus providing a tentative roadmap towards co-adaptive training able to facilitate both learning agents of the BCI loop, to accommodate their interactions, to enable universal BCI accessibility and, through that, to fuel translational and commercial BCI applicability.

Session 4- Adaptive BCI invasive, 4:30 pm – 5:00pm

Presented by Amy Orsborn, University of Washington

Design considerations for closed-loop decoder adaptation algorithms in invasive BCIs


Closed-loop decoder adaptation (CLDA) is a powerful technique in invasive brain-computer interfaces. CLDA is frequently used to train and optimize decoding algorithms “in situ,” reducing errors in performance that arise due to differences between open loop decoder training and closed loop BCI operation. CLDA has also proven useful as a strategy to maintain performance over time despite non-stationary measurements (e.g. signal drift). Many design choices must be made when creating a CLDA algorithm, which will influence its overall performance and utility. These choices include the form of error signal used to guide algorithm retraining, the learning rules used to update the decoder, the timescale of decoder updates, and the decoder parameters to update. User-decoder interactions must also be considered when designing CLDA algorithms. In this talk, I will briefly survey these details. I will present a case-study of CLDA algorithms optimized to robustly and rapidly initialize a BCI decoder independent of initialization, and review the algorithm design choices that led to the success of these algorithms. I will also briefly touch on the use of CLDA for promoting co-adaptation between the brain and decoder.

Thank you to our 2021 Sponsors!