Section: New Results
Adaptive BCI training and operation
Participants : Jelena Mladenović, Jérémy Frey, Fabien Lotte
A generic framework for adaptive EEG-based BCI training and operation
There are two main approaches engaged in improving BCI systems: (i) improving the machine learning techniques, and the newly introduced (ii) improving human learning, by using the knowledge from instructional design and positive psychology. Both agree that the system needs to be adapted to the user but rely on different sources of adaptation: the machine for the former and the brain for the latter. In particular, machine learning algorithms should adapt to non-stationary brain signals, while human learning approaches should adapt the system to the various users' skills and profiles. Including both aspects of adaptation would give rise to a system ready to be used in real life conditions. However, a major obstacle lies in the large spectrum of sources of variability during BCI use, ranging from (i) imperfect recording conditions (e.g., environmental noise, humidity, static electricity etc. to (ii) the fluctuations in the user's psychophysiological states, due to e.g., fatigue, motivation or attention. For these reasons a BCI has not yet proved to be reliable enough to be used outside the laboratory. Particularly, it is still almost impossible to create one BCI design effective for every user, due to large inter subject variability. Therefore, the main concerns are to create a more robust system with the same high level of success for everyone, at all times, and to improve the current usability of the system. This calls for adaptive BCI training and operation.
We propose a conceptual framework which encompasses most important approaches to fit them in such a way that a reader can clearly visualize which elements can be adapted and for what reason. In the interest of having a clear review of the existing adaptive BCIs, this framework considers adaptation approaches for both the user and the machine, i.e., referring to instructional design observations as well as the usual machine learning techniques. It provides not only a coherent review of the extensive literature but also enables the reader to perceive gaps and flaws in current BCI systems, which would, hopefully, bring novel solutions for an overall improvement.
The framework (see Figure 16) has a hierarchical structure, from the lowest level elements which endure rapid changes, to the highest level elements which change at a much slower rate. It is comprised of: (i) one or several BCI systems/pipelines; (ii) a user model, whose elements are arranged according to different time scales ; (iii) a task model, enabling the system adaptation with respect to the user model; (iv) the conductor, an intelligent agent which implements the adaptive control of the whole system. A book chapter on this framework was submitted to a handbook on BCI.
Adapting BCI Feedback based on Flow Theory
Using BCI systems can be very frustrating for people because it is not trivial and so it takes time to master. Differently from other learning procedures, BCIs do not have enough, if any explanatory feedback in assisting the learning of users. Also, as the feedback is not engaging the user's mind might easily wander off, which highly affects the system's accuracy as well as the person's learning pace. For this reason it takes more time to train a user to understand the procedure and have control over the system. Hence, we want to create an immersive and playful environment to attract the user's attention and help them learn with less effort and frustration.
We rely on the theory of Flow, introduced by Csikszentmihalyi in the 1970s. Flow is a state of enjoyment while effortlessly focused on a task so immersive that one looses the perception of time. In order to fulfil these requirements, we choose the users to be involved in an open-source video game called Tux Racer. Also, to ensure the maximal attention of the users, the game difficulty adapts according to users performance in real-time.