Machine learning and experimental tools towards EEG-based NeuroAdaptive Technologies
Designing NeuroAdaptive Technologies (NAT) requires to be able to estimate one or more mental states from a user’s brain signals, to understand the impact of this mental state on the way the user interact with the system, and to adapt this interaction accordingly. This in turns requires 1) suitable experimental protocols to induce the targeted mental states in order to obtain a ground truth, 2) robust machine learning tools to decode this mental state from brain signals and 3) computational models of the user that relates this mental state to the user behaviour and performances, and can take decisions accordingly. In this talk, I will present some of the tools we designed and used for each of these three points. In particular, I will present some experimental protocols to induce mental states such as workload, attention, intrinsic motivation or visual comfort, as well as machine learning tools (e.g., based on spatial fitlering, Riemannian Geometry or Convolutional Neural Networks) to robustly decode such states from EEG signals, despite EEG non-stationarity. I will then present our recent work with the active inference computational model, a framework that can be used to simultaneously infer the user’s state and take suitable decisions accordingly, in order to design robust NAT.