Transient Generalization Over Different Presentation Frequencies: Lagged Stimuli and Superposition
Attempts to make brain-computer interfaces more generalizable have seen some success. The generative model of Thielen et al. has been able to accurately predict responses to unseen stimuli. However, this model still requires the unseen stimuli to be presented at the same frequency. This is because the transients are assumed to stay the same in the gen- erative model, which is a false assumption when the presentation rate changes. In this thesis, two methods of generalizing transient responses across di erent presentation rates are investigated. The rst method uses lagged stimuli: stimuli where the rate is changed by leaving the length of the stimulus-ON the same, but changing the length of the stimulus-OFF. The second method uses superposition, i.e. linear addition of responses, in order to predict transient responses at di erent presentation rates. The rst method using lagged stimuli showed promise. Classi cation accura- cies were signi cantly better than chance when predicting responses to stimuli at a rate that was di erent from the training rate. At the same time, however, the correlations between the transients of the same event at di erent rates were not signi cant. The second method using super- position showed no signi cant results. Some correlations between real transients and transients that were predicted using superposition were signi cant, but the accuracies were at chance level.
Faculteit der Sociale Wetenschappen