In this project, we utilized optimization to discriminate brain data. Participants completed 2 cognitive tasks while ongoing brain activity was recorded from electrodes on their scalp. Our analysis examined whether we could identify what task the participant was performing from differences in the recorded brain time series. We modeled the relationship between input data (brain time series) and output labels (task A and task B) as an unknown function, and we found an optimal approximation of that function from among a family of functions. We employed stochastic gradient descent to minimize the estimation error known as the loss function. The optimal function from among our family of approximate functions, EEGNet, successfully discriminated brain data from a single participant with approximately 90% accuracy. Future research will apply EEGNet on data from more participantsas well as develop approaches to adapt its architecture for the non-Euclidean domains.
展开▼
