Semisupervised Seizure Prediction in Scalp EEG Using Consistency Regularization

Figure 2

The overall structure of CSSPM. The loss function consists of two components. The first cross-entropy loss is evaluated for labeled inputs only, where the ground truth is only given for these data. With the stochastic augmentation and dropout in the network, the entire neural network is considered as a stochastic model. The same input would yield different results at different epochs. Hence a mean square error loss, evaluated for all training data, is applied to penalize the bias between the current prediction and the ensemble prediction . A ramp-up weighting function is added to control the weight of the unsupervised mean square error loss.