Figure 5.

Inference values by sampling the output of Bayesian convolutional neural network 500 times. (A) Correct predictions. Left: ground truth is 0 (interictal), most of the model's output samples are close to 0, indicating a correct prediction with high confidence. Right: ground truth is 1 (preictal), most of the model's output samples are close to 1, indicating a correct prediction with high confidence. (B) Wrong predictions. Most of the model's output samples are close to the wrong value with high confidence. This is an undesirable case. (C) Low-confident predictions with high uncertainty levels. The model's output samples spread randomly between 0 and 1, indicating the high uncertainty of the model.