epistemic uncertainty
Variational Inference (VI)
방식을 활용해 GS를 Bayesian
framework에서 학습하게 하여 uncertainty prediction을 자연스럽게 결합했다.Area Under Sparsification Error (AUSE)
라는 loss를 제안한다reparameterization trick
을 이용해서, gradient가 끝까지 전달될 수 있다.If the estimated variance is a good representation of the model uncertainty, and the pixels with the highest variance are removed gradually, the error should monotonically decrease