UncertaintyEstimationforMulti-viewData: ThePowerofSeeingtheWholePicture Appendix

The third view with radius 0.3 was further translated to make the points overlapping, representing anoisyview. Then,theentiremodelwas trained with the learning rate of 0.003. We usedβ = 1 for the regularization coefficient. During inference, 100 samples were used to make a prediction. We used the early fusion method to concatenate multi-view features into unimodalfeature.