Uncertainty Estimation for Multi-view Data: The Power of Seeing the Whole Picture Appendix A Proofs and Derivations

However, the log likelihood term is not analytical yet. The third view with radius 0.3 was further translated to make the points overlapping, We used 128 units as the hidden dimension. SNGP's GP layer: We followed the settings provided by its authors in their tutorial Our framework is based on GPflow [9]. The uncertainty surfaces of each view are shown in Figure 2. Implementation Details We used the same datasets of the TMC's datasets (Handwritten, CUB, In addition to the methods used in B.1, we implemented MC Dropout and DE(EF) MC Dropout: We used a dropout layer with the dropout rate of 0.2 and a fully connected During inference, 100 samples were used to make a prediction. Figure 4: Domain-shift test accuracy where Gaussian noise is added to half of the views.