Figure 4.

Opinion comparison between NN₁ and NN₂ under undamaged MNIST data and opinion of $N{N}_1^D$ under damaged MNIST data. (A) Opinion of NN₁ with topology 784-1000-10. Belief reaches 0.78 when training data's uncertainty and disbelief are zero, i.e., belief is maximized. (B) Opinion of NN₂ with topology 784-500-500-10. Belief reaches 0.68 when training data's belief is maximized. Base rate is calculated using computational strategy presented in section 4. (C) Projected trust probability comparison between NN₁ and NN₂. Topology impacts projected trust probability. More precisely, NN₁ outperforms NN₂ while both NN₁ and NN₂ are trained with same dataset (same opinion assigned to dataset: {1, 0, 0, 0.5}, and same training process). (D–M) $N{N}_1^D$ with the same topology as NN₁, i.e., 784-1000-10, is trained with damaged data. Randomly take 10–100% training data, alter labels to introduce uncertainty and noise into dataset. Set opinion of damaged data point to have maximum uncertainty: {0, 0, 1, 0.5}. Belief is sparser while disbelief becomes denser in (D–M), but there is still belief even the dataset is 100% damaged. (N,O) normalized cumulative belief and disbelief of $N{N}_1^D$ under 10% to largest data damage, averaged over 10 runs. Note that the cumulative belief is not zero and increases during training even for a completely damaged data. Also note that both disbelief and belief increase as a function of number of episodes.