Figure 1.

Example of the evolution of plausible labels for an image. Without evidence, the probability distribution ${\mathbf{q}}_0\left(y\right)$ assigns equal plausibility to all outcomes. A machine learning model processes the image and produces predictions ${\mathbf{q}}_1\left(y\right|x)$. The incorrect label 3 is represented by $\mathbf{r}\left(y\right|\check{y})$. After observing the image, shown on the right, the label is corrected to 0 in $\mathbf{r}\left(y\right|\widehat{y})$.