Consider a batch containing $2$ examples where a model outputs predicted probabilities over $3$ classes, represented by the matrix $\hat{\mathbf{y}} = \begin{bmatrix} 0.1 & 0.3 & 0.6 \\ 0.3 & 0.2 & 0.5 \end{bmatrix}$. If the true class labels for these examples are $0$ and $2$ respectively, the label vector is $\mathbf{y} = [0, 2]$. Using array indexing, we can efficiently extract the predicted probabilities corresponding to the true labels without writing a for-loop. For the first example (true class $0$), we select the first element of the first row ($0.1$). For the second example (true class $2$), we select the third element of the second row ($0.5$). The resulting array of selected probabilities is $[0.1, 0.5]$.