support.qmd

---
format: 
    pdf:
        fontsize: 10pt
        geometry: 
            - top=10mm
            - bottom=10mm
            - left=10mm
            - right=10mm
---

\pagenumbering{gobble}

::: {#fig-fmnist layout="[[-1,10,-1], [-1,10,-1], [-1,10,-1], [20,-1,20,-1,20]]"}

![Turning a 9 into a 7.](www/mnist_9to7_1.png){#fig-mnist-1 width="50%"}

![Turning a 2 into a 3. ](www/mnist_2to3_14.png){#fig-mnist-2 width="50%"}

![Turning a 2 into a 3.](www/mnist_4to1_12.png){#fig-mnist-3 width="50%"}

![Turning a dress (left) into a trouser (right): a gap between the legs appears as expected.](www/fmnist_dress.png){#fig-fmnist-dress width="5cm"}

![Turning a pullover (left) into a coat (right): it looks like a V-neck has formed.](www/fmnist_pullover.png){#fig-fmnist-pullover width="5cm"}

![Turning a boot (left) into a sandal (right): pixels are darkened in the right places.](www/fmnist_boot.png){#fig-fmnist-boot width="5cm"}

Qualitative examples for MNIST and Fashion-MNIST. **MNIST** (@fig-mnist-1 to @fig-mnist-3): Top row: ECCCo. Bottom row: Wachter. The different underlying models across columns: (a) MLP, (b) Small Ensemble $n=5$, (c) Large Ensemble $n=50$, (d) LeNet-5, (e) JEM, (f) JEM Ensemble. **Fashion-MNIST** (@fig-fmnist-dress to @fig-fmnist-boot): The underlying classifier is a small ensemble of 5 MLPs with one hidden layer. Counterfactuals are generated by ECCCo. 
:::