Optional $-$ Bonus$)$

In Creative Adversarial Netwoks $($CANs$)[3],$ an additional loss is added to en$-$

courage producing novel images with the loss defined as

where $x$ and hat$(c)$ are a real image and its corresponding art style label from the

data distribution $p_{\text{d}\text{a}\text{t}\text{a}}.D_r(*)$ is the transformation function that tries to dis$-$

criminate between real art and generated images. $D_c(*)$ is the the function that

discriminates between different style categories and estimates the style class

posteriors $($i$.$e$.,(*((*|)|)\}.$

$(7\mathrm{.}5$ points$)$ Comment on the use of Binary Cross Entropy$($BCE$)$ in the

deviation loss summed up over each style class $($i$.$e$.$ terms in ${\displaystyle \underset{k=1}{\overset{K}{}}}(cdots)$ above

$.$ What part of the loss function enables the model not to produce meaningless

art images since they may also have high entropy.

$(7\mathrm{.}5$ points$)$ Why Multiclass Cross Entropy improved the performance

over BCE in Holistic$-$CAN$[4]($CAN$-$H$)?$