679d8bbd776e0bbf3b044306c5be94ae-AuthorFeedback.pdf

There are many methods to train and evaluate generative models,15 the most popular of which is the average log likelihood that the model assigns to a sample set. For example, in16 GANs (Goodfellow et al. 2014) the focus is on average likelihood maximization,whichisequivalent to logloss17 minimization. Clarification of main contributions: The reviewer mentions that we study log-loss with respect to the above20 properties. Ontheotherhand,nofunction22 (not even log-loss) possess these properties without calibration. As we showed in Figure 1, in some applications23 log loss is not the optimal loss function to be used.