Using logloss in ensembles

If the criteria on which the competition is being judged is log loss, then how does taking an ensemble make sense?

Does it make sense to take mean of the score (the score of a picture being a dog etc) for ensemble?

No. any deviation from the logloss accuracy may cost you a lot. However some Kaggle winners use essemble of Convolutional Neural Network to win,

Then why are we using ensembles in dogs_cats_ensemble notebook?