The bulk of machine learning models have a tendancy to rely too strongly to the distribution of the data on which they have been trained. Through this review paper I propose to discuss about ways to design an image classifier able to generalize well on a different but related distribution from its training one.