The invention notably relates to a computer-implemented method for learning an autoencoder. The method comprises providing a dataset of images. Each image includes a respective object representation.The method also comprises learning the autoencoder based on the dataset. The learning includes minimization of a reconstruction loss. The reconstruction loss includes a term that penalizes a distancefor each respective image. The penalized distance is between the result of applying the autoencoder to the respective image and the set of results of applying at least part of a group of transformations to the object representation of the respective image. Such a method provides an improved solution to learn an autoencoder.