Upcoming deep surveys (e.g. LSST, JWST, EUCLID) will provide high quality imaging at unprecedentedly high red shifts, allowing the study of galaxy morphology at different cosmic times. The processing of such data will be necessarily automatized due to its enormous volume. Deep Learning has proven to be a powerful tool in these situations. Previous publications ([1] [2] [3] [4]) have shown the effectiveness of supervised learning algorithms for galaxy morphology classification. In the case of future surveys there is an additional challenge: the data will be completely unlabelled. This limits the use of a standard supervised approach since a subsample of data of which its classification is known will not be available for training and labelling data is expensive. Therefore, we explore an unsupervised approach: Simple framework for Contrastive Learning of visual Representations [5]. We test this algorithm on SDSS images of galaxies with z<0.15. We present preliminary results.