Most recent unsupervised learning methods explore alternative objectives, often referred to as self-supervised tasks, to train convolutional neural networks without the supervision of human annotated labels. This paper explores the generation of surrogate classes as a self-supervised alternative to learn discriminative features, and proposes a clustering algorithm to overcome one of the main limitations of this kind of approach. Our clustering technique improves the initial implementation and achieves 76.4% accuracy in the STL-10 test set, surpassing the current state-of-the-art for the STL-10 unsupervised benchmark. We also explore several issues with the unlabeled set from STL-10 that should be considered in future research using this dataset.