Robustness of Unsupervised Representation Learning without Labels