A prominent technique for self-supervised representation learning has been to
contrast semantically similar and dissimilar pairs of samples. Without access
to labels, dissimilar (negative) points are typically taken to be randomly
sampled datapoints, implicitly accepting that these poi