Weakly Supervised Learning via Relational Comparisons

It is observed that the Siamese few-shot learning model is the most efficient at us-ing the pairwise-relationship data for classification, which is similar to the accuracy achieved with 50k datapoints in a fully supervised setting.