“…In this context, we aim to design a novel self-supervised learning pipeline to handle tasks that require a robust distance measure and fine-grained analysis. We start by considering a common approach: clustering steps to propose pseudolabels to unlabeled samples and optimization steps to update the model supervised by those pseudo-labels [10], [11], [12]. However, prior methods that consider this approach often overlook two aspects: the quality of the features and the choice of hyper-parameters for the clustering algorithm.…”