Contrastive Representation Through Angle and Distance Based Loss for Partial Label Learning

Abstract

Partial label learning (PLL) is a form of weakly supervised learning which aims to train a deep network from training instances and its corresponding label set. The label set, also known as the candidate set, is a group of labels associated with each training instance, out of which only one label is the ground truth for the training instance. Contrastive learning is one of the popular techniques used to learn from a partially labeled dataset, intending to reduce intra-class while maximizing inter-class distance. In this paper, we suggest improving the contrastive technique used in PiCO. The proposed C ontrastive R epresentation via A ngle and D istance based L oss (CRADL) segregates the contrastive loss into two parts, the angle based loss and the distance based loss. The former angle based loss covers the angular separation between two contrastive vectors. However, we showcase a scenario where such angular loss prefers one contrastive vector over the other despite having the same angle. Thus, the second loss term built on distance fixes the issue. We show experiments on CIFAR-10 and CIFAR-100, where the corresponding PLL databases are generated using uniform noise. The experiments show that the PLL algorithms learn better using the proposed CRADL-based learning and generate distinguishing representations, as observed by compact cluster formation with CRADL. This eventually results in CRADL outperforming the current state-of-the-art studies in PLL setup at different uniform noise rates.