Ranking Info Noise Contrastive Estimation: Boosting Contrastive Learning via Ranked Positives

This paper introduces Ranking Info Noise Contrastive Estimation (RINCE), a new member in the family of InfoNCE losses that preserves ranked ordering positive samples. In contrast to standard loss, which requires strict binary separation training pairs into similar and dissimilar samples, RINCE can exploit information about similarity ranking for learning corresponding embedding space. We show proposed loss function learns favorable embeddings compared whenever at least noisy be obtained or when definition positives negatives is blurry. demonstrate this supervised classification task with additional superclass labels scores. Furthermore, we also applied unsupervised experiments on representation from videos. particular, yields higher accuracy, retrieval rates performs better out-of-distribution detection than loss.