This paper explores the properties of contrastive representation learning, focusing on two key aspects: alignment and uniformity. Alignment refers to the closeness of features from positive pairs, while uniformity refers to the distribution of normalized features on the hypersphere. The authors prove that the contrastive loss asymptotically optimizes these properties and analyze their positive effects on downstream tasks. They introduce metrics to quantify alignment and uniformity, which are empirically shown to strongly agree with downstream task performance. Directly optimizing these metrics leads to representations with comparable or better performance compared to contrastive learning. The paper provides theoretical motivations, empirical validation, and discusses the implications for future research.This paper explores the properties of contrastive representation learning, focusing on two key aspects: alignment and uniformity. Alignment refers to the closeness of features from positive pairs, while uniformity refers to the distribution of normalized features on the hypersphere. The authors prove that the contrastive loss asymptotically optimizes these properties and analyze their positive effects on downstream tasks. They introduce metrics to quantify alignment and uniformity, which are empirically shown to strongly agree with downstream task performance. Directly optimizing these metrics leads to representations with comparable or better performance compared to contrastive learning. The paper provides theoretical motivations, empirical validation, and discusses the implications for future research.