This paper introduces Prototypical Contrastive Learning (PCL), an unsupervised representation learning method that integrates contrastive learning with clustering. PCL learns low-level features for instance discrimination and encodes semantic structures discovered through clustering into the embedding space. The method uses prototypes as latent variables to find the maximum-likelihood estimation of network parameters in an Expectation-Maximization (EM) framework. The E-step involves clustering to find prototype distributions, while the M-step optimizes the network via contrastive learning. The proposed ProtoNCE loss, a generalized version of InfoNCE, encourages representations to be closer to their assigned prototypes. PCL outperforms state-of-the-art instance-wise contrastive learning methods on multiple benchmarks, particularly in low-resource transfer learning. The method also leads to better clustering results. PCL is formulated as an EM algorithm, where the goal is to find the parameters of a Deep Neural Network (DNN) that best describes the data distribution. The method uses prototypes as additional latent variables and estimates their probability in the E-step via k-means clustering. In the M-step, the network parameters are updated by minimizing the ProtoNCE loss. The paper shows that minimizing ProtoNCE is equivalent to maximizing the estimated log-likelihood under the assumption of isotropic Gaussian distributions around prototypes. PCL is shown to be a special case of prototypical contrastive learning, where the prototype for each instance is its augmented feature. The contributions of this paper include the proposal of PCL as a novel framework for unsupervised representation learning, a theoretical framework that formulates PCL as an EM-based algorithm, and the introduction of ProtoNCE, a new contrastive loss that improves InfoNCE by dynamically estimating the concentration of feature distributions around prototypes. PCL outperforms instance-wise contrastive learning on multiple benchmarks with substantial improvements in low-resource transfer learning. The paper also demonstrates that PCL leads to better clustering results. The method is evaluated on transfer learning tasks, and it is shown to outperform MoCo and SimCLR in low-shot classification and semi-supervised learning. PCL is also shown to outperform MoCo and supervised training in object detection and instance segmentation on COCO. The paper concludes that PCL is a generic unsupervised representation learning framework that maximizes the log-likelihood of the observed data by learning an embedding space that encodes the semantic structure of data.This paper introduces Prototypical Contrastive Learning (PCL), an unsupervised representation learning method that integrates contrastive learning with clustering. PCL learns low-level features for instance discrimination and encodes semantic structures discovered through clustering into the embedding space. The method uses prototypes as latent variables to find the maximum-likelihood estimation of network parameters in an Expectation-Maximization (EM) framework. The E-step involves clustering to find prototype distributions, while the M-step optimizes the network via contrastive learning. The proposed ProtoNCE loss, a generalized version of InfoNCE, encourages representations to be closer to their assigned prototypes. PCL outperforms state-of-the-art instance-wise contrastive learning methods on multiple benchmarks, particularly in low-resource transfer learning. The method also leads to better clustering results. PCL is formulated as an EM algorithm, where the goal is to find the parameters of a Deep Neural Network (DNN) that best describes the data distribution. The method uses prototypes as additional latent variables and estimates their probability in the E-step via k-means clustering. In the M-step, the network parameters are updated by minimizing the ProtoNCE loss. The paper shows that minimizing ProtoNCE is equivalent to maximizing the estimated log-likelihood under the assumption of isotropic Gaussian distributions around prototypes. PCL is shown to be a special case of prototypical contrastive learning, where the prototype for each instance is its augmented feature. The contributions of this paper include the proposal of PCL as a novel framework for unsupervised representation learning, a theoretical framework that formulates PCL as an EM-based algorithm, and the introduction of ProtoNCE, a new contrastive loss that improves InfoNCE by dynamically estimating the concentration of feature distributions around prototypes. PCL outperforms instance-wise contrastive learning on multiple benchmarks with substantial improvements in low-resource transfer learning. The paper also demonstrates that PCL leads to better clustering results. The method is evaluated on transfer learning tasks, and it is shown to outperform MoCo and SimCLR in low-shot classification and semi-supervised learning. PCL is also shown to outperform MoCo and supervised training in object detection and instance segmentation on COCO. The paper concludes that PCL is a generic unsupervised representation learning framework that maximizes the log-likelihood of the observed data by learning an embedding space that encodes the semantic structure of data.