This paper provides an extensive review of self-supervised learning methods that follow the contrastive learning approach. Contrastive learning aims to embed augmented versions of the same sample close to each other while pushing embeddings from different samples apart. The paper discusses common pretext tasks used in contrastive learning, different architectures proposed, and a performance comparison of various methods across multiple downstream tasks such as image classification, object detection, and action recognition. It also highlights the limitations of current methods and the need for further research to make substantial progress. Contrastive learning is a discriminative approach that groups similar samples together and separates diverse samples. It uses a similarity metric to measure how close two embeddings are. In computer vision tasks, a contrastive loss is evaluated based on the feature representations of images extracted from an encoder network. The paper explains how contrastive learning works, including the use of pretext tasks such as image-inpainting, colorizing greyscale images, jigsaw puzzles, super-resolution, video frame prediction, and audio-visual correspondence. These tasks help in learning features using pseudo-labels. The paper also discusses different architectures used in contrastive learning, including end-to-end learning, using a memory bank, using a momentum encoder, and clustering feature representations. Each architecture is explained with examples of successful methods. The paper highlights the importance of choosing the right pretext task for a model to perform well with contrastive learning. It also discusses the challenges of using large batch sizes and the need for effective optimization strategies. In natural language processing, contrastive learning has been used to learn word representations and sentence embeddings. The paper discusses various pretext tasks used in NLP, such as center and neighbor word prediction, next and neighbor sentence prediction, auto-regressive language modeling, and sentence permutation. These tasks help in learning representations that can be used for downstream tasks such as classification, detection, and understanding. The paper concludes by discussing the open problems in contrastive learning and the need for further research to address these issues. It emphasizes the importance of theoretical analysis and the need for new techniques and paradigms to improve the performance of contrastive learning methods.This paper provides an extensive review of self-supervised learning methods that follow the contrastive learning approach. Contrastive learning aims to embed augmented versions of the same sample close to each other while pushing embeddings from different samples apart. The paper discusses common pretext tasks used in contrastive learning, different architectures proposed, and a performance comparison of various methods across multiple downstream tasks such as image classification, object detection, and action recognition. It also highlights the limitations of current methods and the need for further research to make substantial progress. Contrastive learning is a discriminative approach that groups similar samples together and separates diverse samples. It uses a similarity metric to measure how close two embeddings are. In computer vision tasks, a contrastive loss is evaluated based on the feature representations of images extracted from an encoder network. The paper explains how contrastive learning works, including the use of pretext tasks such as image-inpainting, colorizing greyscale images, jigsaw puzzles, super-resolution, video frame prediction, and audio-visual correspondence. These tasks help in learning features using pseudo-labels. The paper also discusses different architectures used in contrastive learning, including end-to-end learning, using a memory bank, using a momentum encoder, and clustering feature representations. Each architecture is explained with examples of successful methods. The paper highlights the importance of choosing the right pretext task for a model to perform well with contrastive learning. It also discusses the challenges of using large batch sizes and the need for effective optimization strategies. In natural language processing, contrastive learning has been used to learn word representations and sentence embeddings. The paper discusses various pretext tasks used in NLP, such as center and neighbor word prediction, next and neighbor sentence prediction, auto-regressive language modeling, and sentence permutation. These tasks help in learning representations that can be used for downstream tasks such as classification, detection, and understanding. The paper concludes by discussing the open problems in contrastive learning and the need for further research to address these issues. It emphasizes the importance of theoretical analysis and the need for new techniques and paradigms to improve the performance of contrastive learning methods.