The paper introduces Contrastive Predictive Coding (CPC), a novel unsupervised learning approach designed to extract useful representations from high-dimensional data. The key insight of CPC is to learn representations by predicting future samples in a latent space using powerful autoregressive models. The model employs a probabilistic contrastive loss, which encourages the latent space to capture information that is maximally useful for predicting future samples. This approach is demonstrated to be effective across four distinct domains: speech, images, text, and reinforcement learning in 3D environments. The authors show that CPC learns interesting high-level information, outperforming other approaches in these domains. The paper also discusses related work, including predictive coding and contrastive losses, and provides experimental results to validate the effectiveness of CPC.The paper introduces Contrastive Predictive Coding (CPC), a novel unsupervised learning approach designed to extract useful representations from high-dimensional data. The key insight of CPC is to learn representations by predicting future samples in a latent space using powerful autoregressive models. The model employs a probabilistic contrastive loss, which encourages the latent space to capture information that is maximally useful for predicting future samples. This approach is demonstrated to be effective across four distinct domains: speech, images, text, and reinforcement learning in 3D environments. The authors show that CPC learns interesting high-level information, outperforming other approaches in these domains. The paper also discusses related work, including predictive coding and contrastive losses, and provides experimental results to validate the effectiveness of CPC.