Contrastive Predictive Coding (CPC) is an unsupervised learning method that extracts useful representations from high-dimensional data by predicting future observations in a latent space. The key idea is to use powerful autoregressive models to learn representations that maximize the mutual information between the input and the context. The method uses a probabilistic contrastive loss, which is tractable through negative sampling. CPC has been shown to be effective across various domains, including speech, images, text, and reinforcement learning. The approach involves encoding input data into a latent space, then using an autoregressive model to predict future states. The loss function is based on Noise-Contrastive Estimation, which allows for end-to-end training. The model learns to capture shared information between different parts of the data, discarding irrelevant details. This results in compact, meaningful representations that can be used for various downstream tasks. Experiments on four domains—speech, images, natural language, and reinforcement learning—demonstrate that CPC achieves strong performance. For speech, the model successfully captures speaker identities and speech content. In image tasks, CPC representations are used for classification and show competitive results. In natural language, CPC outperforms other methods on several benchmarks. In reinforcement learning, the model improves learning efficiency by incorporating an auxiliary contrastive loss. CPC is a flexible framework that can be applied to various data modalities. It combines autoregressive modeling with predictive coding principles to learn abstract representations without supervision. The method is computationally efficient and has shown promising results in challenging tasks, making it a valuable tool for unsupervised learning.Contrastive Predictive Coding (CPC) is an unsupervised learning method that extracts useful representations from high-dimensional data by predicting future observations in a latent space. The key idea is to use powerful autoregressive models to learn representations that maximize the mutual information between the input and the context. The method uses a probabilistic contrastive loss, which is tractable through negative sampling. CPC has been shown to be effective across various domains, including speech, images, text, and reinforcement learning. The approach involves encoding input data into a latent space, then using an autoregressive model to predict future states. The loss function is based on Noise-Contrastive Estimation, which allows for end-to-end training. The model learns to capture shared information between different parts of the data, discarding irrelevant details. This results in compact, meaningful representations that can be used for various downstream tasks. Experiments on four domains—speech, images, natural language, and reinforcement learning—demonstrate that CPC achieves strong performance. For speech, the model successfully captures speaker identities and speech content. In image tasks, CPC representations are used for classification and show competitive results. In natural language, CPC outperforms other methods on several benchmarks. In reinforcement learning, the model improves learning efficiency by incorporating an auxiliary contrastive loss. CPC is a flexible framework that can be applied to various data modalities. It combines autoregressive modeling with predictive coding principles to learn abstract representations without supervision. The method is computationally efficient and has shown promising results in challenging tasks, making it a valuable tool for unsupervised learning.