Machine learning is a paradigm of methods to learn patterns from data. A fundamental challenge in machine learning is how we learn proper representations from data. Supervised learning relies on human labeled data samples to train models. Their training objectives are well-designed losses against input and ground-truth labels. A proper data representation can be learned through such supervised training. However, labeled data is scarce or costly to obtain in many applications. Meanwhile, there is often abundant unlabeled data, such as text and images on the public Internet. How can we automatically learn useful representations from unlabeled data? Autoencoders are a family of methods to learn proper representations. Variational autoencoders (VAEs) are a class of methods to learn autoencoders by introducing probabilistic latent variables and optimizing them with stochastic gradients (Kingma and Welling, Auto-encoding variational bayes. In: ICLR, 2014). Since its inception in 2014, VAE has been widely used for image, speech, and text data modeling and generation. Later, there are extensive follow-up works to improve VAEs. VAEs belong to broader deep generative models generation process. Generative models are models for characterizing probability distributions of data generation process. Generative models serve two important functions: estimating the probability density of data and generating data close to reality. Generative models can be described in different frameworks. Deep generative models are generative frameworks that combine deep neural networks and generative model frameworks. Other deep generative models will be introduced in later chapters.