Variational Autoencoders are a kind of generative model that use a neural network-based approach to learn a probabilistic mapping of data to a latent space, which is a lower-dimensional representation. Unlike traditional autoencoders that focus on deterministic compression, VAEs enforce a probabilistic structure by introducing a separate distribution for each latent variable. This approach allows them to generate new, similar data samples, making them useful for tasks such as image generation, anomaly detection, or even complex probability distributions learning.