The encoder maps input data into a latent space, typically following a Gaussian distribution. It captures the underlying structure of the data and enables efficient encoding. The decoder, on the other hand, reconstructs the original input from the latent variables. During training, VAEs optimize a loss function that balances reconstruction accuracy with the KL divergence between the learned latent distribution and a prior distribution, usually a unit Gaussian. This unique characteristic allows VAEs to generate diverse outputs from a compact latent space representation, making them valuable for tasks like image synthesis, text generation, and anomaly detection.
