VAE (Variational Autoencoder)

A VAE (Variational Autoencoder) is a type of generative model — a probabilistic autoencoder that compresses and reconstructs input data while also being able to generate entirely new data. It is used primarily for image generation, anomaly detection, and learning latent representations of data, and is known as one of the theoretically stable approaches among deep generative models. The basic structure of a VAE consists of two neural networks: 1. Encoder: Maps input data to a probabilistic latent space (a distribution over latent variables) 2. Decoder: Reconstructs data from those latent variables, approximating the data distribution through training The key difference from a standard autoencoder (AE) is that the VAE treats the latent space as a probability distribution rather than a fixed point. This probabilistic approach produces smoother and more diverse data generation, and improves robustness to noise. The VAE's loss function combines a reconstruction loss with a KL divergence term (Kullback-Leibler Divergence), which regularizes the learned latent distribution to stay close to a standard normal distribution. Main applications: • Generating handwritten digits and face images • Anomaly detection in medical images (detecting deviations from normal patterns) • Feature compression and clustering via latent variables • Generative models for time series and audio (e.g., Sequence VAEs) Compared to GANs (Generative Adversarial Networks), VAEs offer more stable training and better interpretability, making them actively used in both academic research and practical applications.