A Variational Autoencoder (VAE) is a generative model that learns to encode data into a probability distribution rather than a fixed point, enabling controlled generation of new data.
Key Differences from Standard Autoencoders:
• Probabilistic encoding - outputs mean and variance
• Continuous latent space - smooth interpolation
• Generative capability - create new samples
• Regularized - prevents overfitting
🎯 The Big Idea
Instead of mapping to a single point, VAEs map to a distribution. This uncertainty allows for generation and ensures similar inputs have similar representations.
VAE vs Standard Autoencoder
Aspect
Standard AE
VAE
Encoding
Deterministic
Probabilistic
Latent Space
Point estimates
Distributions
Generation
Limited
Excellent
Regularization
None inherent
KL divergence
Interpolation
May be discontinuous
Smooth & meaningful
Key Insight:
The "variational" part comes from variational inference - approximating complex distributions with simpler ones!
VAE Architecture Overview
📊
Encoder
Input → μ, σ
→
🎲
Sampling
z ~ N(μ, σ²)
→
🎨
Decoder
z → Output
→
📉
Loss
Recon + KL
Click on each component to learn more about how VAEs work!