Types of Autoencoders & When to Use Them

Autoencoder TypeKey FeatureWhen to Use?
Vanilla Autoencoder (AE)Simple encoder-decoder architectureBasic feature compression
Denoising Autoencoder (DAE)Adds noise to input and learns to reconstructHandles noisy financial data
Sparse Autoencoder (SAE)Enforces sparsity in latent representationSelects key features from many inputs
Variational Autoencoder (VAE)Learns a probabilistic latent spaceUsed in generative modeling, Bayesian inference
Contractive Autoencoder (CAE)Adds Jacobian penalty to prevent overfittingUseful for robust feature learning
Convolutional Autoencoder (CAE)Uses CNNs instead of MLPsIdeal for time-series and structured data
Transformer AutoencoderUses attention mechanismsCaptures long-range dependencies in time-series
LSTM/GRU AutoencoderUses recurrent layers (LSTM/GRU)Best for sequential data like market trends