Understanding Latent Space Learning

What is a Latent Space?

A latent space is a compressed representation of data learned by a neural network.

Properties of Good Latent Spaces

Continuity

Similar inputs should map to nearby points in latent space.

Completeness

Every point in the latent space should correspond to a valid output.

Disentanglement

Different factors of variation should be represented by different dimensions.

Learning Latent Spaces

Autoencoders

• Encode input to latent space
• Decode back to reconstruction
• Minimize reconstruction error

Variational Autoencoders (VAEs)

• Add probabilistic structure
• Regularize latent space to be Gaussian
• Enable sampling and generation