Short Definition
Latent space is the internal representation space where a model encodes hidden factors or learned structure.
Intuition
It is the model's compressed map of what matters about the input, even if those dimensions are not directly human-labeled.
Technical Definition
A latent space consists of learned variables or vectors that represent unobserved factors inferred from data by a model.
Example
An autoencoder may place visually similar images near each other in latent space.
Common Misunderstandings
Latent dimensions do not always have clean human meanings.
Distance in latent space is only meaningful relative to the training objective.