An internal coordinate
A latent representation is a set of variables used to describe underlying structure in a model. In an autoencoder, an encoder maps input into a representation and a decoder maps it back toward an output.
Move a coordinate
Our toy decoder maps z to [z, 2z]. Changing z moves along a line in two-dimensional output space. This is illustrative; learned latent spaces can be nonlinear and high-dimensional. Follow the latent coordinate through the fixed decoder to its two output coordinates.
INTERACTIVE EXPLORATION
Move a latent coordinate
Change z for the toy decoder output [z, 2z].
Interpretation needs care
A latent coordinate does not automatically correspond to a human concept such as color or size. Different models and training objectives can produce different representations.
Check your understanding
For z = 2 and decoder [z, 2z], what is the output?
[2, 4].