Latent space, in machine learning, is an abstract, multi-dimensional space serving as a compressed representation of complex data. This space is not directly observable but is learned by a neural network, typically an encoder, which maps high-dimensional input data (an image or a sound, for instance) to a single point, or vector, within it. The main property of a latent space is its organisation – semantically similar inputs are represented in it by points lying close together. This structure reveals hidden relationships in the data and makes possible not only efficient dimensionality reduction but also the generation of new data samples.
Imagine an enormous collection of photographs of human faces. Rather than remembering them pixel by pixel, the computer builds a kind of orderly “map” for them. A latent space is exactly such a clever map. Every face gets its own coordinates on it. The model learns to arrange the map so that photographs of similar faces (smiling women wearing glasses, say) lie close together. This lets the model not only compare similarity but even create new faces – simply by picking an as yet unoccupied spot on the map and seeing what face ought logically to appear there.