Latent space

Latent space is the internal coordinate system a model builds, where each input is a point, similar inputs sit near each other, and directions can correspond to meaningful concepts.

"Latent" means hidden — not observed directly, but underlying what is observed. Think of how you mentally organise film songs. Nobody gave you axes, yet in your head some songs are "close" (same mood, same era) and some far apart, and you can move along invisible directions: more energetic, older, sadder. Your mind compressed thousands of songs into a private map. A model's latent space is that map made numeric: every input becomes a point — a vector of a few hundred numbers — placed by learned structure rather than surface details.

The geometry is what makes it useful. Nearness is similarity, which powers search and recommendations — an embedding is precisely a point in such a space. Directions can be concepts: the classic word-vector result king − man + woman ≈ queen is latent-space arithmetic. And smooth paths are transformations: in a generative model, walking from one face's point toward another's yields faces that gradually morph — evidence the space between real examples is filled with plausible inventions.

Where you meet the term: autoencoders compress inputs into a latent code by construction; diffusion-models generate inside a latent space for efficiency; "interpolating in latent space" explains face-morphing apps. One honest caution: the axes themselves are rarely human-readable — meaningful directions exist, but usually as learned combinations, found by probing rather than by design.

Where to go next#

• Full lesson: Embeddings
• Related terms: embedding, autoencoder, diffusion-model, pca