Wiki · AI & Machine Learning · concept
Latent space
The compressed internal 'map' a model learns, where similar things sit close together.
Latent space is the internal map a model builds of its data. Instead of storing raw pixels, letters, or amino acids, a trained model represents each input as a point in a high-dimensional space it has learned — and it arranges that space so that things which behave similarly land near each other. “Latent” just means hidden: you never designed these axes, the model discovered them from data.
How it works
During training, a model is pushed to compress its input into a vector of numbers (an embedding) and still do its job from that vector alone. To succeed, it has to throw away noise and keep the structure that matters. The result is a space where distance means something: two proteins with similar folds, or two sentences with similar meaning, end up close together — even if their raw sequences look nothing alike. Directions in the space often line up with real concepts.
Why it matters (for bio × AI)
This is the “Latent” in LatentCell. Nearly every modern bio-AI method runs through a latent space: protein language models embed sequences so you can search for functional cousins; single-cell foundation models put every cell on one map; structure predictors reason in a learned geometric space. When a model “generalizes,” it’s usually because a new input landed in a well-populated neighborhood of its latent space.
Related
See embeddings for the vectors themselves, and foundation model for the models that learn these spaces at scale.