Latent Space

Definitions

Latent Space is a compressed, abstract representation of data learned by a neural network, where each dimension captures a meaningful feature or pattern discovered during training. When an autoencoder, generative adversarial network (GAN), or large language model processes high-dimensional input (like images or text), it maps that input into a lower-dimensional latent space where similar items are located near each other. In this space, mathematical operations correspond to semantic transformations: in word embeddings, vector arithmetic like king minus man plus woman yields a point near queen. In image generation models like Stable Diffusion, navigating the latent space produces smooth transitions between images. Variational autoencoders (VAEs) structure the latent space to follow a known distribution, enabling generation of new data by sampling from it. The quality of a model's latent space determines how well it can represent, interpolate between, and generate data. Understanding latent spaces is key to working with generative AI, recommendation systems, and dimensionality reduction techniques.
In plain English: An AI's internal map of concepts — a hidden space where similar ideas are near each other, allowing the AI to smoothly blend or transition between concepts.

1990s: Autoencoders introduce the concept of a compressed 'latent' representation — hidden variables that capture essential structure
2013: Variational Autoencoders formalize the latent space as a smooth, navigable manifold where nearby points produce similar outputs
2022: Stable Diffusion's latent diffusion approach makes latent space a practical concept — images are generated by navigating it