A compressed, multi-dimensional mathematical space where a machine learning model maps its input data. In this space, similar data points are positioned closer together. It represents the model's internal 'understanding' of the data's features. For example, in a latent space for faces, moving in a certain direction might gradually change a face from smiling to frowning, or male to female. Generative models (like GANs and VAEs) work by sampling points from this space and decoding them back into new, realistic data.
A core concept in manifold learning and dimensionality reduction (e.g., PCA), heavily utilized in modern deep generative models.
The fundamental workspace for Generative AI, allowing for interpolation, style transfer, and semantic manipulation of data.