Latent Space
A compressed, abstract representation space learned by a model where each point represents a meaningful encoding of the input data.
Intuition
Think of latent space as the model's internal 'understanding' of data. When an autoencoder compresses an image to 128 numbers and reconstructs it, those 128 numbers exist in latent space. Similar images map to nearby points.
Navigating Latent Space
In generative models, you can interpolate between points in latent space to smoothly transform one output into another. You can also perform arithmetic: latent(man with glasses) - latent(man) + latent(woman) = latent(woman with glasses).
Applications
Image generation (Stable Diffusion works in latent space), drug discovery (navigating molecular latent spaces), music generation, and any domain where you want a compact, continuous representation of complex data.