noun
A latent variable is an unobservable factor that explains patterns in visible data, like a hidden lever controlling multiple outcomes. In AI, models like large language models or image generators use latent variables to compress complex information into simpler, learnable patterns.
For example, in a chatbot, a latent variable might represent ‘user intent’—not directly measurable but inferred from words, tone, and context. This allows the model to predict responses without needing explicit labels for every possible input.
For developers, latent variables reduce training costs by simplifying data representation. However, they can also introduce ambiguity, requiring careful tuning to avoid misleading outputs. A poorly calibrated latent space might cause a text-to-image tool to generate distorted faces, as seen in early diffusion models.
A language model uses a latent variable to guess that 'bank' refers to a financial institution, not a river, based on surrounding words.