Gaussian Mixture Models

A Gaussian Mixture Model represents a distribution as a weighted set of Gaussian components, making it useful for finding clusters and describing uncertainty or multiple plausible outcomes. Researchers fit these models with expectation-maximization and gradient methods, and use spectral, Fourier, moment-based, and variational Bayesian methods to estimate parameters or choose the number of components. Neural mixture models extend them to conditional prediction and large datasets, with applications ranging from offline reinforcement learning on D4RL to trajectory forecasting on nuScenes and Argoverse. A central challenge is balancing expressive multimodal representations and scalable computation with reliable component recovery, especially when components overlap or data are high-dimensional.