Amortized Inference

Amortized inference is the strategy of learning a mapping from observations to approximate posterior parameters, rather than optimizing the posterior anew for each observation. In traditional variational inference, every new data point requires a separate optimization of the variational free energy. Amortized inference trains a neural network or other function approximator to predict the posterior parameters directly from the data, making inference orders of magnitude faster. This is the core idea behind variational autoencoders and the recognition model in the Free Energy Principle's account of brain function. The brain, on this view, does not solve inference from scratch on each sensory sample; it has learned — through development and evolution — a mapping from sensory patterns to posterior beliefs that is fast, automatic, and precision-weighted.