the process of summing (or integrating) over variables we do not observe. In AI, marginalization appears whenever we compute predictions that account for all possible values of latent variables, as in mixture models and variational autoencoders (Chapter 16).