We propose conjugate energy-based models (EBMs), a class of deep latent-variable models with a tractable posterior. Conjugate EBMs have similar use cases as variational autoencoders, in the sense that they learn an unsupervised mapping between data …