Abstract
How can we learn to do probabilistic inference in a way that generalizes between models? Amortized variational inference learns for a single model, sharing statistical strength across observations. This benefits scalability and model learning, but does not help with generalization to new models. We propose meta-amortized variational inference, a framework that amortizes the cost of inference over a family of generative models. We apply this approach to deep generative models by introducing the MetaVAE: a variational autoencoder that learns to generalize to new distributions and rapidly solve new unsupervised learning problems using only a small number of target examples. Empirically, we validate the approach by showing that the MetaVAE can: (1) capture relevant sufficient statistics for inference, (2) learn useful representations of data for downstream tasks such as clustering, and (3) perform meta-density estimation on unseen synthetic distributions and out-of-sample Omniglot alphabets.
Abstract (translated by Google)
URL
http://arxiv.org/abs/1902.01950